BenchCouncil Transactions on Benchmarks, Standards and

Evaluations (TBench) is an open-access multi-disciplinary

journal dedicated to benchmarks, standards, evaluations,

optimizations, and data sets. This journal is a peer-reviewed,

subsidized open access journal where The International Open

Benchmark Council pays the OA fee. Authors do not have to

pay any open access publication fee. However, at least one of

the authors must register BenchCouncil International

Symposium on Benchmarking, Measuring and Optimizing

(Bench) (https://www.benchcouncil.org/bench/) and present

their work. It seeks a fast-track publication with an average

turnaround time of one month.

BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/b5qtx140

Research Article

RESEARCH ARTICLE

A Hybrid MCDM Framework for Assessing Financial

Resilience and Trend Dynamics in Indian Commercial

Banks

Priya Das

1,∗

and Subir Kumar Sen

Research Scholar, Department of Commerce, Tripura University, Agartala, 799022, India and

Professor, Department of Commerce,

Tripura University, Agartala, 799022, India

∗

Corresponding author. priya568das@gmail.com

Received on 1 January 2026; Accepted on 20 May 2026

Abstract

Assessing ﬁnancial resilience in the banking sector requires an integrated framework that captures both cross-sectional

strength and long-term resilience dynamics. In this study, ﬁnancial resilience of Indian commercial banks during the

period 2013–2024 is assessed by using a hybrid MCDM and non-parametric trend analysis approach. Eleven ﬁnancial

indicators covering solvency, asset quality, eﬃciency, and proﬁtability are included in a composite resilience framework,

and criterion weights were objectively determined using the MEREC method. The technique RAM is used to calculate

annual composite resilience scores and ranks the 29 commercial banks. A Mann–Kendall time-series analysis is also

applied to the ﬁnal RAM scores to analyze long-term monotonic trends in bank-level and sector-wide resilience. The

results showed that RAM scores are tightly clustered across banks, suggesting structural convergence in resilience levels.

However, Kruskal-Wallis non-parametric test showed statistically signiﬁcant diﬀerences in the banks’ relative ﬁnancial

resilience across the study period. The MEREC-RAM ranking result showed Kotak Mahindra Bank Ltd. and Tamilnad

Mercantile Bank Ltd. consistently appeared among top at the rankings. While, the Mann-kendall trend test revealed

signiﬁcant Improvement in the resilience of CSB Bank Ltd. and Bank of Maharashtra over the study period. Overall,

combining year-wise relative rankings and monotonic resilience dynamics enables a comprehensive assessment of the

stability of the Indian banking sector, which can oﬀer key insights for regulators, policymakers, and bank management

in strengthening the long-term ﬁnancial resilience of the sector.

Key words: Banks, Financial Resilience, MCDM, MEREC, RAM, Mann-Kendall Trend Analysis

JEL co des: G21, G28, G32, C44, D81

1. Introduction

Financial resilience is the ability of an institution to manage,

respond to, and bounce back from ﬁnancial challenges such as

economic downturns, reduced earnings, or unforeseen costs. For

banks and other ﬁnancial institutions, the ﬁnancial resilience

refers to the ability to withstand shocks, adapt to changing

economic conditions, and maintain stability while continuing to

meet obligations and support economic activities[1, 2]. In the

context of emerging economies, where ﬁnancial systems often

face volatility, weak governance, and high exposure to non-

performing assets (NPAs), evaluating the resilience of banks

is particularly crucial[3, 4]. The resilience of ﬁnancial systems

is inﬂuenced by several factors, such as ﬁrm characteristics,

capital adequacy, liquidity management, risk governance, and

their ability to respond to external shocks[5]. Assessing ﬁnan-

cial resilience enables regulators and policymakers to identify

vulnerabilities and design regulatory frameworks that enhance

risk management and recovery mechanisms[2].

The Indian banking sector has experienced signiﬁcant struc-

tural changes in the past decades. The heavily regulated regime

of the 1980s constrained were limiting operational eﬃciency

of Indian banks and their ability to respond to a rapidly ex-

panding economy[6–8]. However, the economic reforms of the

early 1990s, which included deregulation, partial privatisa-

tion of public sector banks, and interest rate lib eralisation,

brought improvements in pro ductivity, competitiveness, and

risk management practices[6, 9–11] . Conversely, the post-

global ﬁnancial crisis period highlighted persistent weaknesses

in banking systems around the world, including India, in the

form of huge NPAs and governance issues[12, 13]. The evo-

lution of the banking sector in India has built upon previous

structural reforms and has moved towards a more risk sen-

sitive and resilient regulatory framework. This evolution has

P. Das and S. K. Sen

been guided by the adoption of Basel I and II capital ade-

quacy standards, the introduction of risk-based supervision,

and governance reforms in public sector banks aimed at enhanc-

ing asset quality, transparency, and prudential discipline[14].

These eﬀorts laid the groundwork for more speciﬁc post-crisis

reforms, such as the Asset Quality Review (AQR) and the

Insolvency and Bankruptcy Code (IBC) in 2015–2016 that en-

hanced balance-sheet recognition and co diﬁed mechanisms for

stress resolution[15, 16]. Although greater regulation has led

to better governance and operational eﬃciency, there is some

evidence that heavier compliance requirements can reduce prof-

itability and internal capital generation, particularly in times

of economic stress[17]. In 2018, the Punjab National Bank

(PNB) scam revealed serious gaps in internal controls, gov-

ernance mechanisms, and risk oversight in Indian banks. It

demonstrated the need for stronger processes, transparency,

and institutional resilience to avoid systemic vulnerabilities

[18]. At the same time, the digitalization of banking services

has brought new operational risks[19].

These developments point to some of the ongoing chal-

lenges faced by the Indian banking sector in sustaining ﬁnancial

resilience throughout cyclical stress, rising NPAs, and balance-

sheet pressures. Commercial banks, particularly public sector

banks, which account for a substantial share of total banking as-

sets, have remained vulnerable to capital adequacy constraints,

asset quality deterioration, and volatility in proﬁtability. Such

vulnerabilities underline the importance of continuously mon-

itoring the stability of banks and their capacity to absorb

shocks arising from macroeconomic ﬂuctuations and structural

adjustments.

Therefore, banking sector resilience should b e systemati-

cally and multidimensionally assessed to explore how banks

cope with ﬁnancial stress and adapt to regulatory and eco-

nomic changes. In this context, this study aims to evaluate the

ﬁnancial resilience of Indian commercial banks over 12 years

(2013–2024) using a multi-criteria decision-making (MCDM)

framework that incorporates the Method based on the Removal

Eﬀects of Criteria (MEREC)-based objective weighting and the

Root Assessment Method (RAM), supported by non-parametric

trend analysis. The Method RAM is relatively new and ﬁrst ap-

plied in measuring banking sector performance. The MEREC-

RAM approach can be used to constructs composite resilience

scores based on deviations from reference performance levels. It

allows consistent aggregation of heterogeneous ﬁnancial indica-

tors without excessive sensitivity to scaling assumptions[20, 21].

This framework can be combined with the Mann–Kendall trend

test to compare the relative resilience rankings and their evo-

lution over time, extending MCDM applications beyond static

performance comparisons. The unique methodology of the cur-

rent study is the integration of the impact-driven weighting

scheme with a transparent composite scoring approach and a

dynamic time-series evaluation that does not rely on the sub-

jective judgments or correlation-based weights used in other

studies. Instead this model built on a method that assigns

weights by evaluating the contribution of each indicator to the

overall system performance[22].

Accordingly, the present study is organized into ﬁve major

sections. Section 1 presents a brief introduction of the study;

Section 2 outlines the background of the study, mentioning the

relevant literature, as well as hypotheses development based on

relevant theories. Section 3 presents the methodological steps

and pro cedures, Section 4 discusses the major ﬁndings using

tables and ﬁgures, and ﬁnally, Section 5 ends with the study

conclusions mentioning managerial and policy implications, and

future research directions.

2. Background of the Study

2.1. Financial Resilience Background

Financial resilience has become a key determinant in un-

derstanding how individuals, households, institutions, and

economies handle and recover from ﬁnancial shocks. It generally

refers to the ability to maintain stability and well-being during

disruptions by combining ﬁnancial resources, skills, and institu-

tional support. This idea encompasses proactive strategies, in-

cluding accumulating savings, diversifying income sources, and

utilizing eﬀective ﬁnancial tools, as well as having the capacity

to recover or even advance to a better ﬁnancial position[23].

Salignac et al.[1] proposed a foundational framework that ex-

plored the concept of ﬁnancial resilience at the household and

community levels in Australia. The study demonstrated the

role that resources, skills, and access to opportunities play in

resilience. This work was further extended to a global scale by

Klapper and Lusardi[24], who showed that ﬁnancial literacy is

a signiﬁcant predictor of ﬁnancial resilience at b oth the indi-

vidual and household levels, as well as over time[25]. Hamid

et al.[4] investigated the determinants of ﬁnancial resilience in

the context of an emerging economy in Malaysia, highlighting

the key role that savings, income diversiﬁcation, and access to

credit play in enabling individuals and households to build ﬁ-

nancial resilience. Salignac et al.[2] extended this to developing

economies, where institutional and policy supp orts for national

ﬁnancial resilience were found to contribute to inclusive eco-

nomic development. Jansson[26] linked ﬁnancial resilience to

ﬁrms, noting that ﬁnancial position, proﬁtability, and owner-

ship structures impact a ﬁrm’s ability to withstand downturns.

Sreenivasan and Suresh[27] examined start-ups and found that

good liquidity management, innovation, and preparedness help

new entrepreneurs to manage change. In a study of U.S. house-

holds during the COVID-19 pandemic, Clark and Mitchell[28]

found that ﬁscal assistance programs and savings buﬀers were

important for individual ﬁnancial resilience.

Financial resilience in banking and other ﬁnancial insti-

tutions has been studied extensively, including international

banks during the global ﬁnancial crisis[29], banking systems

around the world[30], the U.S. banking system[31], and a loan-

level analysis of ﬁnancial institutions in Mexico[32]. Alam et

al. [33] conducted research on corp orate resilience among ﬁrms

in Bangladesh, speciﬁcally the relationship between corporate

ownership patterns and the resilience of ﬁrms, Daadmehr[34]

proposed a composite ﬁnancial resilience index for workplaces

and ﬁrms, which combines the multi-dimensional matrices (liq-

uidity, leverage, and sustainability practices). Chen and Sun[35]

presented an econometric approach to measuring ﬁnancial

resilience across institutions and economies that combine dy-

namic panel modelling and stress-testing indicators to capture

the temporal response of ﬁnancial entities to shocks.

2.2. MCDM in Bank Performance Evaluation

MCDM techniques have b een extensively used for assess-

ing bank performance in diﬀerent institutional and regional

contexts[36]. Most empirical studies have group ed MCDM

approaches into criteria weighting metho ds and outranking

methods. Studies have combined the two types of MCDM

tools to get a robust performance ranking. AHP–TOPSIS and

IV–TOPSIS was used to evaluate listed private banks in India,

which consistently showed the superiority of HDFC Bank[37],

and a CRITIC-based TOPSIS framework was incorporated to

evaluate public sector banks in India, which showed persistent

performance diﬀerences[38]. Such hybrid frameworks have been

used internationally. Nguyen et al. [39] and Yazdi et al. [40]

showed that bank performance rankings during the COVID-

19 period are contingent on the weighting and aggregation

methods, and that objective weighting schemes are important.

Various alternative weighting and ranking methods have

been compared to improve methodological robustness (e.g.,

Unl¨u et al. [41],

Unvan & Ergen¸c [42], Wanke et al. [43], Sama et

al. [44]). Recent contributions have examined hybrid and inte-

grated models that combine subjective and ob jective factors to

increase discrimination power and ranking stability[45–47]. Mo-

han and Irfan[8] also explained how artiﬁcial intelligence with

MCDM tools was used to assess the performance of banks in

India.

The literature conﬁrms the eﬀectiveness of MCDM frame-

works in assessing bank performance; however, most studies

remain conﬁned to static, cross-sectional p erformance rankings.

Limited attention has been paid to integrating MCDM-based

composite indices with dynamic analyses that capture per-

formance evolution over time, particularly in the context of

ﬁnancial resilience. This gap motivates the present study’s use

of an ob jective weighting scheme and a composite assessment

framework that can support both ranking and temporal trend

analysis.

2.3. Hypotheses Development

Financial resilience can b e conceptualized as a multi-

dimensional construct reﬂecting a bank’s ability to absorb

shocks, adapt to adverse conditions, and maintain core func-

tions over time[48, 49]. While the theory of resilience posits

that a bank’s solvency strength, asset quality, operational eﬃ-

ciency, and proﬁtability will result in resilience. However, the

theory does not suggest that resilience will improve in a lin-

ear or monotonic way, especially in banking systems sub ject to

regulatory reform, economic cycles, and changing risk proﬁles.

From a regulatory perspective, the Capital Buﬀer

Theory[50, 51], as formalized by the Basel framework, states

that capital adequacy and provisioning requirements are ﬁrst

and foremost intended to maintain minimum stability thresh-

olds rather than to continually improve performance. Stricter

asset classiﬁcation standards, supervisory reviews, and correc-

tive action frameworks are some of the regulatory interventions

that force banks to rebalance their capital structures and

portfolios[52]. These measures may increase system-wide stabil-

ity, but it is uncertain how they will ultimately aﬀect composite

resilience measures because improvements in capital or as-

set quality may put short-term pressure on proﬁtability and

eﬃciency.

Moreover, Dynamic Capability Theory[53, 54] provides a

strategic view, which states that resilience can only be im-

proved over the long term if a bank has the ability to perceive

upcoming threats and reorganize its resources in the face of

evolving circumstances[54]. Due to diﬀerences in ownership

structures, managerial skills, and strategic adaptability, not

all banks tend to react to environmental change or regulatory

compliance with higher overall resilience. Consequently, banks

may exhibit heterogeneous and non-uniform resilience trajecto-

ries over time. Combined, these theoretical perspectives suggest

that, at the aggregate level, bank ﬁnancial resilience may not

follow a systematic or statistically discernible long-term trend.

Accordingly, the following null hypothesis is formulated:

: There is no statistically signiﬁcant monotonic trend in the

composite RAM-based ﬁnancial resilience scores of Indian com-

mercial banks over the study period.

: There is no statistically signiﬁcant diﬀerence between

private and public sector banks in India over the study period.

3. Methodology

This study examines the ﬁnancial resilience of selected com-

mercial banks in India from 2013-2014 to 2024-2025.The study

employs an integrated MEREC-based RAM approach within

an MCDM framework to determine the weights of criteria,

rank banks based on their ﬁnancial resilience comp osite scores,

and then applies the Mann-Kendall trend analysis to iden-

tify the performance dynamics and resilience trends of banks

over the years. It also employed Welch two-sample t-test and

Kruskal–Wallis non-parametric test to examine bank speciﬁc

and year-wise signiﬁcant diﬀerences in resilience scores. The

study used equal weights and entropy-based weighting to check

for the sensitivity of weight change on the ranking outcomes. It

further employed TOPSIS, RATMI, and MARCOS techniques

to examine the robustness of the proposed MEREC-RAM

model. Finally, the Kendall’s tau coeﬃcients are determined to

analyse the concordance of ranking outcomes across the meth-

ods. Figure 1 presents the conceptual framework designed for

conducting the research.

Figure 1. Conceptual Framework of the Study. (Source: Authors’

Compilation)

Accordingly, the study integrated secondary data from the

CMIE database and banks’ annual reports from 2013 to 2024.

The data collected for a sample of 29 commercial banks out

of which 17 are private and 12 public sector counterparts. The

study performed robust data cleaning, accounting for any miss-

ing data. Companies with continuous missing data for multiple

variables were excluded from the analysis. To enhance the ac-

curacy of the data, no approximation or rounding has been

performed. Table 1 outlines the study sample, including bank

codes.

Financial resilience in banks refers to the ability to absorb

shocks, maintain core ﬁnancial functions, and adapt to adverse

economic and regulatory conditions. In line with the banking

resilience and ﬁnancial stability literature, resilience is concep-

tualised as a multidimensional construct encompassing solvency

P. Das and S. K. Sen

Code Name of Bank Ownership Co de Name of Bank Ownership

CB1 Axis Bank Ltd. Private CB16 Indusind Bank Ltd. Private

CB2 Bank of Baroda Public CB17 Jammu & Kashmir Bank Ltd. Private

CB3 Bank of India Public CB18 Karnataka Bank Ltd. Private

CB4 Bank of Maharashtra Public CB19 Karur Vysya Bank Ltd. Private

CB5 CSB Bank Ltd. Private CB20 Kotak Mahindra Bank Ltd. Private

CB6 Canara Bank Public CB21 Punjab & Sind Bank Public

CB7 Central Bank of India Public CB22 Punjab National Bank Public

CB8 City Union Bank Ltd. Private CB23 RBL Bank Ltd. Private

CB9 DCB Bank Ltd. Private CB24 South Indian Bank Ltd. Private

CB10 Dhanlaxmi Bank Ltd. Private CB25 State Bank of India Public

CB11 Federal Bank Ltd. Private CB26 Tamilnad Mercantile Bank Ltd. Private

CB12 HDFC Bank Ltd. Private CB27 UCO Bank Public

CB13 ICICI Bank Ltd. Private CB28 Union Bank of India Public

CB14 Indian Bank Public CB29 Yes Bank Ltd. Private

CB15 Indian Overseas Bank Public

Table 1. Selected Commercial Banks (CB). (Source: Authors’ Compilation)

Resilience Code Attributes Descriptions Expected Outcome

Solvency

C1 CAR

Tier I Capital + Tier II Capital

Risk – Weighted Assets

Max

C2 Tier-1 CAP

Tier I Capital

Risk – Weighted Assets

Max

C3 D/E

Total debt

Shareholders’ Equity

Min

Assets Quality

C4 GNPA

Gross NPAs

Gross Advances

Min

C5 NNPA

Net NPAs

Net Advances

Min

C6 LGR

Total Advances

− Total Advances

t−1

Total Advances

t−1

Min

C7 PCR

Total Provisions

Gross NPAs

Max

Eﬃciency

C8 C-I

Operating Expenses

Total Income

Min

C9 C-D

Total Advances

Total Deposits

Min

Proﬁtability

C10 ROA

Net Proﬁt

Total Assets

Max

C11 ROE

Net Proﬁt

Shareholders’ Equity

Max

Table 2. Description of Selected Variables (Criteria). (Source: Authors’ Compilation)

strength, asset quality, operational eﬃciency, and proﬁtability.

Therefore, the criteria or resilience indicators are grouped into

four major dimensions: Solvency, Asset Quality, Eﬃciency, and

Proﬁtability, and treated as indicators of banks’ ﬁnancial re-

silience. Eleven ﬁnancial ratios are incorporated under the four

dimensions of the resilience framework. Table 2 presents the de-

scriptions of the resilience indicators (variables). The variables

are selected and established based on the available literature

and related concepts and theories on banks’ p erformance and

ﬁnancial resilience[3, 4, 37, 38, 43, 55–63].

3.1. MEREC Approach

Keshavarz et al.[22] proposed a new ob jective weighting

method, MEREC, which utilizes the removal eﬀect on alter-

natives to determine attribute weights. In contrast to previous

criteria weight calculation techniques, the MEREC method

relies on how the removal of conditions aﬀects the total ef-

fects of substitutes[64, 65]. Higher weights are assigned to

attributes with greater performance eﬀects, and smaller weights

are given to attributes with smaller performance eﬀects. It

gives an innovative approach to the criteria weight calculation

technique.

The method follows the following six steps[22, 64, 65]:

Step 1: Construct the decision/evaluation matrix. The

multi-criteria decision-making incorporates an m × n matrix

where m is the number of alternatives and n is the number of

criteria

X = x







·· · x







(1)

The elements of the matrix are denoted by x

, where i =

1, . . . , m, and j = 1, . . . , n. The values in the matrix should

be greater than zero (x

> 0) as well as a positive integer.

Subsequently, any negative criteria values will be adjusted using

the proper method.

Step 2: Normalize the decision matrix n

Simple linear normalization procedure is applied to scale the

elements of the decision matrix.

max x

if j ∈ B (2)

min x

if j ∈ C (3)

Where B is the beneﬁt group of criteria (Max-type), and C

represents the cost criteria (Min-type)

Step 3: Calculate the overall performance P

In this step, a logarithmic aggregation measure with equal

weights for all criteria is employed to compute the overall per-

formance values P

. This measure is based on a non-linear

function. Based on the normalized decision matrix obtained

in the previous step, larger values of n

indicate better

performance of the alternatives.

= ln





1 +





j=1

|ln(n









(4)

Step 4: Determine the performance of alternatives by

removing each criterion

This step evaluates the eﬀect of each criterion on the over-

all performance of the alternatives by excluding one criterion

at a time. Let P

′

denote the overall performance of the i

alternative when the j

criterion is removed.

The p erformance value after removing the j

criterion is

computed as:

′

= ln





1 +





n − 1

k=1, k=j

|ln(n









(5)

Step 5: The summation of absolute deviations

In this step, we calculate the removal eﬀect of the j

crite-

rion based on the values obtained in steps 3 and 4. Let R

the removal eﬀect of the j

criterion

i=1

′

− P

| (6)

Step 6: Determine the ﬁnal weights of the criterion

j=1

(7)

3.2. RAM Method

The Root Assessment Method (RAM) aims to derive the utility

value of each alternative by aggregating its scores over deci-

sion criteria. The Method was ﬁrst applied to sustainability-

focused multi-criteria decision problems, addressing complex

evaluations with transparent ranking logic[20]. Subsequent

studies extended its application to energy technology selec-

tion, manufacturing systems, urban and cost-of-living assess-

ment, and comparative methodological analysis, including

fuzzy, spherical fuzzy, and neutrosophic environments to man-

age uncertainty[21, 66–68]. The RAM method is selected in

this bank resilience study, considering its low computational

complexity, stable rankings[20], and eﬀective integration with

MEREC-based objective weighting, which may outperform con-

ventional outranking MCDM techniques in complex ﬁnancial

evaluations.

The method involves the following major steps[20, 21, 68]:

Step 1: The ﬁrst step involves forming an initial matrix X.

This is preceded by deﬁning the m set of alternatives A

and a

set of n criteria C

D = x







·· · x

. ·· ·

·· · x

m×n







(8)

Where alternatives i = 1, 2, . .. , m and criterion j = 1, 2,

. . . ,n canbe expressed based on the nature of the criteria, i.e.,

cost (C) or b eneﬁt (B).

Step 2: Normalization of the initial matrix (x) using the

linear sum normalization

i=1

(9)

Where elements x

represents the elements of the matrix x.

Step 3: Determination of the weighted normalized matrix

= n

× w

(10)

Where n

represents the normalized values calculated in

step 2, and w

represents the MEREC weights.

Step 4: Calculate the aggregates of weighted normal-

ized values of beneﬁcial and non-beneﬁcial criteria for each

alternative using the following equations

j=1

+ij

(11)

−i

j=1

−ij

(12)

Whereas K

is the sum of weighted normalized values

of the beneﬁcial criteria and K

−i

is the sum of weighted

normalized values of the cost criteria.

Step 5: Determination of the aggregate relative resilience

scores of alternatives using the following function Q

2+K

−i

2 + K

(13)

The ranking of alternatives relies on the ﬁnal values Q

descending order.

3.3. Mann–Kendall Trend Analysis

To determine whether the ﬁnancial resilience of individual

banks exhibited a monotonic trend over the study perio d, the

univariate Mann–Kendall non-parametric trend test was ap-

plied to the composite RAM-based ﬁnancial resilience scores for

each bank from 2013 to 2024. The Mann–Kendall test is widely

used for trend detection in time-series data because it does not

require assumptions of normality and is robust to outliers.

P. Das and S. K. Sen

The test statistic is computed as:

S =

n−1

i=1

j=i+1

sgn(x

− x

) (14)

Where n is the number of years, x

, and x

denote the com-

posite RAM scores observed at time points i and j, respectively,

and the sign function sgn(·) is deﬁned as:

sgn(x

− x

) =











+1, if x

− x

> 0,

0, if x

− x

= 0,

−1, if x

− x

< 0.

For sample sizes greater than ten, the statistic S is stan-

dardized to a normal variate Z ,

Where, Z =











S−1

√

Var(S)

if S > 0

0 if S = 0

S+1

√

Var(S)

if S < 0

Here,

Var(S) =

n(n − 1)(2n + 5)

It tests the following hypotheses:

: No monotonic trend exists in the series

: A monotonic trend exists in the series

A positive and statistically signiﬁcant Z-value indicates an

increasing trend in ﬁnancial resilience, whereas a negative and

statistically signiﬁcant Z-value indicates a decreasing trend.

Statistical signiﬁcance is evaluated at the 5% level (p < 0.05).

4. Results and Discussions

4.1. Analysis using MEREC

The study applies a multi-stage MCDM-integrated Mann-

Kendall model. Initially, a 29 × 11 (m × n) decision matrix

is developed based on 11 criteria related to 29 banks (alterna-

tives), for each year from 2013-2014 to 2024-2025, using Eq. (1).

The relative importance of each criterion or ﬁnancial resilience

indicator is determined using the MEREC weighting method.

Then, the weights obtained are integrated into the RAM ap-

proach to generate the resilience index. The ranking of banks is

determined based on the aggregate function calculated at the

end of the RAM approach.

The values in the decision matrix (Eq. (1)) were normalized

for each period using Eq. (2) using the linear min-max normal-

ization metho d to convert all the cost-beneﬁt criteria into the

same measurement scale. Subsequently, a logarithmic aggrega-

tion measure based on equal criterion weights is applied to the

normalized values n

to compute the P

values in Eq. (4). The

values represent the initial overall performance scores of the

alternatives, taking into account all criteria. Accordingly, the

modiﬁed performance values P

′

are calculated using Eq. (5) by

removing one criterion at a time from the set of criteria associ-

ated with alternatives, i.e., column-wise elimination. Equation

(6) is then used to evaluate the criterion removal eﬀect R

measuring the absolute deviation between P

′

and P

for each

alternative i and criterion j . Finally, the criterion weights W

for each study period are determined using Eq. (7) and are

presented in Table 3.

The MEREC results, as shown in Table 3, reveals moderate

variation in criteria weights over time. CAR (C1) increases from

0.046 (2013) to 0.123 (2022), reﬂecting rising regulatory em-

phasis. Loan growth (C6) ﬂuctuates between 0.064 and 0.092,

indicating cyclical sensitivity. Cost eﬃciency (C8) also varies

(0.069–0.098), while ROA (C10) and ROE (C11) remain rela-

tively stable (0.088–0.102). This suggests limited diﬀerentiation

and a consistently lower role in resilience assessment.

Figure 2. Relative Importance of Financial Resilience Indicators

(2013-2024). (Source: Estimated by the authors)

Fig 2 highlights temporal shifts in criteria importance as well

as ﬂuctuations in loan growth and eﬃciency indicators. This

reﬂects impact of real-world events such as the AQR and IBC

phases. On the other hand relative weight stability in proﬁtabil-

ity indicators suggests consistent regulatory and operational

frameworks.

4.2. Analysis using RAM

The method RAM evaluated banking resilience by aggregat-

ing multiple ﬁnancial indicators in a structured manner. The

ﬁrst step is to build the initial decision matrix D (Eq. (8)),

which summarizes the performance of each banking alternative

over all selected resilience indicators. Next it applies linear sum

normalization (Eq. (9)) to enable logical comparison across in-

dicators measured in diﬀerent units. The normalized matrix

is then weighted (Eq. (10)) with the MEREC-derived weight

coeﬃcients shown in Table 3 to produce the weighted normal-

ized matrix V , which ensures that criteria with higher systemic

relevance have a stronger eﬀect on the evaluation.

RAM explicitly accounts for the asymmetric roles of beneﬁ-

cial and non-beneﬁcial criteria by aggregating them separately,

as expressed in Equations(11) and (12). K

is the sum of

weighted normalized values of the beneﬁcial criteria and K

−i

is the sum of weighted normalized values of the cost criteria

which are presented in Table 4. Beneﬁcial criteria contribute

positively to resilience, whereas non-beneﬁcial criteria capture

vulnerability channels. This separation enables the ﬁnal RAM

utility score Q

, computed using Eq. (13), to reﬂect the balance

between shock-absorption capacity and risk exposure.

Table 5 shows the RAM-based aggregate relative resilience

scores (Q

) of each bank from 2013 to 2024. The (Q

) scores

show a high level of consistency. The values generally fall

between about 1.41 and 1.42.

The Q

values are normalized and presented in Table 6

which shows the ﬁnancial resilience ranges between 0.995 and

C1 C2 C3 C4 C5 C6 C7 C8 C9 C10 C11

2013 0.046 0.101 0.095 0.095 0.091 0.082 0.101 0.082 0.103 0.102 0.102

2014 0.060 0.100 0.083 0.093 0.090 0.079 0.097 0.098 0.101 0.100 0.100

2015 0.043 0.103 0.094 0.098 0.096 0.091 0.103 0.069 0.103 0.100 0.101

2016 0.073 0.099 0.084 0.091 0.089 0.077 0.099 0.096 0.098 0.097 0.097

2017 0.047 0.101 0.093 0.097 0.095 0.084 0.102 0.078 0.102 0.100 0.100

2018 0.041 0.103 0.096 0.098 0.097 0.092 0.087 0.080 0.102 0.102 0.102

2019 0.054 0.100 0.093 0.096 0.094 0.087 0.102 0.074 0.101 0.101 0.101

2020 0.090 0.096 0.092 0.089 0.087 0.064 0.098 0.093 0.097 0.096 0.097

2021 0.065 0.099 0.095 0.094 0.093 0.064 0.101 0.097 0.100 0.096 0.096

2022 0.123 0.091 0.078 0.086 0.085 0.084 0.095 0.088 0.093 0.088 0.090

2023 0.048 0.101 0.098 0.100 0.098 0.077 0.103 0.073 0.103 0.100 0.099

2024 0.100 0.093 0.080 0.090 0.090 0.080 0.097 0.091 0.096 0.092 0.091

Table 3. Year-Wise Obtained Criteria Weights using the MEREC Approach. (Source: Estimated by the authors)

2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024

−i

CB1 0.019 0.015 0.019 0.015 0.021 0.018 0.018 0.018 0.017 0.022 0.014 0.020 0.017 0.019 0.018 0.016 0.019 0.017 0.014 0.014 0.018 0.018 0.018 0.013

CB2 0.015 0.018 0.015 0.014 0.011 0.019 0.015 0.018 0.015 0.021 0.013 0.019 0.015 0.024 0.015 0.018 0.014 0.020 0.017 0.014 0.015 0.019 0.016 0.018

CB3 0.013 0.023 0.013 0.020 0.009 0.026 0.013 0.025 0.012 0.024 0.012 0.023 0.015 0.023 0.016 0.018 0.013 0.020 0.014 0.019 0.013 0.025 0.014 0.023

CB4 0.016 0.021 0.019 0.023 0.012 0.023 0.009 0.027 0.013 0.021 0.005 0.025 0.016 0.019 0.016 0.014 0.015 0.014 0.019 0.010 0.020 0.013 0.021 0.015

CB5 0.012 0.020 0.010 0.063 0.008 0.010 0.015 0.018 0.013 0.012 0.012 0.011 0.020 0.015 0.020 0.015 0.027 0.014 0.026 0.008 0.021 0.015 0.019 0.022

CB6 0.014 0.020 0.014 0.017 0.011 0.021 0.015 0.022 0.013 0.022 0.011 0.021 0.015 0.020 0.015 0.023 0.014 0.020 0.017 0.018 0.016 0.022 0.017 0.020

CB7 0.010 0.027 0.013 0.022 0.009 0.023 0.010 0.024 0.008 0.024 0.007 0.022 0.014 0.025 0.015 0.020 0.012 0.024 0.012 0.017 0.013 0.021 0.015 0.018

CB8 0.018 0.012 0.019 0.011 0.021 0.013 0.021 0.012 0.021 0.013 0.016 0.014 0.017 0.014 0.019 0.013 0.020 0.019 0.020 0.017 0.018 0.024 0.019 0.020

CB9 0.018 0.016 0.019 0.015 0.021 0.014 0.019 0.014 0.020 0.015 0.016 0.014 0.018 0.015 0.018 0.017 0.016 0.019 0.016 0.015 0.014 0.023 0.014 0.027

CB10 0.006 0.022 0.008 0.032 0.011 0.011 0.016 0.026 0.016 0.013 0.013 0.016 0.017 0.014 0.015 0.030 0.009 0.021 0.011 0.022 0.008 0.017 0.009 0.021

CB11 0.019 0.012 0.019 0.012 0.018 0.014 0.018 0.014 0.019 0.016 0.015 0.015 0.017 0.014 0.017 0.012 0.017 0.015 0.017 0.012 0.016 0.017 0.016 0.015

CB12 0.019 0.014 0.020 0.011 0.022 0.015 0.021 0.012 0.021 0.014 0.017 0.015 0.021 0.014 0.020 0.011 0.023 0.012 0.021 0.009 0.018 0.020 0.018 0.014

CB13 0.019 0.020 0.019 0.020 0.020 0.022 0.019 0.020 0.019 0.021 0.015 0.019 0.017 0.016 0.019 0.012 0.023 0.013 0.022 0.010 0.020 0.016 0.020 0.013

CB14 0.015 0.018 0.015 0.014 0.015 0.016 0.017 0.016 0.017 0.019 0.013 0.018 0.016 0.018 0.017 0.025 0.014 0.017 0.015 0.013 0.014 0.017 0.017 0.015

CB15 0.012 0.024 0.011 0.024 0.008 0.034 0.009 0.032 0.007 0.028 0.011 0.030 0.010 0.020 0.017 0.016 0.015 0.018 0.014 0.019 0.014 0.020 0.016 0.018

CB16 0.019 0.016 0.018 0.016 0.022 0.017 0.021 0.014 0.021 0.017 0.015 0.020 0.017 0.017 0.018 0.013 0.019 0.014 0.020 0.011 0.018 0.017 0.011 0.017

CB17 0.019 0.012 0.015 0.014 0.017 0.018 0.009 0.017 0.016 0.016 0.013 0.019 0.013 0.017 0.015 0.016 0.013 0.022 0.016 0.016 0.015 0.017 0.017 0.017

CB18 0.015 0.017 0.016 0.014 0.017 0.013 0.018 0.014 0.017 0.015 0.014 0.016 0.016 0.016 0.016 0.012 0.014 0.016 0.018 0.013 0.018 0.021 0.016 0.019

CB19 0.017 0.014 0.018 0.011 0.020 0.012 0.018 0.012 0.018 0.015 0.015 0.018 0.017 0.016 0.018 0.016 0.017 0.016 0.019 0.009 0.018 0.013 0.019 0.016

CB20 0.020 0.015 0.020 0.015 0.021 0.022 0.021 0.013 0.022 0.014 0.018 0.015 0.020 0.013 0.021 0.010 0.024 0.012 0.024 0.008 0.021 0.013 0.022 0.012

CB21 0.012 0.022 0.013 0.020 0.015 0.018 0.014 0.020 0.013 0.020 0.076 0.022 0.012 0.025 0.010 0.029 0.014 0.022 0.015 0.022 0.011 0.023 0.014 0.025

CB22 0.014 0.023 0.014 0.021 0.010 0.028 0.014 0.021 0.008 0.026 0.008 0.023 0.016 0.022 0.015 0.024 0.010 0.024 0.011 0.021 0.012 0.020 0.018 0.019

CB23 0.016 0.032 0.017 0.022 0.018 0.024 0.018 0.018 0.019 0.017 0.015 0.019 0.017 0.018 0.017 0.014 0.010 0.016 0.014 0.015 0.013 0.020 0.012 0.015

CB24 0.016 0.014 0.014 0.011 0.015 0.016 0.016 0.013 0.016 0.015 0.013 0.018 0.014 0.019 0.014 0.018 0.009 0.026 0.013 0.025 0.015 0.022 0.016 0.022

CB25 0.015 0.023 0.015 0.017 0.016 0.023 0.016 0.020 0.015 0.023 0.012 0.021 0.016 0.018 0.016 0.018 0.015 0.019 0.016 0.015 0.015 0.021 0.015 0.017

CB26 0.017 0.015 0.018 0.010 0.020 0.012 0.019 0.013 0.019 0.011 0.016 0.015 0.019 0.012 0.020 0.014 0.024 0.014 0.025 0.008 0.022 0.015 0.024 0.021

CB27 0.015 0.024 0.014 0.019 0.008 0.028 0.011 0.024 0.009 0.028 0.008 0.026 0.013 0.022 0.015 0.018 0.010 0.017 0.012 0.092 0.012 0.021 0.012 0.022

CB28 0.014 0.022 0.014 0.021 0.014 0.022 0.014 0.025 0.011 0.024 0.010 0.025 0.014 0.025 0.015 0.029 0.014 0.023 0.015 0.018 0.016 0.019 0.018 0.016

CB29 0.019 0.017 0.019 0.017 0.021 0.019 0.020 0.012 0.019 0.025 0.013 0.026 0.006 0.035 0.012 0.030 0.012 0.039 0.010 0.026 0.008 0.019 0.011 0.019

Table 4. Aggregates of Weighted Normalized Values of the Beneﬁcial (K

) and Cost (K

−i

) Criteria

1. This normalization simpliﬁed it to compare the scores across

banks and over the years. Most banks keep consistently high

normalized scores, reﬂecting strong and stable performance

throughout the study perio d. The limited spread in normal-

ized Q

values highlights the stability of the banking system,

while the observed diﬀerences give valuable insights into how

banks perform relative to each other. This supports the reliabil-

ity of the RAM-based evaluation and suggests that the banking

system exhibits consistent resilience with limited dispersion in

performance levels.

The banks such as CB8, CB11, CB12, and CB20 consistently

reﬂected higher Q

scores across multiple years. These banks

demonstrate minimal ﬂuctuations, reﬂecting strong adaptabil-

ity and sustained resilience under uncertainties. In contrast,

banks such as CB5, CB6, CB10, and CB15 demonstrates

slightly lower Q

values and maximum variation over time. Al-

though their scores remain within a high range, the observed

ﬂuctuations suggest relatively lower stability and sensitivity to

macro-level changes compared to the top-performing banks.

Accordingly, Table 7 presents the year-wise ranking results

obtained using the MEREC-RAM approach.

The annual RAM-based rankings provide additional insight

about banks’ relative ﬁnancial resilience over time. While the

changes in normalized Q

values between banks are almost

negligible, the modest variations discovered have a substantial

impact on year-over-year rankings. The narrow dispersion of

RAM scores reﬂects the standardized and highly regulated na-

ture of the Indian banking system under RBI and Basel norms.

This demonstrates that even slight changes in performance can

aﬀect how banks compare in a competitive market.

A set of banks frequently holds the top ranks. In For

instance, Kotak Mahindra Bank Ltd. (CB20) performed ex-

ceptionally well, frequently ranking among the top places

and achieving ﬁrst position multiple times, particularly in re-

cent years. Similarly, Tamilnad Mercantile Bank Ltd. (CB26),

HDFC Bank Ltd. (CB12) and ICICI Bank Ltd. (CB13) main-

tained stable ranks throughout the study period. City Union

Bank Ltd. (CB8) and Federal Bank Ltd. (CB11) also exhibited

relatively high and stable rankings, with minor ﬂuctuations.

These banks can be considered as consistently resilient per-

formers within the sample. On the other hand, CSB Bank Ltd.

(CB5) reﬂected signiﬁcant improvement in rankings from bot-

tom to top ten over the years. Bank of Maharashtra (CB4) also

demonstrated signiﬁcant improvement in rankings by reaching

to the second place during 2024.

P. Das and S. K. Sen

2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024

CB1 1.417 1.417 1.417 1.416 1.415 1.414 1.416 1.417 1.417 1.416 1.416 1.417

CB2 1.415 1.416 1.414 1.415 1.414 1.414 1.414 1.415 1.414 1.416 1.415 1.415

CB3 1.413 1.414 1.411 1.413 1.413 1.413 1.414 1.415 1.414 1.414 1.413 1.414

CB4 1.415 1.415 1.413 1.411 1.413 1.410 1.415 1.416 1.416 1.419 1.418 1.418

CB5 1.414 1.403 1.414 1.415 1.416 1.416 1.418 1.418 1.420 1.421 1.418 1.416

CB6 1.414 1.415 1.413 1.414 1.413 1.413 1.414 1.414 1.414 1.416 1.415 1.415

CB7 1.411 1.413 1.412 1.412 1.411 1.411 1.413 1.414 1.413 1.414 1.413 1.415

CB8 1.418 1.418 1.418 1.419 1.418 1.417 1.417 1.418 1.417 1.417 1.415 1.416

CB9 1.417 1.417 1.418 1.417 1.418 1.416 1.417 1.416 1.415 1.416 1.413 1.413

CB10 1.411 1.409 1.415 1.413 1.417 1.415 1.417 1.412 1.412 1.413 1.413 1.412

CB11 1.418 1.418 1.417 1.417 1.417 1.416 1.417 1.417 1.417 1.417 1.416 1.416

CB12 1.417 1.418 1.418 1.419 1.418 1.417 1.418 1.419 1.419 1.420 1.416 1.417

CB13 1.416 1.416 1.416 1.416 1.416 1.415 1.416 1.418 1.419 1.420 1.417 1.418

CB14 1.415 1.416 1.416 1.416 1.416 1.414 1.415 1.414 1.415 1.416 1.415 1.417

CB15 1.413 1.412 1.409 1.409 1.410 1.411 1.413 1.416 1.415 1.415 1.414 1.415

CB16 1.417 1.417 1.418 1.418 1.417 1.415 1.416 1.417 1.417 1.418 1.416 1.414

CB17 1.418 1.416 1.416 1.413 1.416 1.414 1.415 1.416 1.413 1.416 1.415 1.416

CB18 1.415 1.416 1.417 1.417 1.416 1.415 1.416 1.417 1.415 1.417 1.415 1.415

CB19 1.417 1.418 1.418 1.418 1.417 1.415 1.416 1.416 1.417 1.419 1.417 1.417

CB20 1.417 1.418 1.416 1.418 1.419 1.417 1.418 1.419 1.420 1.421 1.418 1.419

CB21 1.413 1.414 1.415 1.414 1.414 1.435 1.413 1.411 1.414 1.414 1.412 1.413

CB22 1.414 1.414 1.411 1.414 1.411 1.411 1.414 1.414 1.412 1.413 1.413 1.416

CB23 1.412 1.415 1.415 1.416 1.417 1.415 1.416 1.417 1.414 1.415 1.414 1.415

CB24 1.416 1.416 1.416 1.417 1.416 1.414 1.415 1.415 1.411 1.413 1.414 1.414

CB25 1.414 1.415 1.414 1.415 1.414 1.413 1.415 1.415 1.415 1.416 1.414 1.415

CB26 1.417 1.418 1.418 1.418 1.418 1.416 1.418 1.418 1.419 1.421 1.418 1.417

CB27 1.414 1.414 1.410 1.412 1.411 1.411 1.414 1.415 1.413 1.397 1.413 1.413

CB28 1.414 1.414 1.414 1.413 1.412 1.412 1.413 1.412 1.414 1.415 1.415 1.417

CB29 1.417 1.417 1.417 1.418 1.415 1.412 1.408 1.411 1.409 1.412 1.412 1.413

Table 5. Aggregate resilience scores Q

of banks for each year from 2013-2024. (Source: Authors’ Computation)

In contrast, several banks show persistent lower rankings,

indicating relatively weaker resilience. For instance, Bank of

India (CB3), Central Bank of India (CB7) and Dhanlaxmi

Bank Ltd. (CB10), UCO Bank (CB27) frequently app eared in

the lower ranks. This indicates continuous structural or per-

formance challenges. These banks also exhibit higher ranking

divergence reﬂecting instability in their resilience over time. On

the other hand, Yes Bank Ltd. (CB29) experienced substantial

declines in their rankings especially from 2018, indicative of the

existence of underlying challenges and diﬀerent government’s

correction measures such as AQR and IBC implementations

during 2016-2017, as well as interventions with revised Prompt

Corrective Action (PCA).

Figure 3 presents the heatmap on normalized resilience score

of Qi across diﬀerent period. The colour gradient, ranging from

darker shades (lower scores) to lighter (higher scores) provides

a visual representation of the variation in ﬁnancial resilience

across banks and over time. The colour gradient, ranging from

darker shades (lower scores) to lighter shades (higher scores),

provides a visual representation of the variation in ﬁnancial

resilience across banks and over time. A noticeable deviation

is observed around 2018, where a darker vertical band appears

across a large number of banks. This indicates a sector-wide

decline in resilience during that year, pointing to the presence of

a systemic sho ck aﬀecting the entire banking system rather than

isolated bank-speciﬁc issues. This might be manifested from two

major events in the Indian ﬁnancial system e.g., PNB scam

and the IL&FS crisis during the year 2018 which signiﬁcantly

aﬀected the entire system.

Figure 3. Heatmap on banks year-wise composite ﬁnancial resilience

scores obtained using MEREC-RAM approach. (Source: Authors’

Computation)

Following this period, it shows a gradual return to normal

range, reﬂecting recovery and improved resilience in subsequent

2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024

CB1 0.999 0.999 0.999 0.998 0.997 0.985 0.998 0.998 0.997 0.996 0.998 0.999

CB2 0.998 0.998 0.997 0.998 0.997 0.985 0.997 0.997 0.996 0.997 0.998 0.997

CB3 0.997 0.997 0.995 0.996 0.996 0.984 0.997 0.997 0.995 0.995 0.996 0.996

CB4 0.997 0.998 0.996 0.995 0.996 0.983 0.998 0.998 0.997 0.998 1.000 0.999

CB5 0.997 0.989 0.997 0.997 0.998 0.986 0.999 0.999 1.000 1.000 1.000 0.998

CB6 0.997 0.998 0.996 0.997 0.996 0.985 0.997 0.996 0.996 0.996 0.997 0.997

CB7 0.995 0.996 0.995 0.995 0.995 0.983 0.996 0.997 0.995 0.995 0.997 0.997

CB8 1.000 1.000 1.000 1.000 1.000 0.987 0.999 0.999 0.997 0.997 0.997 0.998

CB9 0.999 0.999 1.000 0.999 0.999 0.987 0.999 0.998 0.996 0.996 0.996 0.996

CB10 0.995 0.994 0.998 0.996 0.999 0.986 0.999 0.995 0.994 0.994 0.996 0.995

CB11 1.000 1.000 0.999 0.999 0.999 0.987 0.999 0.999 0.997 0.997 0.998 0.998

CB12 0.999 1.000 1.000 1.000 1.000 0.987 1.000 0.999 0.999 0.999 0.998 0.999

CB13 0.998 0.998 0.998 0.998 0.998 0.986 0.999 0.999 0.999 0.999 0.999 0.999

CB14 0.998 0.998 0.998 0.998 0.998 0.985 0.998 0.996 0.996 0.996 0.998 0.998

CB15 0.996 0.996 0.993 0.994 0.994 0.983 0.996 0.998 0.996 0.995 0.997 0.998

CB16 0.999 0.999 1.000 1.000 0.999 0.986 0.998 0.999 0.998 0.998 0.999 0.996

CB17 1.000 0.998 0.998 0.996 0.998 0.985 0.997 0.997 0.995 0.996 0.998 0.998

CB18 0.998 0.999 0.999 0.999 0.999 0.986 0.998 0.998 0.996 0.997 0.998 0.997

CB19 0.999 0.999 1.000 0.999 0.999 0.986 0.999 0.998 0.997 0.998 0.999 0.999

CB20 0.999 0.999 0.998 1.000 1.000 0.987 1.000 1.000 1.000 0.999 1.000 1.000

CB21 0.996 0.997 0.998 0.997 0.997 1.000 0.996 0.994 0.995 0.995 0.996 0.996

CB22 0.997 0.997 0.995 0.997 0.994 0.983 0.997 0.996 0.994 0.994 0.997 0.998

CB23 0.996 0.997 0.997 0.998 0.999 0.986 0.998 0.998 0.996 0.996 0.997 0.997

CB24 0.999 0.999 0.998 0.999 0.998 0.985 0.997 0.997 0.993 0.994 0.997 0.997

CB25 0.997 0.998 0.997 0.998 0.997 0.985 0.998 0.997 0.996 0.996 0.997 0.997

CB26 0.999 1.000 1.000 0.999 1.000 0.987 1.000 0.999 0.999 1.000 1.000 0.999

CB27 0.997 0.997 0.994 0.996 0.995 0.983 0.997 0.997 0.995 0.983 0.996 0.996

CB28 0.997 0.997 0.997 0.996 0.995 0.984 0.996 0.995 0.995 0.996 0.998 0.998

CB29 0.999 0.999 0.999 1.000 0.997 0.984 0.993 0.994 0.992 0.993 0.996 0.996

Table 6. Normalized values of aggregate resilience scores (Q

). (Source: Estimated by the authors)

years. This pattern highlights the ability of banks to adapt and

regain stability after adverse conditions.

After this period, it indicates a gradual recovery back to

normal range, which shows that banks have returned to a

more stable situation in later years. Interestingly, while most

other banks’ resilience scores fell signiﬁcantly, as seen in Fig. 3,

the Punjab & Sind Bank (CB21) had the highest ranking in

terms of resilience in 2018 (Table 7), which was likely due to

a one-oﬀ balance-sheet adjustments. Overall, the RAM rank-

ings highlight meaningful improvements in relative resilience

and systemic stress, rather than dramatic shifts in the banks’

absolute strength.

4.3. Hypotheses Testing

The Mann–Kendall (MK) non-parametric trend test was em-

ployed using Eq. (14) to examine both bank-level and sector-

level dynamics in composite RAM-based ﬁnancial resilience

scores over the period 2013–2024. First, annual RAM scores for

29 banks were structured as a balanced time series, enabling

univariate MK tests for each bank to detect statistically sig-

niﬁcant monotonic trends. The test statistic S, Kendall’s tau

(τ ), and associated p-values were computed for all banks which

allow for classiﬁcation of resilience tra jectories as increasing or

decreasing trend.

The Mann–Kendall trend analysis in Fig. 4 demonstrates

the upward (increasing) and downward (decreasing) trend in

ﬁnancial resilience, as well as banks with no signiﬁcant trend

or (stable over the period) highlighted from 2013 to 2024. The

Figure 4. Bank-sp eciﬁc Mann–Kendall Trend Test results (2013–2024).

(Source: Authors’ Computation)

MEREC–RAM rankings presented in Table 6 capture the rel-

ative resilience levels of banks in a particular year, while the

Mann–Kendall test captures the dynamics of resilience.

The results reveal that the majority of banks do not ex-

hibit statistically signiﬁcant trends in their resilience scores

P. Das and S. K. Sen

2013 2014 2015 2016 2017 2018 2019 2020 2021 2022 2023 2024

CB1 5 7 7 14 18 16 14 10 8 17 8 5

CB2 15 16 22 16 19 19 22 18 17 12 15 16

CB3 24 24 26 25 24 22 23 17 20 23 27 24

CB4 17 17 23 28 22 29 17 13 11 6 3 2

CB5 22 29 19 18 13 8 4 5 1 1 4 14

CB6 18 19 24 19 23 21 20 24 18 18 17 20

CB7 28 26 25 27 26 26 25 22 25 22 22 18

CB8 3 2 1 2 2 4 5 6 7 11 16 11

CB9 9 8 5 8 5 5 7 14 14 14 24 28

CB10 29 28 17 22 10 11 6 27 26 27 26 29

CB11 2 3 8 9 7 7 8 7 10 10 9 12

CB12 6 1 2 1 4 3 1 2 3 4 10 6

CB13 13 15 12 15 16 14 9 3 5 5 5 3

CB14 16 14 14 13 15 15 15 23 15 15 14 9

CB15 26 27 29 29 29 27 27 15 13 21 18 15

CB16 7 9 6 5 6 13 11 8 6 8 7 23

CB17 1 13 15 23 14 18 18 16 24 16 12 13

CB18 14 12 9 10 11 9 13 9 12 9 11 19

CB19 8 5 3 7 8 10 10 12 9 7 6 7

CB20 4 6 11 3 1 2 2 1 2 3 1 1

CB21 25 25 16 20 20 1 28 29 21 24 28 27

CB22 23 22 27 21 28 25 21 25 27 26 23 10

CB23 27 20 18 12 9 12 12 11 19 19 21 21

CB24 12 11 13 11 12 17 19 21 28 25 20 22

CB25 19 18 20 17 21 20 16 19 16 13 19 17

CB26 10 4 4 6 3 6 3 4 4 2 2 4

CB27 20 21 28 26 27 28 24 20 23 29 25 26

CB28 21 23 21 24 25 24 26 26 22 20 13 8

CB29 11 10 10 4 17 23 29 28 29 28 29 25

Table 7. Year-wise MEREC-based RAM rankings. (Source: Estimated by the authors)

over time. This suggests that ﬁnancial resilience remains rel-

atively stable for most institutions with no consistent upward

or downward trajectory.

However, a subset of banks demonstrates statistically signif-

icant trends. Speciﬁcally, banks such as CB4 and CB5 exhibit

signiﬁcant improving trends, indicating gradual improvements

in resilience over the study period. These results suggest that

these banks have strengthened their ﬁnancial position over

time. In contrast, several banks display signiﬁcant decreasing

trends in resilience. Notably, CB8, CB9, CB11, CB12, CB16,

and CB24 show statistically signiﬁcant downward trends, with

CB11 exhibiting the strongest decline (p < 0.01). This indicates

a functional decline in ﬁnancial resilience for these institutions,

which may warrant closer attention from the side of regulatory

bodies.

Banks CB4 and CB5 indicate statistically signiﬁcant up-

ward trends in resilience which align with Financial Resilience

Theory and show signs of improvement in solvency, asset qual-

ity, eﬃciency, and proﬁtability. Capital Buﬀer Theory focuses

on the need to maintain suﬃcient capital and Tier-1 buﬀers

to reduce bank fragility and promote long-term stability. This

reﬂected in the banks with positive trends as well as with

the banks showing stable ﬁnancial resilience throughout the

period. Meanwhile, banks that have experienced signiﬁcant de-

clines that correspond with the theory that weak buﬀers make

banks more vulnerable due to problems with NPAs, inadequate

provisioning, and earnings pressure. CB11 display marginally

signiﬁcant declining trend which reﬂect the lingering eﬀects of

capital erosion, governance challenges, and volatility in prof-

itability. The majority of large public and private sector banks

such as CB25, CB1, CB13, CB22, CB25, and so forth, showing

no statistically signiﬁcant trend implies relative stability over

the period but limited monotonic improvement in resilience.

The sector-level resilience is further evaluated by construct-

ing annual mean RAM scores, to which separate MK tests

are applied. To assess heterogeneity in resilience patterns, the

sample was stratiﬁed based on ownership structure (public vs.

private banks). Both parametric and non-parametric tests were

employed to determine whether resilience trajectories diﬀered

signiﬁcantly across subgroups.

Although the initial bank-level Mann–Kendall analysis pro-

vides signiﬁcant evidence on how some banks experienced

increasing and decreasing trends in their ﬁnancial resilience, the

Mann–Kendall test for sector mean resilience indicates no sta-

tistically signiﬁcant monotonic trend (Z = −1.71, p = 0.086).

However, the negative tau (−0.394) suggests a weak declining

tendency over the study period. Therefore, the null hypothesis

) of no monotonic trend cannot b e rejected at the aggregate

level.

A Sen’s slope estimator is used to quantify the magni-

tude and direction of sector-wide change. The estimated slope

is visualised in Fig. 5. Initially, a Welch two-sample t-test

is conducted on average RAM scores over the study period

(2013–2024), which revealed a statistically signiﬁcant diﬀerence

between the two groups (t = 4.64,p < 0.001). A non-parametric

Kruskal–Wallis test is additionally performed to ensure robust-

ness. The results conﬁrm a statistically signiﬁcant diﬀerence

Figure 5. Ownership-Based Heterogeneity Analysis and Sectoral Trend in Financial Resilience. (Source: Authors’ Computation)

in resilience scores, especially between the public and private

sector banks (χ

= 71.096, p < 0.001). The null hypothe-

sis (H

) of no diﬀerence between public and private banks

is rejected, as Welch two-sample t-test and non-parametric

tests Kruskal–Wallis reveal statistically signiﬁcant diﬀerences

in resilience across ownership groups.

As can be seen in the boxplot (Fig. 5), private banks have

a more resilient recovery from the sharp drop around 2018,

which is a sign of greater systemic resilience. The diﬀerent re-

covery paths of public sector banks (PSBs) and private banks

since 2018 show the impact of diﬀerent governance structures,

support for recapitalization, and operational ﬂexibility. Public

sector banks received signiﬁcant recapitalization and gover-

nance changes through the EASE and Indradhanush programs.

However, they also had to manage a larger amount of legacy

non-performing assets. In contrast, private banks usually had

more managerial freedom and better proﬁt margin which al-

low them to recover and adapt more quickly. These diﬀerences

in structure might caused the varying resilience of each sector

after systemic shocks.

These validate that the observed variations in resilience

are statistically signiﬁcant and not an artifact of the MCDM-

based ranking framework. The results are also highly consistent

with Dynamic Capability Theory, which asserts that ﬁrms

can maintain superior performance by sensing risks, seiz-

ing opportunities, and reconﬁguring resources under changing

environments[53]. The gradual recovery of the Indian banking

sector after the structural change around 2018 also shows a

positive adaptive response through a series of balance-sheet

adjustments, capital augmentation, and operational restructur-

ing.

4.4. Sensitivity Analysis and Robustness Test

To examine the eﬀect of changes in criteria weighting, a

sensitivity analysis is conducted by comparing the baseline

MEREC–RAM results with alternative equal-weighting and

Entropy-based weighting scenarios. The equal-weighting was

implemented by applying 1/11, in which all ﬁnancial resilience

indicators were assigned identical importance 0.0909 (i.e., for

11 criteria), thereby eliminating any data-driven or sub jective

prioritisation among criteria. Corresp ondingly, the method en-

tropy is applied to determine objectives which then followed by

entropy-RAM rankings.

The study applied additional outranking techniques such as

TOPSIS, RATMI, and MARCOS for a robustness test incor-

porating the MEREC weights. The Kendall’s tau coeﬃcient

is employed to assess the degree of rank concordance among

the ranking outcomes obtained from variations in weight-

ing schemes and multiple outranking models. Fig. 6 presents

the year-wise p-values of Kendall’s tau coeﬃcient for rank

agreement in both the sensitivity analysis and robustness tests.

The consistently high Kendall’s tau coeﬃcients across all

years (Fig. 5) indicate that the integrated approach produces

stable and reliable rankings. Entropy and MEREC shows near-

perfect agreement (τ = 0.96 − 1.00), while Equal Weighting

and MEREC ranges from 0.71 to 0.94. This demonstrates that

the results are not driven by a speciﬁc weighting technique,

but rather reﬂect inherent data structure. Under the robustness

test, RAM and TOPSIS (0.79–0.95) show strongest agreement,

followed by RAM-MARCOS (0.67–0.89) and RAM-RATMI

(0.67–0.81). The absence of signiﬁcant rank reversals across

methods and time reinforces the robustness of the framework.

5. Conclusions

Assessing banks’ ﬁnancial resilience is vital in emerging

economies, where ﬁnancial systems are characterized by volatil-

ity, governance constraints, and heightened exposure to NPAs.

Thus, in order to assess the ﬁnancial resilience Indian com-

mercial banks between 2013 and 2024, this study developed

an integrated analytical framework that combines the MEREC

objective weighting approach, the RAM-based composite re-

silience index, , and Mann-Kendall trend analysis.

The results showed that the RAM-based resilience rankings

are narrowly clustered among banks, indicating the eﬀect of

structural convergence brought about by regulatory standard-

ization. The statistical signiﬁcance tests Kruskal–Wallis verify

that the observed diﬀerences are meaningful and signiﬁcant.

However, the trend analysis shows that resilience trajecto-

ries at the bank level vary signiﬁcantly. A limited number

of banks have statistically signiﬁcant monotonic trends, but

cross-sectional rankings show year-to-year variability for most

banks. The lack of a noticeable trend at the sector-level indi-

cates that the Indian banking sector resilience develops through

periodic adjustments rather than steady directional improve-

ment. Moreover, while Private sector banks showed greater

adaptive capacity and quicker recovery from systemic shocks,

public banks were more sensitive to regulatory interventions.

The MEREC-RAM ranking result showed Kotak Mahindra

Bank Ltd. and Tamilnad Mercantile Bank Ltd. consistently

showed strong resilience across the study period. Meanwhile the

CSB Bank Ltd. reﬂected signiﬁcant improvement in rankings

P. Das and S. K. Sen

Figure 6. Heatmaps on Kendall’s tau (τ ) across methods. (Source: Authors’ Computation)

from bottom to top ten over the years. Bank of Maharash-

tra also demonstrated signiﬁcant improvement in rankings by

reaching to the second place during 2024. Conversely, Bank

of India, Central Bank of India and Dhanlaxmi Bank Ltd.,

UCO Bank constantly positioned at the bottom in the rankings.

Interestingly, the Mann-kendall trend test showed signiﬁcant

improvement in the resilience of Bank of Maharashtra and CSB

Bank Ltd. during the study period. Whereas, City Union Bank

Ltd., DCB Bank Ltd., Federal Bank Ltd., HDFC Bank Ltd.,

Indusind Bank Ltd., and South Indian Bank Ltd. reﬂected

signiﬁcant decline in their ﬁnancial resilience. However, other

banks demonstrated stable or no signiﬁcant improvement or

deterioration.

This study contributes to Financial Resilience Theory in

showing that resilience is non-linear, heterogeneous, and con-

vergent rather than monotonic. These ﬁndings are also in line

with Institutional Adaptation Theory, which contends that

resilience outcomes depend on banks’ capacity to adapt to shift-

ing macroeconomic and regulatory conditions. The results are

consistent with a convergence-diﬀerentiation approach to bank-

ing resilience, in which statistically signiﬁcant performance

diﬀerences resulting from diﬀerences in banks’ capacity to

adapt, manage risks, and react to changing conditions are con-

trasted with structural similarity brought about by regulatory

frameworks.

The Mann-Kendall analysis shows that banks exhibit dis-

tinct resilience tra jectories, which have signiﬁcant managerial

and policy signiﬁcance. From a managerial perspective, the

ﬁndings suggest that persistent adaptive capability is neces-

sary in addition to maintaining high relative rankings. To

achieve long-term resilience, banks must place the utmost im-

portance on proactive risk management, ongoing balance-sheet

reconﬁguration, and operational eﬀectiveness. Banks with in-

creasing trends demonstrate great adaptive capacity and can

serve as benchmarks for best practices, with a continuous em-

phasis on risk management and capital strengthening. Banks

showing deteriorating trends (e.g., CB8, CB9, CB11, CB12,

CB16, and CB24) should prioritize strengthening asset qual-

ity through stricter credit appraisal, early warning systems,

and faster resolution of non-performing assets. Management

should improve capital adequacy, liquidity buﬀers, and risk gov-

ernance while adopting data-driven monitoring of proﬁtability

and operational eﬃciency. Diversiﬁcation of loan portfolios and

digital transformation can reduce concentration and op erating

risks. Regulators may encourage periodic stress testing, en-

hanced disclosure standards, and corrective action frameworks

for vulnerable banks. From a policy perspective, the lack of a

broad sector-wide trend indicates that while regulatory changes

have stabilized the system, they have not consistently increased

resilience.

Although the study makes important contributions, it is

limited by its analysis to ﬁnancial indicators without incorpo-

rating macroeconomic or market-based variables. Although the

Mann–Kendall test captures monotonic trends but it unable to

conduct a complete structural break or regime shifts. Further

research could expand on this framework by including macro-

ﬁnancial variables, structural break or panel-based approaches,

and identifying causal links between regulatory interventions,

adaptive capacities, and banking system resilience.

Ethical Statement

No ethical approval was required for this study, as it did not in-

volve human or animal subjects. The study relied exclusively on

secondary data obtained from publicly available and authorized

databases and annual reports.

Funding

This research received no speciﬁc grant from any funding

agency in the public, commercial, or not-for-proﬁt sectors.

Declaration of competing interests

The authors declare that they have no known competing ﬁnan-

cial interests or personal relationships that could have appeared

to inﬂuence the work reported in this paper.

Data Availability Statements

The data supporting the ﬁndings of this study are available

from the corresp onding author upon reasonable request. How-

ever, the data are not publicly available due to access restric-

tions associated with the Prowess-IQ database and institutional

data usage policies. Data were obtained from authorized sources

including Prowess-IQ (https://prowess.cmie.com/) and annual

reports of banks.

Credit authorship contribution statement

Priya Das: Conceptualization; Methodology; Investigation;

Data Curation; Formal Analysis; Software; Visualization; Writ-

ing – Original Draft Preparation. Subir Kumar Sen: Concep-

tualization; Supervision; Pro ject Administration; Validation;

Visualization; Resources; Writing – Review & Editing.

References

1. F. Salignac, A. Marjolin, R. Reeve, and K. Muir,

“Conceptualizing and measuring ﬁnancial resilience,”

Social Indicators Research, vol. 145, no. 1, pp. 17–38,

2019, doi: https://doi.org/10.1007/s11205-019- 02100-4.

2. F. Salignac, J. Hanoteau, and I. Ramia, “Financial

resilience: a way forward towards economic development

in developing countries,” Social Indicators Research, vol.

160, no. 1, pp. 1–33, 2022, doi: https://doi.org/10.1007/

s11205-021- 02793-6.

3. S. K. Nkundabanyanga, E. Mugumya, I. Nalukenge,

M. Muhwezi, and G. M. Najjemba, “Firm character-

istics, innovation, ﬁnancial resilience and survival of

ﬁnancial institutions,” Journal of Accounting in Emerging

Economies, vol. 10, no. 1, pp. 48–73, 2020, doi:

https://doi.org/10.1108/JAEE-08-2018-0094.

4. F. S. Hamid, Y. J. Loke, and P. N. Chin, “Determinants of

ﬁnancial resilience: insights from an emerging economy,”

Journal of social and economic development, vol. 25,

no. 2, pp. 479–499, 2023, doi: https://doi.org/10.1007/

s40847-023- 00239-y.

5. A. A. Tinta, I. M. Ou´edraogo, and R. M. Al-

Hassan, “The micro determinants of ﬁnancial inclusion

and ﬁnancial resilience in africa,” African Development

Review, vol. 34, no. 2, pp. 293–306, 2022, doi:

https://doi.org/10.1111/1467-8268.12636.

6. A. Das and S. Ghosh, “Financial deregulation and proﬁt

eﬃciency: A nonparametric analysis of indian banks,”

Journal of Economics and Business, vol. 61, no. 6,

pp. 509–528, 2009, doi: https://doi.org/10.1016/j.jeconbus.

2009.07.003.

7. S. Kumar and R. Gulati, “Did the global ﬁnancial crisis

alter the competitive conditions in the indian banking

industry?” Applied Economics Letters, vol. 26, no. 10,

pp. 857–865, 2019, doi: https://doi.org/10.1080/13504851.

2018.1502865.

8. N. Mohan and M. Irfan, “Applying ai & topsis-

mcdm tool in evaluating top ﬁve private indian bank

performances,” in Applications of Block Chain technology

and Artiﬁcial Intelligence: Lead-ins in Banking, Finance,

and Capital Market. Springer, 2024, pp. 291–303, doi:

https://doi.org/10.1007/978-3-031-47324- 1

15.

9. G. Caprio, Banking on crises: expensive lessons from re-

cent ﬁnancial crises. World Bank, Development Research

Group, Finance, 1998, no. 1979.

10. A. Bhattacharyya, C. K. Lovell, and P. Sahay, “The impact

of liberalization on the productive eﬃciency of indian

commercial banks,” European Journal of operational

research, vol. 98, no. 2, pp. 332–345, 1997, doi:

https://doi.org/10.1016/S0377-2217(96)00351-7.

11. S. Akhtar, S. N. Azmi, P. A. Khan, A. A. Jan,

and Z. Ansari, “Unveiling the ﬁnancial landscape:

analyzing proﬁtability, productivity, and eﬃciency of

banks in an emerging economy using the camels framework

and panel analysis,” Cogent Business & Management,

vol. 11, no. 1, p. 2399747, 2024, doi: https:

//doi.org/10.1080/23311975.2024.2399747.

12. A. Sharma, “Incomplete reform or opportunity: the

role of the banking sector in the credit transmission

mechanism in india,” Journal of Economic Policy

Reform, vol. 11, no. 4, pp. 273–288, 2008, doi:

https://doi.org/10.1080/17487870802567315.

13. B. Eichengreen and P. Gupta, “The ﬁnancial crisis

and indian banks: survival of the ﬁttest?” Journal of

International Money and Finance, vol. 39, pp. 138–152,

2013, doi: https://doi.org/10.1016/j.jimonﬁn.2013.06.022.

14. H. Fujii, S. Managi, and R. Matousek, “Indian bank

eﬃciency and productivity changes with undesirable

outputs: A disaggregated approach,” Journal of banking

& ﬁnance, vol. 38, pp. 41–50, 2014, doi: https:

//doi.org/10.1016/j.jbankﬁn.2013.09.022.

15. K. Srinivasan, K. Ramesh, K. Gunasekaran, and

K. Sivasubramanian, “Reforms in indian banking sector: a

paradigm shift in growth and ﬁnancial inclusion in india,”

in Technology-Driven Business Innovation: Unleashing

the Digital Advantage, Volume 1. Springer, 2024, pp. 433–

439, doi: https://doi.org/10.1007/978-3- 031-51997-0

36.

16. N. Shikha and I. Kapsis, “Bank crisis management

and resolution after svb and credit suisse: Perspectives

from india and the european union,” International

Insolvency Review, vol. 33, no. 1, pp. 55–88, 2024, doi:

https://doi.org/10.1002/iir.1516.

17. A. Nasim, G. Downing, and M. A. Nasir, “The role

of uncertainty, regulatory and economic environment

and quantitative tightening in banks’ performance,”

International Journal of Finance & Economics, vol. 31,

no. 1, pp. 46–69, 2026, doi: https://doi.org/10.1002/ijfe.

3128.

18. K. D. Hanumantu, V. Worlikar, and S. Narayanaswami,

“The punjab national bank scam: Ethics versus robust

processes,” Journal of Public Aﬀairs, vol. 19, no. 4, p.

e1952, 2019, doi: https://doi.org/10.1002/pa.1952.

19. J. Truby, R. Brown, and A. Dahdal, “Banking on

ai: mandating a proactive approach to ai regulation

in the ﬁnancial sector,” Law and Financial Markets

Review, vol. 14, no. 2, pp. 110–120, 2020, doi:

https://doi.org/10.1080/17521440.2020.1760454.

20. A. Sotoudeh-Anvari, “Root assessment method (ram):

A novel multi-criteria decision making method and its

applications in sustainability challenges,” Journal of

Cleaner Production, vol. 423, p. 138695, 2023, doi:

https://doi.org/10.1016/j.jclepro.2023.138695.

21. ——, “Root assessment method (ram) under neutrosophic

environment,” in Neutrosophic Paradigms: Advancements

in Decision Making and Statistical Analysis: Neutrosophic

Principles for Handling Uncertainty. Springer, 2025, pp.

123–138, doi: https://doi.org/10.1007/978-3- 031-78505- 4

22. M. Keshavarz-Ghorabaee, M. Amiri, E. K. Zavadskas,

Z. Turskis, and J. Antucheviciene, “Determination of

objective weights using a new method based on the removal

eﬀects of criteria (merec),” Symmetry, vol. 13, no. 4, p.

525, 2021, doi: https://doi.org/10.3390/sym13040525.

P. Das and S. K. Sen

23. M. S. Tahir and D. W. Richards, “A systematic

literature review of ﬁnancial resilience: antecedents,

consequences and future research agenda,” Journal

of Financial Regulation and Compliance, 2025, doi:

https://doi.org/10.1108/JFRC-10-2024- 0204.

24. L. Klapper and A. Lusardi, “Financial literacy and

ﬁnancial resilience: Evidence from around the world,”

Financial Management, vol. 49, no. 3, pp. 589–614, 2020,

doi: https://doi.org/10.1111/ﬁma.12283.

25. A. Lusardi, A. Hasler, and P. J. Yakoboski, “Building

up ﬁnancial literacy and ﬁnancial resilience,” Mind &

Society, vol. 20, no. 2, pp. 181–187, 2021, doi:

https://doi.org/10.1007/s11299-020-00246-0.

26. C. Jansson, “Financial resilience: the role of ﬁ-

nancial balance, proﬁtability, and ownership,” in

The Resilience Framework: Organizing for Sus-

tained Viability. Springer, 2017, pp. 111–131, doi:

https://doi.org/10.1007/978-981-10-5314- 6

27. A. Sreenivasan and M. Suresh, “Readiness of ﬁnancial

resilience in start-ups,” Journal of Safety Science and

Resilience, vol. 4, no. 3, pp. 241–252, 2023, doi:

https://doi.org/10.1016/j.jnlssr.2023.02.004.

28. R. L. Clark and O. S. Mitchell, “Americans’ ﬁnancial

resilience during the pandemic,” Financial Planning

Review, vol. 5, no. 2-3, p. e1140, 2022, doi:

https://doi.org/10.1002/cfp2.1140.

29. G. M. Markman and M. Venzin, “Resilience: Lessons

from banks that have braved the economic crisis—and

from those that have not,” International Business

Review, vol. 23, no. 6, pp. 1096–1107, 2014, doi:

https://doi.org/10.1016/j.ibusrev.2014.06.013.

30. G. O. Danisman, E. Demir, and A. Zaremba, “Financial

resilience to the covid-19 pandemic: The role of banking

market structure,” Applied Economics, vol. 53, no. 39, pp.

4481–4504, 2021, doi: https://doi.org/10.1080/00036846.

2021.1904118.

31. T. Papadimitriou, P. Gogas, and A. Agrapetidou, “The

resilience of the us banking system,” International Journal

of Finance & Economics, vol. 27, no. 3, pp. 2819–2835,

2022, doi: https://doi.org/10.1002/ijfe.2300.

32. C. Cant´u, R. Lobato, C. L´opez, and F. L´opez-Gallo,

“A loan-level analysis of ﬁnancial resilience in mexico,”

Journal of Banking & Finance, vol. 135, p. 105951, 2022,

doi: https://doi.org/10.1016/j.jbankﬁn.2020.105951.

33. S. Alam, S. K. Das, U. R. Dipa, and S. Z. Hossain,

“Predicting ﬁnancial distress through ownership pattern:

dynamics of ﬁnancial resilience of bangladesh,” Future

Business Journal, vol. 10, no. 1, p. 91, 2024, doi:

https://doi.org/10.1186/s43093-024-00379-5.

34. E. Daadmehr, “Workplace sustainability or ﬁnancial

resilience? composite-ﬁnancial resilience index,” Risk

management, vol. 26, no. 2, p. 7, 2024, doi:

https://doi.org/10.1057/s41283-023-00139-9.

35. Y. Chen and C. Sun, “A new method for measuring ﬁnancial

resilience,” Economics Letters, vol. 242, p. 111883, 2024,

doi: https://doi.org/10.1016/j.econlet.2024.111883.

36. E. K. Zavadskas, Z. Turskis, J. Antucheviciene, and

A. Zakarevicius, “Optimization of weighted aggregated

sum pro duct assessment,” Elektronika ir elektrotechnika,

vol. 122, no. 6, pp. 3–6, 2012, doi: https://doi.org/10.

5755/j01.eee.122.6.1810.

37. S. Gupta, M. Mathew, S. Gupta, and V. Dawar,

“Benchmarking the private sector banks in india using

mcdm approach,” Journal of Public Aﬀairs, vol. 21, no. 2,

p. e2409, 2021, doi: https://doi.org/10.1002/pa.2409.

38. S. Gupta, M. Mathew, G. Syal, and J. Jain, “A hybrid

mcdm approach for evaluating the ﬁnancial performance

of public sector banks in india,” International Journal of

Business Excellence, vol. 24, no. 4, pp. 481–501, 2021, doi:

https://doi.org/10.1504/IJBEX.2021.117648.

39. P.-H. Nguyen, J.-F. Tsai, Y.-C. Hu, and G. V.

Ajay Kumar, “A hybrid method of mcdm for evaluating

ﬁnancial performance of vietnamese commercial banks

under covid-19 impacts,” in Shifting economic, ﬁnancial

and banking paradigm: New systems to encounter

COVID-19. Springer, 2021, pp. 23–45, doi: https:

//doi.org/10.1007/978-3- 030-79610- 5

40. A. Karbassi Yazdi, C. Spulbar, T. Hanne, and R. Birau,

“Ranking performance indicators related to banking

by using hybrid multicriteria methods in an uncertain

environment: A case study for iran under covid-19

conditions,” Systems Science & Control Engineering,

vol. 10, no. 1, pp. 166–180, 2022, doi: https:

//doi.org/10.1080/21642583.2022.2052996.

41. U.

Unl¨u, N. Yal¸cın, and N. Av¸sarlıgil, “Analysis

of eﬃciency and productivity of commercial banks in

turkey pre-and during covid-19 with an integrated mcdm

approach,” Mathematics, vol. 10, no. 13, p. 2300, 2022,

doi: https://doi.org/10.3390/math10132300.

42. Y. A.

Unvan and C. Ergen¸c, “Financial p erformance

analysis with the fuzzy copras and entropy-copras

approaches,” Computational Economics, vol. 59, no. 4,

pp. 1577–1605, 2022, doi: https://doi.org/10.1007/

s10614-021- 10143-4.

43. P. Wanke, M. A. K. Azad, A. K. Yazdi, F. R. Birau,

and C. M. Spulbar, “Revisiting camels rating system

and the performance of asean banks: a comprehensive

mcdm/z-numbers approach,” IEEE Access, vol. 10,

pp. 54 098–54 109, 2022, doi: https://doi.org/10.1109/

ACCESS.2022.3171339.

44. H. R. Sama, S. V. K. Kosuri, and S. Kalvakolanu,

“Evaluating and ranking the indian private sector banks—a

multi-criteria decision-making approach,” Journal of

Public Aﬀairs, vol. 22, no. 2, p. e2419, 2022, doi:

https://doi.org/10.1002/pa.2419.

45. P. Kumar and D. Sharma, “Prioritising the ﬁnancial

performance of indian private sector banks by a hybrid

mcdm approach,” International Journal of Process

Management and Benchmarking, vol. 16, no. 4, pp.

490–511, 2024, doi: https://doi.org/10.1504/IJPMB.2024.

137145.

46.

O. Karada˘g Ak, A. Hazar, and S¸ . Babu¸scu, “Evaluation of

the ﬁnancial performance of development and investment

banks with entropy-based aras method,” Macroeconomics

and Finance in Emerging Market Economies, vol. 18,

no. 2, pp. 441–461, 2025, doi: https://doi.org/10.1080/

17520843.2022.2035523.

47.

O. I¸sık, M. Shabir, G. Demir, A. Puska, and D. Pamucar,

“A hybrid framework for assessing pakistani commercial

bank performance using multi-criteria decision-making,”

Financial Innovation, vol. 11, no. 1, p. 38, 2025, doi:

https://doi.org/10.1186/s40854-024-00728-x.

48. J. Zahedi, M. Salehi, and M. Moradi, “Identifying and

classifying the contributing factors to ﬁnancial resilience,”

foresight, vol. 24, no. 2, pp. 177–194, 2022, doi:

https://doi.org/10.1108/FS-10-2020-0102.

49. Z. Liu, J.-K. Chen, and J. J. Xiao, “Financial

resilience: a scoping review, conceptual synthesis and

theoretical framework,” International Journal of Bank

Marketing, vol. 43, no. 7, pp. 1541–1576, 2025, doi:

https://doi.org/10.1108/IJBM-12-2024-0735.

50. T. Jokipii and A. Milne, “Bank capital buﬀer and risk

adjustment decisions,” Journal of Financial Stability,

vol. 7, no. 3, pp. 165–178, 2011, doi: https:

//doi.org/10.1016/j.jfs.2010.02.002.

51. D. Corbae and P. D’Erasmo, “Capital buﬀers in

a quantitative model of banking industry dynamics,”

Econometrica, vol. 89, no. 6, pp. 2975–3023, 2021, doi:

https://doi.org/10.3982/ECTA16930.

52. C. D. Carroll, R. E. Hall, and S. P. Zeldes, “The buﬀer-

stock theory of saving: Some macroeconomic evidence,”

Brookings papers on economic activity, vol. 1992, no. 2,

pp. 61–156, 1992, doi: https://doi.org/10.2307/2534582.

53. D. J. Teece, G. Pisano, and A. Shuen, “Dynamic

capabilities and strategic management,” Strategic man-

agement journal, vol. 18, no. 7, pp. 509–533, 1997, doi:

https://doi.org/10.1002/(SICI)1097-0266(199708).

54. M. M. H. Chowdhury and M. Quaddus, “Supply chain

resilience: conceptualization and scale development using

dynamic capability theory,” International journal of

production economics, vol. 188, pp. 185–204, 2017, doi:

https://doi.org/10.1016/j.ijpe.2017.03.020.

55. M. Doumpos and C. Zopounidis, “A multicriteria decision

support system for bank rating,” Decision support

systems, vol. 50, no. 1, pp. 55–63, 2010, doi:

https://doi.org/10.1016/j.dss.2010.07.002.

56. R. Gineviˇcius and A. Podviezko, “The evaluation of

ﬁnancial stability and soundness of lithuanian banks,”

Economic research-Ekonomska istraˇzivanja, vol. 26, no. 2,

pp. 191–208, 2013, doi: https://doi.org/10.1080/1331677X.

2013.11517616.

57. B. Gavurova, J. Belas, K. Kocisova, and T. Kliestik, “Com-

parison of selected methods for performance evaluation of

czech and slovak commercial banks,” Journal of Business

Economics and Management, vol. 18, no. 5, pp. 852–876,

2017, doi: https://doi.org/10.3846/16111699.2017.1371637.

58. M. Dash, “A model for bank p erformance measurement

integrating multivariate factor structure with multi-criteria

promethee methodology,” Asian Journal of Finance &

Accounting, vol. 9, no. 1, pp. 310–332, 2017, doi:

https://doi.org/10.5296/ajfa.v9i1.11073.

59. C. Ayadurai and R. Eskandari, “Bank soundness: a

pls-sem approach,” in Partial least squares structural

equation modeling: Recent advances in banking and

ﬁnance. Springer, 2018, pp. 31–52, doi: https:

//doi.org/10.1007/978-3- 319-71691- 6

60. I. Marjanovi´c and

Z. Popovi´c, “Mcdm approach for

assessment of ﬁnancial performance of serbian banks,”

in Business performance and ﬁnancial institutions in

Europe: Business models and value creation across

European industries. Springer, 2020, pp. 71–90, doi:

https://doi.org/10.1007/978-3-030-57517- 5

61. J. Reig-Mullor, J. M. Brotons-Martinez, and M. E.

Sansalvador-Selles, “A novel approach to improve the bank

ranking process: an empirical study in spain,” Journal of

Intelligent & Fuzzy Systems, vol. 38, no. 5, pp. 5323–5331,

2020, doi: https://doi.org/10.3233/JIFS-179626.

62. M. A. Bushashe, “Determinants of private banks p er-

formance in ethiopia: A partial least square structural

equation model analysis (pls-sem),” Cogent Business &

Management, vol. 10, no. 1, p. 2174246, 2023, doi:

https://doi.org/10.1080/23311975.2023.2174246.

63. V. Sharma, M. Gupta, and K. Jangir, “Exploring the

impact of risk factors on proﬁtability in commercial

banking in india: a pls-sem analysis approach,” 2024, doi:

https://doi.org/10.1108/978-1-83753-734- 120241007.

64. S. S. Goswami, S. K. Mohanty, and D. K. Behera,

“Selection of a green renewable energy source in india with

the help of merec integrated piv mcdm tool,” Materials

today: proceedings, vol. 52, pp. 1153–1160, 2022, doi:

https://doi.org/10.1016/j.matpr.2021.11.019.

65. S. Kumar, P. Ahijith Kumar, K. Bharati, L. Patnaik,

S. Ranjan Maity, and M. Lepicka, “Coating material

selection for bulk metal forming dies: A merec-integrated

approach with multiple mcdm methods,” International

Journal on Interactive Design and Manufacturing (IJI-

DeM), vol. 19, no. 6, pp. 4055–4070, 2025, doi:

https://doi.org/10.1007/s12008-024-01983-z.

66. I. M. Hezam, A. M. Ali, K. Sallam, I. A. Hameed,

A. Foul, and M. Abdel-Basset, “An extension of root

assessment method (ram) under spherical fuzzy framework

for optimal selection of electricity pro duction technologies

toward sustainability: a case study,” International Journal

of Energy Research, vol. 2024, no. 1, p. 7985867, 2024,

doi: https://doi.org/10.1155/2024/7985867.

67. T. Van Dua, D. D. Trung, N. T. P. Giang, and

D. Van Duc, “Comparison of two methods: Ram and

aroman,” in International Conference on Sustainability

and Emerging Technologies for Smart Manufacturing.

Springer, 2024, pp. 727–735, doi: https://doi.org/10.1007/

978-981- 97-7083- 0 73.

68. A. Shekhovtsov, M. Gandorc, R. Pawlakb, and

W. Sa labuna, “A novel rancom-ram-based framework

for city assessment based on cost of living,” Procedia

Computer Science, vol. 270, pp. 5776–5786, 2025, doi:

https://doi.org/10.1016/j.procs.2025.10.046.

Multidimensional Arabic Commonsense Benchmark

BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/byjwsd23

Research Article

RESEARCH ARTICLE

JAMAL: A Multidimensional Benchmark for Arabic

Commonsense Reasoning Across Life-Domains and

Cognitive Axes

Basma Sayah ,

1,∗

Attia Nehar,

1,2

Hadda Cherroun,

Slimane Bellaouar

and Firoj Alam

Laboratoire d’Informatique et de Mathématiques (LIM), Amar Telidji University, Laghouat, Algeria ,

Computer Science Department,

Ziane Achour University of Djelfa, Djelfa, Algeria ,

Laboratoire de Mathématiques et Sciences Appliquées (LMSA), Université de

Ghardaia, Ghardaia, Algeria and

Qatar Computing Research Institute (QCRI), Hamad Bin Khalifa University, Doha, Qatar

∗

Corresponding author: b.sayah@lagh-univ.dz

Received on 10 January 2026; Accepted on 21 May 2026

Graphical Abstract

Highlights

• JAMAL, a novel language-agnostic commonsense

framework instantiated as a concrete Arabic bench-

mark.

• Establishes a functional dimension covering 56 life-

domain categories.

• Deﬁnes a cognitive dimension encompassing everyday

situations, general knowledge, and problem-solving.

• Introduces a cultural grounding categorization of

items into universal, western/global, and Arabic-

speciﬁc knowledge for controlled analysis.

• Conducts a ﬁne-grained evaluation of ﬁve language

models across the three axes, revealing behavioral er-

ror patterns.

• Publicly releases the dataset to foster Arabic NLP

development and support future multilingual exten-

sions.

Abstract

Commonsense is a broad and multifaceted concept, making its evaluation a persistent challenge in natural language

processing (NLP). This paper introduces JAMAL (Arabic for “Camel”), a multidimensional framework and benchmark

for Arabic commonsense reasoning. JAMAL is structured along three complementary axes: (i) a life-domain axis com-

prising a taxonomy of 56 functional categories informed by the World Health Organization’s International Classiﬁcation

of Functioning, Disability, and Health (ICF), capturing diverse aspects of daily human experience; (ii) a cognitive axis

organizing commonsense into three reasoning types: everyday situations, general knowledge, and problem-solving; and

(iii) a cultural grounding axis distinguishing between universal, western/global, and Arabic-speciﬁc knowledge. To oper-

ationalize this framework, benchmark items are constructed using psycholinguistically inspired principles of constrained

contextual prediction. We evaluate ﬁve Arabic language models using JAMAL and observe consistent diﬀerences in their

performance across all axes. Notably, FANAR-27B achieves the strongest overall results among all evaluated models,

outperforming FANAR-9B and smaller baselines. Overall, JAMAL provides a structured and interpretable benchmark

for evaluating commonsense reasoning in Arabic, supporting the development of more robust language models through

systematic analysis of their behavioral limitations.

Key words: Commonsense Reasoning, Arabic NLP, Language Model Evaluation, Psycholinguistically Grounded

Benchmarking, WHO-ICF Framework

1. Introduction

Neural language models have made remarkable strides in re-

cent years, achieving strong performance across a wide range of

NLP tasks, including text generation, sentiment analysis, and

machine translation [1]. In particular, large-scale transformer-

based models such as GPT, Llama, Qwen, and Gemini can

generate ﬂuent and coherent text [2–5].

As these models grow in capability and complexity, rigorous

evaluation becomes increasingly important. Standard bench-

mark suites probe abilities such as commonsense reasoning,

general knowledge, and reading comprehension, including Hel-

laSWAG [6], MMLU [7], and RACE [8]. However, most of these

benchmarks are designed primarily for English, limiting their

direct applicability to other languages without adaptation.

In Arabic NLP, evaluation has advanced in recent years

through general benchmarking eﬀorts [9, 10], dedicated re-

sources such as ArabicMMLU [11] and AraDiCE [12], and

leaderboards such as BALSAM [13]. However, benchmarks

targeting other speciﬁc capabilities remain limited [14].

Among these capabilities, commonsense reasoning is partic-

ularly important, as it underpins eﬀective understanding and

interaction with the world [15]. Despite its importance, exist-

ing Arabic commonsense evaluation resources remain limited in

both scope and granularity. Early eﬀorts, such as the translated

Is This Sentence Valid? dataset [16], provide only coarse, task-

level evaluation and do not capture the multifaceted nature of

commonsense reasoning.

At the same time, the emergence of Arabic large language

models—including Jais [17], ALLaM [18], and Fanar [19] has in-

creased the need for more structured and diagnostic evaluation

frameworks. These models diﬀer in training data composition,

dialectal coverage, and intended use cases, and their capabilities

are not fully reﬂected by existing general-purpose benchmarks.

This motivates the need for structured evaluation frameworks

that provide deeper insight into model behavior.

To address these challenges, we introduce JAMAL

, a mul-

tidimensional framework and Arabic benchmark for evaluating

commonsense reasoning in language models.

JAMAL is structured along three complementary axes. The

ﬁrst is a life-domain axis comprising 56 functional categories

informed by the World Health Organization’s International

Classiﬁcation of Functioning, Disability, and Health (ICF) [20],

enabling systematic coverage of everyday human experiences.

The second is a cognitive axis that organizes commonsense

into three reasoning types: everyday situations, general knowl-

edge, and problem-solving. The third is a cultural grounding

axis distinguishing between universal, western/global, and

Arabic-speciﬁc knowledge.

Together, these axes enable ﬁne-grained and interpretable

evaluation of model performance across functional domains,

reasoning types, and cultural contexts.

This work makes four contributions: (1) a language-agnostic

multidimensional framework for commonsense evaluation; (2)

an Arabic benchmark grounded in functional, cognitive, and

cultural axes; (3) a comprehensive empirical analysis of model

behavior across these dimensions; and (4) diagnostic insights re-

vealing systematic strengths and weaknesses across ﬁve Arabic

language models.

The name JAMAL is a wordplay: in Arabic, jamal means

“Camel,” a culturally salient symbol, and it is also phoneti-

cally related to jumal (“sentences”), reﬂecting the benchmark’s

sentence-based design.

The remainder of this paper is organized as follows. Sec-

tion 2 reviews related work on commonsense reasoning eval-

uation, covering global benchmarks, multilingual eﬀorts, and

Arabic-speciﬁc resources. Section 3 describes the design and

construction of JAMAL. Section 4 presents the evaluation of

ﬁve Arabic language models using JAMAL. Finally, Section 5

discusses conclusions, limitations, and future directions.

2. Related Work

This section reviews the evolution of global and Arabic com-

monsense evaluation benchmarks to situate the proposed JA-

MAL within the existing literature.

2.1. Global trends in commonsense evaluation

At the global level, commonsense reasoning (CSR) evalua-

tion has evolved through a series of inﬂuential benchmarks,

reﬂecting a shift from performance-oriented assessment to-

ward more structured and diverse methodologies. Early large-

scale datasets, such as SWAG (2018) [21], CommonsenseQA

(2019) [22], and HellaSwag (2019) [6], predominantly adopt

multiple-choice question answering formats, where common-

sense is treated as a single uniﬁed capability and evaluated

using accuracy-based metrics.

In the following years, evaluation was extended to more spe-

cialized reasoning domains. Social IQA (2019) [23] focuses on

social and emotional reasoning, while PIQA (2020) [24] tar-

gets physical reasoning about everyday object use, material

properties, and aﬀordances. This shift was motivated by the

need to better capture diﬀerent facets of commonsense beyond

general-purpose reasoning. However, most benchmarks still rely

on answer selection as the primary evaluation paradigm.

In parallel, multilingual benchmarks were introduced to

study cross-lingual transfer of commonsense knowledge. X-

CODAH (2019) [25] and X-CSQA (2021) [26] extend existing

datasets to multiple languages, typically via translation-based

approaches, enabling evaluation of multilingual generalization.

More recently, research has moved toward more structured

and interpretable evaluation frameworks. TG-CSR (2023) [27]

introduces a theory-driven approach that decomposes com-

monsense reasoning into nine interpretable dimensions, such

as temporal, spatial, and causal reasoning, enabling more

ﬁne-grained analysis of model behavior.

A further trend is the shift toward generative and reasoning-

centric evaluation. For example, ExplaGraphs (2021) [28]

requires models to generate structured explanations to jus-

tify predictions, while emerging frameworks such as SCoRE

(2025) [29] emphasize multi-hop reasoning and reasoning-chain

evaluation. Overall, these developments reﬂect a movement

from static accuracy-based benchmarks toward more inter-

pretable and reasoning-aware evaluation paradigms.

Despite these advances, a tension remains between scala-

bility and interpretability: large-scale benchmarks oﬀer broad

coverage and simple evaluation, while structured approaches

provide deeper insights but are less scalable.

2.2. Commonsense evaluation for Arabic language

Research on Arabic commonsense reasoning has produced a

growing number of benchmarks, establishing important foun-

dations for Arabic-centric evaluation while also revealing chal-

lenges in achieving structured and ﬁne-grained assessment.

Early benchmarks, such as the Arabic Commonsense

Dataset (ArCD) (2019) [30], Is This Sentence Valid?

Multidimensional Arabic Commonsense Benchmark

Benchmark Year Lang. Methodology Taxonomy / Structure Evaluation

SWAG [21] 2018 EN MCQ (sentence comple-

tion)

None Accuracy

CSQA [22] 2019 EN MCQ (knowledge-based) ConceptNet relations Accuracy

HellaSwag [6] 2019 EN Adversarial MCQ None Accuracy

Social IQA [23] 2019 EN MCQ (social reasoning) Social scenarios Accuracy

PIQA [24] 2020 EN MCQ (physical reasoning) Physical interactions Accuracy

X-CODAH [25] 2019 Multi Multilingual MCQ Translated dataset Accuracy

X-CSQA [26] 2021 Multi Multilingual MCQ Translated taxonomy Accuracy

ExplaGraphs [28] 2021 EN Generative explanation Argument graphs NLI + explanation quality

+ human eval

TG-CSR [27] 2023 EN Theory-driven MCQ 9 reasoning dimensions Accuracy per dimension

SCoRE [29] 2025 EN Multi-hop reasoning Scenario-based logic CoT audit + human eval

ArCD [30] 2019 AR MCQ (Wikipedia-based) None Accuracy

Is This Sentence Valid? [16] 2020 AR Sentence classiﬁcation Binary validity Accuracy

Arabic Winograd [31] 2020 AR Coreference resolution Pronoun resolution Accuracy

ArabicSense [32] 2025 AR MCQ + generation Implicit / synthetic Accuracy, F1, BERTScore

Commonsense in Arab Culture [33] 2025 AR MCQ + completion 12-domain taxonomy Accuracy

JAMAL 2026 AR CPARG-based cloze-

style text completion

(psycholinguistically

motivated); hybrid

construction (manual

curation + semi-

automatic generation)

Three-axis structure:

cognitive (3 reasoning

types), life-domain

(56 categories), and

cultural grounding (3

categories)

Accuracy across axes

Table 1. Comparison of existing global and Arabic commonsense reasoning benchmarks across language, methodology, taxonomy, and

evaluation metrics, including the proposed JAMAL benchmark.

(2020) [16], and the Arabic Winograd Dataset (2020) [31],

introduced initial testbeds for Arabic commonsense evalua-

tion. These datasets mainly rely on multiple-choice or clas-

siﬁcation formats and treat commonsense as a general and

undiﬀerentiated capability.

More recent work has attempted to enrich both task design

and evaluation ob jectives. ArabicSense (2025) [32] incorporates

both classiﬁcation and explanation generation, using metrics

such as accuracy, F1, and BERTScore, although it relies heav-

ily on synthetic data. Commonsense Reasoning in Arab Culture

(2025) [33] introduces a culturally grounded taxonomy covering

12 daily life domains and 54 subtopics, combining multiple-

choice and sentence completion tasks to evaluate reasoning

across Arab cultural contexts. While this improves coverage,

the taxonomy is largely derived from corpus-driven topic mod-

eling with manual reﬁnement, rather than an explicit cognitive

or functional framework.

To address these limitations, there is a growing need

for theory-driven benchmarks that implement structured tax-

onomies grounded in established cognitive, psychological, or

behavioural models. Such benchmarks should also ensure care-

ful item construction to avoid synthetic artifacts or superﬁcial

lexical cues, instead capturing meaningful reasoning processes.

Motivated by these limitations, we introduce JAMAL,

a language-agnostic framework designed around a taxonomy

spanning functional, cognitive, and cultural dimensions. Built

through a controlled human-in-the-loop construction process,

JAMAL enables ﬁne-grained evaluation of commonsense rea-

soning across multiple complementary axes, supporting more

systematic analysis of model behavior than existing Arabic

benchmarks. Table 1 summarizes the key characteristics of

global and Arabic benchmarks and situates JAMAL within the

broader evaluation landscape.

3. The JAMAL Benchmark: Design,

Development and Validation

This section describes the design and construction of JAMAL, a

language-agnostic framework for ﬁne-grained, multidimensional

commonsense evaluation, and its Arabic instantiation as a con-

crete benchmark dataset. The overall process is illustrated in

Figure 1.

3.1. Deﬁne the theoretical framework

In this ﬁrst step, we aim to establish a theoretical foundation for

understanding commonsense, in order to evaluate it eﬀectively.

3.1.1. Operationalize commonsense (Cognitive dimension)

Since the goal was to build an eﬀective benchmark for evalu-

ating the commonsense knowledge of language models and to

assess the extent to which this knowledge is acquired during

training, it was essential to establish a clear and precise deﬁ-

nition of commonsense to delineate what constitutes common-

sense and what does not. After reviewing several deﬁnitions,

we adopted one that explicitly distinguishes commonsense from

domain-speciﬁc expertise: “Commonsense is practical good

sense gained through life experience, not through specialized

study” [34].

To operationalize this concept, we identiﬁed key themes

from the literature. Ilievski et al. [35] emphasize the dimension

of everyday situational knowledge, which enables navigation

of routine scenarios. In contrast, Lenat [36] and Whiting

et al. [37] treat commonsense as a body of general factual

knowledge about the world. Complementing these, Newell and

Simon [38] link it to the problem-solving processes used to

address everyday challenges.

Basma Sayah et al.

Figure 1. JAMAL Construction Process: From deﬁning commonsense and its dimensions to benchmark

validation.

Synthesizing these perspectives, we derived three core cat-

egories of commonsense: (i) everyday situations, (ii) general

knowledge, and (iii) problem-solving.

The dimension of everyday situations encompasses the

ability to understand typical scenarios, predict events based

on prior experience [39], make sound judgments, and respond

appropriately to common occurrences [40]. For example, un-

derstanding that people typically stand in a queue to wait for

their turn or that one usually checks if an appliance is plugged

in when it does not turn on. The general knowledge dimen-

sion refers to the factual and conceptual understanding that

individuals typically possess, including basic facts, causal rela-

tionships, and general principles (e.g., “the sky is blue” [36]).

This category also includes knowledge about physical objects

and their properties, such as understanding that an egg consists

of a yolk, egg white, and shell [37]. The problem-solving di-

mension involves applying logical reasoning and cognitive skills

to solve structured tasks and address real-world challenges that

require decision-making and pattern recognition [38]. This in-

cludes scenarios such as prioritizing which tasks to complete

ﬁrst when facing multiple deadlines. Furthermore, this dimen-

sion is shaped by interactions with both the environment and

social contexts [41]. While these three categories capture the

cognitive aspects of commonsense reasoning, its evaluation also

requires grounding it in diverse areas of everyday life. To this

end, we introduce a complementary life-domain dimension,

described in the following subsection.

3.1.2. Deﬁne Life Domains (Functional Dimension)

We adopt the International Classiﬁcation of Functioning,

Disability and Health (ICF) as the conceptual foundation for

deﬁning the functional dimension of JAMAL, leveraging its

comprehensive coverage of human activities, participation, and

environmental context. However, our goal is not to reproduce

the clinical taxonomy, but to derive a simpliﬁed and function-

ally meaningful set of life domains suitable for commonsense

reasoning evaluation.

The taxonomy is derived by clustering related ICF com-

ponents (primarily d- and e-codes) into higher-level functional

domains that reﬂect how humans organize everyday experience.

Rather than enforcing a one-to-one mapping between ICF codes

and items, we aggregate multiple codes when they correspond

to the same commonsense reasoning context.

For example, mobility-related activities (d4) inform Trans-

portation and Movement, interpersonal interactions (d7) in-

form Personal and Social Life, and leisure-related activities

(d920) are distributed across Games, Sports, and Recreation,

Arts, Music, and Creativity, and Fiction and Entertainment.

Environmental factors such as products, nature, and animals

(e1, e2, e245) are integrated into domains describing physical

and ecological context.

This procedure yields 15 top-level life domains covering

social, physical, environmental, cognitive, occupational, and

recreational aspects of human life, further decomposed into 56

subdomains (see Table A).

We introduce three deliberate simpliﬁcations relative to

the original ICF: (i) aggregating multiple activity codes that

correspond to the same everyday commonsense reasoning con-

text into uniﬁed functional domains, (ii) elevating perceptual

attributes such as Colors, Shapes, and Forms from sen-

sory functions to a standalone commonsense domain, and (iii)

separating goal-directed activities (Human Actions and be-

haviours) from institutional contexts (Work and Productivity)

to better reﬂect diﬀerences in reasoning demands.

Overall, the resulting taxonomy preserves the breadth of

the ICF while providing an evaluation-oriented organization of

everyday human experiences for commonsense reasoning tasks.

3.2. JAMAL Design

This design phase translates the theoretical framework into

practical speciﬁcations. We ﬁrst identify the target language

models for evaluation, which informs the selection of suit-

able benchmark item formats. We then select foundational

knowledge sources aligned with the functional and cognitive

dimensions of the framework.

Multidimensional Arabic Commonsense Benchmark

3.2.1. Select target architectures for evaluation

Given the diversity of available architectures, we selected

two main categories of language models for evaluation: base

language models and instruction-tuned models.

Base language models are pre-trained on large, general-

purpose corpora using objectives such as masked or next-token

prediction. They typically require task-speciﬁc ﬁne-tuning to

achieve strong performance, as exempliﬁed by models like

BERT [42] and GPT [43].

instruction-tuned models are large language models that

undergo additional ﬁne-tuning to follow natural language in-

structions. This paradigm, popularized by models such as

ChatGPT [2], enables zero-shot and few-shot task execution

without the need for explicit task-speciﬁc ﬁne-tuning.

3.2.2. Select task and item formats

We adopted the CPARG assessment format, a psycholin-

guistically grounded approach employed in “What BERT Is

Not?” [44], which was originally introduced in human stud-

ies by Federmeier and Kutas [45]. In this format, each item

presents a two-sentence context in which the task is to predict

the ﬁnal word of the second sentence. The task requires using

implicit cues from the ﬁrst sentence to infer this missing word

through commonsense reasoning.

An example of a context in the CPARG format is shown

below:

• The child blew out the candles. Everyone shouted happy

____.

In this example, the target word to be predicted is birthday.

Solving this item relies on commonsense knowledge typically

shared through a familiar social script of blowing out candles

at a birthday celebration.

The CPARG format allows for the evaluation of the internal

commonsense knowledge that language models acquire during

training. Since the target word in each item is not explicitly

stated, models cannot rely on simple text matching or selecting

from predeﬁned options. Instead, they must draw on internal

reasoning and inference to predict the correct word.

To ensure proper alignment with the CPARG format, we

established the following criteria for creating the benchmark

items. In line with CPARG terminology, we refer to these items

as contexts :

• Contexts must reﬂect commonsense rather than specialized

knowledge, following the established deﬁnition 3.1.1.

Example of specialized knowledge (invalid):

Sentence 1: The patient’s echocardiogram revealed severe

mitral valve regurgitation.

Sentence 2: The cardiologist decided the best course of

action was to perform a [blank].

Target word: annuloplasty

• Each context must consist of exactly two sentences.

• The target word to be predicted must not be explicitly

mentioned in the context.

Example of an explicitly mentioned target word in the

ﬁrst sentence (invalid):

Sentence 1: He couldn’t ﬁnd his keys anywhere.

Sentence 2: After searching for an hour, he ﬁnally found

the missing [blank].

Target word: keys

• The dataset must cover a diverse range of commonsense

scenarios representing diﬀerent aspects of daily life.

3.2.3. Select commonsense data sources

To construct commonsense items in the CPARG format, we

required source material that could be reformulated into two-

sentence contexts. We drew upon three main sources: existing

commonsense benchmarks, scenarios inspired by our every-

day experiences, and text generated by ChatGPT-3.5 and

DeepSeek.

This multi-source approach was strategically chosen to lever-

age the unique advantages of each while mitigating their inher-

ent biases. Existing benchmarks provide a validated foundation

but risk inheriting cultural biases and data contamination from

their original creation. Scenarios from daily life introduce cru-

cial realism and a human perspective, yet they are limited in

scalability and can reﬂect the subjectivity of the researchers’

own experiences. Finally, LLM generated text ensures scala-

bility and diversity but may amplify social biases present in

the models’ training data or produce superﬁcially ﬂuent but

conceptually shallow items. By combining these sources and

reﬁning them manually, JAMAL aims to encompass a broad

spectrum of commonsense knowledge, counterbalancing the

weaknesses of any single source to create a more robust and

nuanced evaluation set.

3.3. Develop benchmark items

In this section, we describe the process of creating and curating

items from diverse sources in accordance with the prescribed

CPARG format.

3.3.1. Create Benchmark’s items from existing

commonsense benchmarks

We searched for existing benchmarks using the three categories

of the cognitive dimension of commonsense as keywords, fo-

cusing on assessments from both human and machine studies

that incorporated these categories. For everyday situations,

we identiﬁed the HellaSWAG [6] and Winogrande [46] bench-

marks. HellaSWAG contains short narrative scripts drawn from

video captions. Winogrande focuses on pronoun resolution in

multiple-choice questions, testing comprehension of context and

references. For general knowledge, we selected the “Is This Sen-

tence Valid?” benchmark [16], which evaluates the ability to

determine the factual and logical validity of statements. For

problem-solving, we selected the Cornell Conditional Reasoning

Test [47], a standardized assessment from cognitive psychol-

ogy designed to evaluate logical and conditional reasoning in

humans.Figure 2 illustrates the selected benchmarks.

Common sense

Probl em

Solvi ng

General

Know ledge

Every d ay

Situat ion

HellaSWAG

Winogrande

Cornell Class

Reasoning

Test

Is This

Sentence Valid?

benchmark

Figure 2. Overview of selected commonsense benchmarks categorized into three

core cognitive dimensions: everyday Situations, general Knowledge, and problem

Solving.

Basma Sayah et al.

No. Processing Stage Context  Issue

1 Original He crashed into a brick wall and

fell to the ground. As he lifted his

wrist, someone’s shoe came down

on it.

     

     



Unusable context; no

clear commonsense in-

ference can be derived.

1 Discarded Discarded sentence   Rejection due to

irrelevance.

2 Original The person lifts the violin to their

chin and prepares. They play a

song on the violin.

    

   

The word “violin” can

serve as the inference

target.

2 Rened context Ahmad lifts his instrument to

his chin and gets ready. He

plays a sad song on the violin.

    

   

Removed the target

word from the rst

sentence; added the

emotional cue ”sad”

and the contextual

cue ”instrument”.

3 Original A ery ball throws a person back-

ward. He crashes through a brick

wall and lands on the ground.

   



   

     

Unrealistic scenario;

metaphorical descrip-

tion lacking grounded

commonsense interpre-

tation.

3 Newly crafted His eyes were swollen and

bruised. He got into a ght

yesterday and received a

punch.

   

   

Inspired by phys-

ical harm, we re-

placed the unrealis-

tic scenario with a

grounded everyday

situation involving a

ght and injury.

Table 2. Examples of items manually created or reﬁned from existing commonsense benchmarks.

We translated the selected benchmark sentences into Ara-

bic using a Python script and the Google Statistical Machine

Translation (SMT) API

. Sentences were manually adapted to

the CPARG format, discarding many captions, questions, and

dialogues that did not reference a speciﬁc word or concept.

Three main adaptation cases were observed:

• Unusable sentences: Could not be modiﬁed to CPARG

format as no target word could be inferred from context

(see sentence (1), Table 2).

• Easily adaptable sentences: Required minimal edits,

such as removing the ﬁnal word to designate it as the target

(see sentence (2), Table 2).

• Sentences requiring full rewriting: Retained only the

core idea; new sentences were crafted to ﬁt the CPARG

format (see sentence (3), Table 2).

A total of 600 items were adapted. The process was highly

selective, except for the Cornell Critical Thinking Test, Level

X, whose 65 items were included in full. These items assess rea-

soning and judgment through everyday scenarios and explicitly

aim to evaluate “the ability to apply logical principles to every-

day problems” [47], making them well-suited for commonsense

evaluation.

3.3.2. Create benchmark’s items from daily life experience

To expand the benchmark’s scope and realism, we supple-

mented it with 400 new items inspired by our own everyday

https://cloud.google.com/translate

experiences. Many of these ideas emerged from reviewing exist-

ing benchmarks or from personal observations. Each item was

carefully constructed to conform to the CPARG format, result-

ing in a benchmark of 1, 000 contexts. Examples of contexts

from our daily life experience are shown in table 3:

3.3.3. Benchmark Item Construction using ChatGPT-3.5

and DeepSeek

We used two LLM-based generation strategies to construct

CPARG-format contexts: (i) interactive prompting with

ChatGPT-3.5, and (ii) an automated multi-step pipeline using

the DeepSeek API.

ChatGPT-3.5 interactive generation:

We ﬁrst generated CPARG-format contexts using ChatGPT-3.5

through iterative prompt engineering. We experimented with

diﬀerent prompting strategies, ranging from minimal prompts

(providing only CPARG examples) to more explicit instruc-

tions describing the task, and ﬁnally to detailed step-by-step

prompts encouraging constraint checking before generation. A

representative prompt is:

"Generate 10 Arabic two-sentence contexts: the ﬁrst sen-

tence provides a hint about a food item, and the second

sentence ends with the target item as the ﬁnal word. En-

sure that the ﬁrst sentence uniquely identiﬁes the item and

that the second sentence reveals the answer only in the ﬁnal

word."

This strategy was not fully error-proof, given the inherent

diﬃculty of implicitly eliciting target words without explicit

mention. Common issues included the target word appearing

Multidimensional Arabic Commonsense Benchmark

No. Example (Source) Translation

1 .ﺎﻬﺤﺴﻤﺑ ﺕﺃﺪﺑﻭ ﺵﺎﻤﻗ ﺔﻌﻄﻗﻰﻠﻴﻟ ﺕﺮﻀﺣﺃ .ﺭﺎﺒﻐﻟﺎﺑ ﺓﺎﻄﻐﻣ ﺔﻟﻭﺎﻄﻟﺍﺖﻧﺎﻛ The table was covered in dust. Leila brought a cloth and

started wiping.

2 ﻊﻤﺠﺑ ﺁﺪﺑ ﺎﻤﻬﺘﻬﺟﻭ ﻰﻟﺇ ﺎﻤﻬﻟﻮﺻﻭ ﺭﻮﻓ .ﻢﻴﻴﺨﺘﻟﺍ ﻥﺎﺒﺤﻳ ﻡﺎﺸﻫﻭ ﺭﺩﺎﻘﻟﺍ ﺪﺒﻋﻥﺎﻛ

.ﺐﻄﺤﻟﺍ

Abdelkader and Hisham loved camping. As soon as they

arrived, they started collecting rewood.

3 .ﺭﺰﺠﻟﺍ ﻞﻛﺄﺑ ﺎﻬﻣﺃ ﺎﻬﺘﺤﺼﻧ .ﺮﻈﻨﻟﺍ ﻲﻓﻒﻌﺿ ﻦﻣ ﻲﻧﺎﻌﺗﺔﻤﺴﺑ ﺖﻧﺎﻛ Basma suered from poor eyesight. Her mother advised

her to eat carrots.

Table 3. Examples of CPRAG-style contexts derived from daily-life experiences.

No. Proceesing Stage Context ﻕﺍ Issue

1 Generated I drew a new card and remembered

the rules. There was no room for

mistakes in UNO.

 󰇶󰇉 .ﺍﺍ ﺕﻭ ﺓ  ﺕﺃ

.ﻭﺍ 󰆳󰆎  ﻝ



ﻙ

Too vague, applies to

multiple card games.

1 Rened She looked at the colored cards

and had to nd a way to get

rid of them quickly. There was

no room for error in playing

UNO.

 ﻥﺃ 󰉼󰉡 ﻥ󰁅 ،ﺍ ﺕﺍ 󰆭󰆠ﺇ ﺕ

ﻝ



ﻙ  󰇶󰇉 . 󰉼󰉫 󰊎󰊈 

.ﻭﺍ  󰆳󰆎 

Added gameplay-

specic cues (col-

ored cards and card-

elimination strat-

egy) to better dis-

ambiguate the UNO

context.

2 Generated He to ok his racket in hand, real-

izing that hitting with power and

precision was the key to victory.

The match was all about exchang-

ing hits on the tennis court.

ﺩﻭ ﺓ ﺏﺍ ﻥﺃً 󰁅ﺭ ،ﻩ  ﺃ

ﻝﺩ ﻝ ﺭﻭ ﺓﺍﺭﺍ 󰁅 .ﺯﺍ ﺡ 

.ﺍ   ﺕﺍ

Ambiguous terms:

racket and court; could

refer to other racket

sports.

2 Rened The match was about exchang-

ing the bright green ball with

the racket. Every day, he

would go to practice on the

tennis court.

ﻥﺍ ﺕﺍﺫ ﺓـﺍ ﻝﺩ ﻝ ﺭﻭ ﺓﺍﺭﺍ 󰁅

 ﻡ 󰁀 ﻥ󰁅 .ﺏ ﺍ ﺍ

.ﺍ  󰆳󰆎 ﺏﺭ󰊎󰊈

Added a concrete

visual cue (green

tennis ball) and

habitual practice

context to better

disambiguate the

tennis scenario.

3 Generated When Ahmed’s family prepares a

healthy dinner, they like to add

dark green leafy vegetables rich in

minerals, like spinach.

󰉼󰉂 ، ﺀ ﻭ 󰄨󰄖ﺃ 󰃎󰃆  

ﺀﺍ ﻕﺍﺭﻭﺃ ﺕﺍﺫ ﺕﺍﻭ ﺇ ﻥ

.ﺍ  ،ﻥﺩ ﻭ ﺍﺩ

Fits multiple vegeta-

bles.

3 Rened Ahmed’s family prepared a

meal that Popeye always eats

to become strong. But Ahmed

did not like eating spinach.

ً ﺍﺩ ﻱ 󰀐 ﻭ 󰄨󰄖ﺃ 󰃎󰃆 ﺕﺃ 

ﻝﻭ   ﻥ󰁅 󰄨󰄖ﺃ ـ .ً  

.ﺍ

Added strong cul-

tural cue (Pop-

eye reference) to

uniquely identify

spinach as the in-

tended vegetable.

Table 4. Examples of CPARG-format contexts generated by ChatGPT-3.5 and their subsequent manual reﬁnement.

before the ﬁnal position in the sentence, ambiguous or under-

speciﬁed hints (e.g., “she steamed this green vegetable. . . ”), and

reduced quality when generating multiple examples simultane-

ously. These limitations are consistent with known challenges

in constraint adherence and hallucination in large language

models [48].

All outputs were therefore manually reviewed, and non-

compliant instances were corrected to ensure strict adherence

to the CPARG format. Examples of this reﬁnement process are

shown in Table 4.

DeepSeek API pipeline generation:

To improve scalability and reduce manual interaction, we addi-

tionally used the DeepSeek API with an automated multi-step

generation pipeline. This pipeline decomposed the task into

three stages:

1. Generate target words belonging to functional categories

(see Appendix A).

2. Generate short descriptive hint sentences for each target

word.

Basma Sayah et al.

No. Processing Stage Context ﻕﺍ Issue

1 Generated On the slopes of the mountains, tall

green plants rose, famous for their

sturdy wood. We took souvenir

pictures next to the cedar tree.

ﺀﺍ  ﺕ ﺭﺍ ﻝﺍ ﺡ 

ﺍًﺭ ﺃ .ﺍ 󰉼󰉨 󰉼󰉮 ،ﻡﺍ 󰃎󰃆

.ﺯﺭﺍ ﺓ



 ﺭ󰁅

Could refer to other

types of trees.

1 Rened On the slopes of the Lebanese

mountains, tall green plants

rose, famous for their sturdy

wood. We took souvenir pic-

tures next to the cedar tree.

 ﺕ ﺭﺍ ﻥ ﻝ ﺡ 

ﺃ .ﺍ 󰉼󰉨 󰉼󰉮 ،ﻡﺍ 󰃎󰃆 ﺀﺍ

.ﺯﺭﺍ ﺓ



 ﺭ󰁅 ﺍًﺭ

Added a strong

geographic cue

(Lebanon), since

the cedar tree is a

national symbol of

the country.

2 Generated I wanted to grow plants that

tolerate salinity and give sweet

fruits near the sea. The expert

advised me to plant a palm tree.

ﺍًﺭ ﻭ ﺍ  ﺕ ﺍﺭﺯ ﺕﺩﺭﺃ

ﺍﺭ ﺍ  .ﺍ  ﺏ ﺓ

.ﺍ

Ambiguous plant con-

text, as it allows mul-

tiple interpretations of

salt-tolerant plants.

2 Rened I wanted to produce dates rich

in benets. I have decided to

plant a palm tree.

ﺍﺭﺯ ﺕﺭ .ﺍ ﺍ ﺍ ﺝﺇ ﺕﺩﺭﺃ

.ﺍ

Added “dates” cue

to disambiguate the

target.

3 Generated In the fertile elds, tall plants were

grown with stems full of a sweet-

tasting liquid from which a natural

sweetener is extracted. To produce

natural sugar, The farmers planted

a lot of sugarcane.

󰃎󰃆  ﺕ ﻉﺭ 󰁅 ،ﺍ ﻝﺍ 󰆳󰆎

ﺝ ﻕﺍﺍ    ﻥ ﺕﺍﺫ

ﻉﺭﺯ ،ﺍ ﺍ ﺝ .ﺍ ﺍ 

  ـﺍ ﻥﺍ

CPRAG should rely

on the full context

rather than the second

sentence alone.

3 Rened In the fertile elds, tall plants

were grown with stems full

of a sweet-tasting liquid from

which a natural sweetener

is extracted. The farmers

planted a lot of sugarcane.

󰃎󰃆  ﺕ ﻉﺭ 󰁅 ،ﺍ ﻝﺍ 󰆳󰆎

 ﺝ ﻕﺍﺍ    ﻥ ﺕﺍﺫ

  ـﺍ ﻥﺍ ﻉﺭﺯ .ﺍ ﺍ

Removed the addi-

tional cue (“to pro-

duce natural sugar”)

to ensure that infer-

ence is based on the

full context rather

than the second sen-

tence alone.

Table 5. Examples of CPARG-style contexts generated by the DeepSeek pipeline, together with their subsequent manual reﬁnement.

3. Generate an intermediate sentence that semantically links

the hint sentence (ﬁrst sentence) to the target word, which

appears as the ﬁnal word of the second sentence.

While this structured pipeline improved consistency, manual

validation and correction were still required to ensure full com-

pliance with CPARG constraints. Representative examples of

reﬁned outputs are shown in Table 5.

Overall, combining ChatGPT-3.5 and DeepSeek-based gen-

eration, we constructed a total of 1823 contexts.

It is worth noting that in English translations of the gener-

ated examples, the target word may not always appear in ﬁnal

position due to diﬀerences in syntactic structure. However, this

does not aﬀect the original Arabic instances used in the bench-

mark, where the target word is strictly constrained to appear

as the ﬁnal token in accordance with the CPARG format.

3.4. Systemize and validate

In this ﬁnal phase, we systematize the benchmark by labeling

each item according to its functional and cognitive category

assignments. We then conduct human veriﬁcation and valida-

tion of both the labels and the benchmark items to ensure the

reliability and internal consistency of JAMAL.

3.4.1. Categorize items

To enable ﬁne-grained evaluation, each item in JAMAL was

labeled with one of the 56 life-domain categories from the func-

tional dimension (Appendix A) and with one or more of the

three cognitive branches (everyday situations, general knowl-

edge, and problem-solving). Because the cognitive branches can

overlap, items could receive multiple cognitive lab els when ap-

propriate. All automatically assigned labels were subsequently

veriﬁed and corrected by human annotators.

Appendix B details the ﬁnal distribution of items across cat-

egories. Across its 56 subdomains, JAMAL contains between

17 and 54 items per subcategory. This density aligns with es-

tablished benchmark practices: BIG-bench tasks often contain

10–100 examples per task, CommonsenseQA and HellaSwag in-

clude roughly 15–50 examples per category, Social IQA contains

approximately 15–50 questions per subcategory, and psycholin-

guistic diagnostic suites such as CPARG consist of 102 items

distributed across all categories.

Each item was also annotated for cultural grounding to

enable post-hoc analysis of model performance across diﬀer-

ent types of cultural knowledge. Items were labeled into three

categories: universal (knowledge shared across cultures), west-

ern/global (concepts commonly encountered in mainstream

media), and Arabic-speciﬁc (knowledge grounded in Arabic

cultural contexts, including traditions and local practices), as

detailed in Section 4.3.2.

Multidimensional Arabic Commonsense Benchmark

3.4.2. Verify benchmark through expert review

The veriﬁcation process consisted of two stages. First, one

of the authors reviewed the items to ensure compliance with

the benchmark design and corrected any inconsistencies. Sec-

ond, an external native Arabic speaker independently evaluated

the items and ﬂagged contexts that did not meet the CPARG

requirements.

The reviewer received detailed written instructions (see

Appendix C), which included:

• Carefully examining each context in the benchmark.

• Identifying contexts that violated the established require-

ments.

• Providing justiﬁcation for each ﬂagged context.

• Noting ambiguous or unclear formulations.

During the veriﬁcation process, 83 contexts (4.6%) were

ﬂagged as incorrect and categorized into four error types:

• Target word already mentioned in the context

(37 out of 1823): This was the largest error category,

comprising 44.6%of all errors.

• Multiple possible target words (16 out of 1823): This

category includes cases where more than one target word

could be inferred, comprising 19.3%of all errors.

• Incorrect target word (27 out of 1823): This category

reﬂects mismatches between the context and the intended

target word, comprising 32.5%of all errors.

• Sentence-level errors (3 out of 1823): This category

includes cases of poor phrasing or ambiguity, comprising

3.6% of all errors.

Overall, the veriﬁcation process conﬁrmed the overall qual-

ity of the benchmark, and ﬂagged items were corrected accord-

ingly.

3.4.3. Validate benchmark

After veriﬁcation, we validated JAMAL using a stratiﬁed 25%

sample (455 items) drawn proportionally from all 56 functional

categories. Two native Arabic speakers with diﬀerent aca-

demic backgrounds independently evaluated the items, ensuring

a diverse perspective since commonsense benchmarks target

shared everyday reasoning rather than specialized knowledge.

All reviewers followed the instructions detailed in App endix D,

assessing each item across four criteria: Cloze Predictabil-

ity, Agreement with Reference, Category Validation,

and Inferential Consistency. The structured evaluation form

ensured a systematic and consistent assessment of item quality.

• Cloze predictability: This criterion measures whether the

reviewer’s predicted word exactly matches the gold target

word. The ﬁrst reviewer achieved exact matches for 403

items (88.57%). Allowing for morphological variants shar-

ing the same root increases this to 408 items (89.67%),

and including valid synonyms further raises it to 417 items

(91.65%). (In the evaluation of language models in later

sections, we account for exact matches, root-based variants,

and synonym matches.)

The remaining 38 responses (8.35%) were distributed as

follows: 1 item (0.22%) was left blank (target word: knit-

ting), primarily due to reviewer oversight; 27 items (5.93%)

were incorrect due to inattention; and 10 items (2.20%)

reﬂected plausible alternative completions, indicating po-

tential ambiguity in item design.

For the second reviewer, exact matches were achieved

for 409 items (89.89%). Allowing for root-related variants

increases this to 411 items (90.33%), and including syn-

onyms raises this to 415 items (91.21%). The remaining

40 responses (8.79%) were distributed as follows: 4 items

(0.88%) were left blank; 30 items (6.59%) were incorrect

due to inattention; and 6 items (1.32%) reﬂected plausible

alternative completions.

• Agreement with reference: In this criterion, review-

ers evaluated whether the expected word represented the

most plausible continuation of the sentence. The expected

word was considered correct even if the reviewer predicted

a synonym or root variant (accounted for in the evaluation

metrics), made a minor error, or left the item blank. An item

was marked incorrect only when the reviewer identiﬁed a dif-

ferent plausible completion that was not synonymous with

the target word.

The ﬁrst reviewer judged 445 items as correct (97.80%),

while the second reviewer judged 449 items as correct

(98.68%), resulting in an inter-annotator agreement of

96.7%.

• Category validation: Only one category misclassiﬁcation

was identiﬁed by the second reviewer. The word em-

broidery, in this context, was incorrectly classiﬁed under

Patterns and Design instead of Crafts and Hobbies. The

corresponding item was:

“Maryam brought a piece of fabric and colored threads

to decorate her old clothes. She decided to learn the art of

. . . ” (Crafts and Hobbies ).

• Inferential consistency: To test compliance with the

CPARG format, both reviewers conﬁrmed that the primary

contextual cue is present in the ﬁrst sentence and that suc-

cessful inference requires the full context rather than only

the second sentence. The target word is not explicitly men-

tioned, conﬁrming that all items in the validation sample

adhere to the CPARG format.

The validation results demonstrate strong inter-annotator

agreement and support the semantic and categorical reliability

of JAMAL.

3.4.4. Cross-cultural adaptation and migration pipeline

After the design, creation, and validation of JAMAL, we in-

troduce a cross-cultural adaptation and migration pipeline to

enable its extension to new cultural and linguistic settings. JA-

MAL is language-agnostic, as its structural organization across

the functional, cognitive, and cultural axes is designed to gen-

eralize beyond Arabic. Adaptation to a new language or culture

can therefore be achieved through three complementary strate-

gies: translation of existing items, creation of new instances,

and culture-speciﬁc generation of novel concepts, with all steps

supported by human-in-the-loop validation.

Translation-based adaptation:

For cross-lingual transfer, automatic translation can be used

to accelerate initial dataset construction. However, translation

may alter syntactic structure and aﬀect constraints such as the

requirement that the target word appears in ﬁnal position in

the CPARG format. Therefore, all translated instances undergo

post-veriﬁcation to ensure compliance with the cloze structure

and to preserve unambiguous target predictability.

Basma Sayah et al.

Creation of new items:

When translation is insuﬃcient or when extending JAMAL,

new items are constructed directly following the CPARG for-

mat. This process consists of three steps:

1. Target word selection: Target words are selected for each

category (e.g., food, music, clothing) using lexical and

knowledge resources such as WordNet, ConceptNet, and

BabelNet, as well as localized sources such as Wikidata and

language model outputs. These resources provide culturally

relevant candidate concepts.

2. Context construction: Two-sentence contexts are con-

structed around each target word. The ﬁrst sentence

introduces a contextual cue, while the second sentence is de-

signed to end with the target word, which is removed during

evaluation. This step can be automated using the pipeline

described in Section 3.3.3, which generates CPARG-style

contexts.

3. Human validation and reﬁnement: Annotators verify that

each instance satisﬁes CPARG structural constraints and is

culturally appropriate for the target setting. When neces-

sary, they reﬁne wording or add minimal contextual cues to

ensure naturalness and unambiguous target predictability.

Culture-speciﬁc adaptation:

The cultural axis of JAMAL requires additional care during

adaptation, particularly for categories that contain culture-

dependent knowledge. These include celebrations, ﬁlms, arts

and music. In such cases, adaptation is not strictly transla-

tional but involves functional cultural mapping, where concepts

are replaced with culturally equivalent or culturally salient

alternatives in the target setting.

For example, a celebration such as Eid may be mapped to

Christmas or Diwali, a traditional dish to a locally equivalent

food item, and a reference to a popular ﬁlm to a culturally

corresponding iconic movie. New culturally grounded concepts

can also be introduced based on localized knowledge bases (e.g.,

Wikidata), native-speaker expertise, or culture-sp eciﬁc textual

resources.

This strategy enables ﬁne-grained adaptation not only

across languages but also across dialects and regional varieties,

supporting more localized and culturally sensitive evaluation

settings.

4. Experiments and Discussion

We evaluated ﬁve Arabic language models on JAMAL using

a Python script: AraGPT-base, MARBERT, AraBERT-large,

FANAR-9B, and FANAR-27B.

We evaluate model performance using four complementary

metrics:

• Exact match: assigns a score of 1 if the predicted word

exactly matches the gold target, and 0 otherwise.

• Synonym match: assigns a score of 1 if the predicted word

is either identical to or a synonym of the target, based on

Arabic WordNet, and 0 otherwise.

• Same root: uses the ISRI stemmer to determine whether

the predicted and target words share the same morphologi-

cal root.

The evaluation scripts and JAMAL are available at https:

//github.com/BasmaSayah/An-ICF- guided- commonsense- ben

chmark

• Cosine similarity: computes the semantic similarity be-

tween predicted and target words using FastText embed-

dings, yielding a score in [0, 1].

Together, these metrics provide a multi-faceted evalua-

tion of commonsense reasoning, capturing lexical accuracy,

morphological similarity, and semantic relatedness.

4.1. Overall commonsense evaluation

As shown in Figure 3, same-root scores consistently exceed

exact-match scores across all models, with AraGPT-base im-

proving from 13.93% to 17.27%, MarBERT from 16.34% to

18.91%, AraBERT-large from 4.77% to 10.20%, FANAR-9B

from 35.31% to 57.73%, and FANAR-27B from 62.88% to

72.53%. This pattern indicates that models frequently produce

the correct lexical item with diﬀerent conjugations rather than

the exact target word.

Similarly, synonym-match scores exceed exact-match scores

for all models: AraGPT-base rises from 13.93% to 17.87%, Mar-

BERT from 16.34% to 19.30%, AraBERT-large from 4.77% to

10.86%, FANAR-9B from 35.31% to 58.77%, and FANAR-27B

from 62.88% to 72.75%. This pattern is logical, as the syn-

onym metric assigns a positive score when the model predicts

either the exact word or one of its synonyms. Cosine similar-

ity yields the highest scores overall, 37.32% for AraGPT-base,

42.99% for MarBERT, 26.67% for AraBERT-large, 61.85% for

FANAR-9B, and 77.44% for FANAR-27B, suggesting that even

when models do not output the expected word or a direct syn-

onym, they often generate semantically related terms. These

cases are analyzed further in subsection 4.2.

In model comparisons, the FANAR models achieve the high-

est performance across all metrics, with FANAR-27B substan-

tially outperforming FANAR-9B, suggesting consistent beneﬁts

from scaling within the same architecture. Among smaller

models, MARBERT (165M parameters) surpasses the larger

AraBERT-large (370M parameters) across all metrics, and

AraGPT-base (135M parameters) also outperforms AraBERT-

large despite its smaller size.

These ﬁndings suggest that model size alone is not a

reliable predictor of commonsense reasoning performance. In-

stead, diﬀerences in pretraining data, architectural design, and

training objectives likely play an important role, although

their individual contributions cannot be disentangled in our

experiments.

4.2. Error analysis

In this subsection, we present a qualitative analysis of model er-

rors, followed by a theoretical interpretation of these behaviours

grounded in prior research on transformer-based language

models.

Error patterns:

We conducted a qualitative analysis of cases in which models

received a score of zero under the Exact Match, Same Root, and

Synonym Match metrics. Across AraGPT-base, MARBERT,

AraBERT-large, FANAR-9B, and FANAR-27B, we identiﬁed

several recurring error patterns. While not strictly mutually

exclusive, these patterns capture distinct dimensions of model

errors, ranging from semantic speciﬁcity and reasoning failures

to cultural bias and syntactic interference.

• Semantic neighbor substitution: Models often predict

semantically related concepts instead of the target word,

Multidimensional Arabic Commonsense Benchmark

Models

Score (%)

AraGPT-base

(135M)

MARBERT

(165M)

AraBERT-

Large

(370M)

FANAR

(9B)

FANAR 27B

(Parma)

13.93%

16.34%

4.77%

35.31%

62.88%

(a) Exact-match

Models

Score (%)

AraGPT-base

(135M)

MARBERT

(165M)

AraBERT-

Large

(370M)

FANAR

(9B)

FANAR 27B

17.27%

18.91%

10.2%

57.73%

72.53%

(b) Same-root

Models

Score (%)

AraGPT-base

(135M)

MARBERT

(165M)

AraBERT-

Large

(370M)

FANAR

(9B)

FANAR 27B

17.87%

19.3%

10.86%

58.77%

72.75%

Models

Score (%)

AraGPT-base

(135M)

MARBERT

(165M)

AraBERT-

Large

(370M)

FANAR

(9B)

FANAR 27B

37.32%

42.99%

26.67%

61.85%

77.44%

(d) Cosine similarity

Figure 3. Overall accuracy, same-ro ot, synonym, and cosine similarity scores across evaluated models.

reﬂecting partial domain awareness. Example: For the con-

text “When he traveled to Brazil, he saw a woman wearing

a colorful dress moving her hands in a rapid rhythm. She

was performing a [MASK]”, the expected answer is samba.

AraGPT-base predicts the dancers, MarBERT predicts

ballet, and AraBERT-large predicts hair.

• Hypernym substitution: Speciﬁc items are replaced by

broader categories, indicating a lack of precision. Exam-

ple: “The melodies moved between diﬀerent instruments in

astonishing harmony. The audience listened attentively to

the [MASK].” Expected: symphony. AraGPT-base predicts

song, FANAR-27B predicts playing.

• Hyponym confusion: Mo dels struggle to distinguish

closely related hyponyms in specialized domains. Exam-

ple: “She used a hooked needle and yarn to make a pillow

cover. She learned the craft of [MASK].” Expected: crochet.

FANAR-9B predicts embroidery, FANAR-27B predicts tai-

loring.

• Causal overextension: Models occasionally replace likely

outcomes with extreme or abstract consequences. Example:

“He walked on a thin rope above the circus. Everyone feared

that he would [MASK].” Expected: fall. AraGPT-base

predicts die.

• Associative / Cultural bias: Corpus-level or cultural

associations can override contextual constraints. Exam-

ple: “The boy listened to the song three times in a

row. He was trying to memorize the [MASK].” Expected:

lyrics. AraGPT-base, MarBERT, and FANAR-9B predict

the Quran.

• Context insensitivity / Syntactic priming: Predictions

may follow grammatical patterns while ignoring semantic

ﬁt. Example: “Ahmad raises his instrument to his chin and

gets ready. He plays a sad song on the [MASK].” Expected:

violin. AraGPT-base predicts then, FANAR-27B predicts

oud.

These error patterns are consistent across all models, al-

though they occur less frequently in FANAR-27B. Notably,

high cosine similarity often reﬂects semantically related but

Basma Sayah et al.

incorrect predictions rather than plausible completions, under-

scoring the limitations of embedding-based evaluation. Overall,

the results highlight persistent weaknesses in lexical precision,

causal reasoning, cultural awareness, and ﬁne-grained semantic

discrimination.

Interpretation of error patterns

To situate these ﬁndings within existing literature, we provide

theoretical explanations grounded in prior work on transformer-

based language models, highlighting interacting mechanisms

involving semantic representation, data distribution eﬀects, and

syntactic , as well as decoding biases.

• Semantic similarity errors (Semantic neighbor / Hy-

pernym substitution)): These errors arise when models

select tokens that are semantically related to the target

but diﬀer in speciﬁcity. This includes both overgeneral-

ized predictions (hypernyms) and contextual or associative

neighbors. The behavior reﬂects the organization of se-

mantic information in continuous embedding spaces, where

related concepts are embedded in nearby regions [49, 50].

As a result, next-token prediction is inﬂuenced by competi-

tion among semantically similar candidates, often favoring

higher-frequency or more generic alternatives with stronger

prior probability [51].

• Hyponym confusion: This error type reﬂects diﬃculty

in distinguishing closely related concepts within narrow

or specialized domains. Models often capture the broader

semantic ﬁeld but fail to resolve ﬁne-grained lexical distinc-

tions between near-hyponyms. Prior work suggests that dis-

tributional representations compress ﬁne-grained semantic

diﬀerences into shared latent regions [6]. This eﬀect is fur-

ther ampliﬁed by the Zipﬁan distribution of language, where

speciﬁc terms occur infrequently in training corpora [52].

Consequently, such terms receive fewer training signals,

leading to weaker and less reliable representations [53].

• Causal overextension: This pattern reﬂects a tendency

to generate salient or prototypical outcomes rather than

contextually constrained causal consequences. Instead of ex-

plicitly modeling intermediate causal steps, models rely on

high-probability continuations conditioned on surface con-

text (i.e., shallow lexical and syntactic cues rather than

deeper semantic or pragmatic understanding) [54]. This can

result in overly extreme or exaggerated predictions when

ﬁne-grained causal constraints are not strongly represented

in the training data.

• Associative / Cultural bias errors: These errors oc-

cur when strong corpus-level associations override local

contextual constraints. Frequent co-occurrence patterns in

pre-training data induce strong prior probabilities that

may dominate contextual conditioning. As a result, cul-

turally salient or high-frequency terms may be produced

even when they are not contextually appropriate, reﬂecting

well-documented bias propagation eﬀects in large language

models [55].

• Context insensitivity / Syntactic priming: These cases

indicate reliance on surface-level syntactic or lexical pat-

terns rather than deeper semantic integration. Models may

produce locally ﬂuent but globally inconsistent outputs

when semantic constraints are weak or ambiguous. Prior

studies show that transformer models often over-rely on lo-

cal dependencies learned during training, at the expense of

long-range semantic coherence [44].

Taken together, these theoretical perspectives suggest that the

observed error patterns are not isolated failures but system-

atic behaviours arising from the interaction between embedding

geometry, data distribution, and decoding biases. They also

reﬂect a broader tension between reliance on high-probability

training priors and the need for precise, contextually con-

strained reasoning.

4.3. Strengths and weaknesses of the models

Figures 4 and 5 show the root-match performance of AraGPT-

base, MARBERT, AraBERT-large, FANAR-9B, and FANAR-

27B across the 15 higher-level functional categories. Since the

full set of 56 categories is too dense for visual presentation,

we aggregate results at this higher level for clarity. The same

ﬁgures also report performance across the three cognitive cat-

egories (everyday situations, general knowledge, and problem

solving). Table 6 shows model performance across cultural

scope categories (universal, mainstream, and Arabic).

4.3.1. Performance across functional and cognitive

categories

AraGPT-base demonstrates relatively balanced performance

across the three commonsense branches (everyday situations,

general knowledge, problem solving), suggesting uniform treat-

ment of diﬀerent reasoning types. At the functional level, it

performs better in behavioural and environmentally grounded

domains (e.g., Human Actions & behaviours, Nature & Ecosys-

tems) but struggles in perceptual or visually grounded cate-

gories such as Colors and Shapes & Forms.

MarBERT achieves the strongest overall performance

among the smaller models, particularly in human-centric and

interaction-oriented domains (Personal & Social Life). This

aligns with its pretraining on diverse social media text. How-

ever, like AraGPT-base, it underperforms in ﬁne-grained per-

ceptual categories. In the cognitive branches, MarBERT excels

in Problem Solving compared to Everyday Situations and

General Knowledge.

AraBERT-large shows lower performance across most cate-

gories despite being pretrained on the same dataset as AraGPT-

base v2 [56]. AraBERT-large performs moderately in Materials

& Substances, Games, Sports & Recreation, and Clothing &

Accessories, and exhibits relatively better results in Problem

Solving, suggesting that increased capacity may aid abstract

reasoning.

Overall, these results conﬁrm that training data, model ar-

chitecture, pretraining objectives, and model size all inﬂuence

commonsense reasoning. Models pretrained on socially rich and

contextually diverse corpora, like MarBERT, exhibit stronger

and more consistent performance, especially in human-centered

domains.

Figure 5 shows that FANAR-9B substantially outperforms

smaller models across both cognitive and functional dimensions.

It achieves balanced scores across the three cognitive categories:

everyday situations, general knowledge, and problem solv-

ing; and demonstrates strong performance in human-centered

and practical functional domains such as Food & Nutrition,

Materials & Substances, Human Actions & behaviours, and

Transportation & Movement. Perceptual and functional cat-

egories, such as Colors and Shapes & Forms, remain compara-

tively weaker, indicating room for improvement in ﬁne-grained,

concrete knowledge representation. Animals & Living Beings

exhibits a comparable weakness.

Multidimensional Arabic Commonsense Benchmark

Score (%)

17.77

Personal & Social Life

6.25

Animals & Living Beings

15.51

Transportation & Movement

1.45

Colors, Shapes & Forms

16.25

Home & Living Environment

26.44

Nature & Ecosystems

17.23

Food & Nutrition

19.09

Clothing & Accessories

26.82

Materials & Substances

31.53

Human Actions & Behaviors

18.53

Physical & Mental Health

13.49

Arts, Music & Creativity

19.88

Games, Sports & Recreation

13.76

Fiction & Entertainment

20.50

Work & Productivity

(A1) AraGPT-base ro ot-match scores across life domains

Score (%)

16.78

Everyday Situation

16.94

General Knowledge

17.39

Problem Solving

(A2) AraGPT-base ro ot-match scores across cognitive dimension

Score (%)

31.03

Personal & Social Life

8.96

Animals & Living Beings

13.15

Transportation & Movement

7.59

Colors, Shapes & Forms

22.32

Home & Living Environment

27.15

Nature & Ecosystems

13.30

Food & Nutrition

14.70

Clothing & Accessories

16.21

Materials & Substances

25.13

Human Actions & Behaviors

21.68

Physical & Mental Health

14.27

Arts, Music & Creativity

18.51

Games, Sports & Recreation

19.67

Fiction & Entertainment

22.81

Work & Productivity

(B1) MarBERT root-match scores across life domains

Score (%)

19.07

Everyday Situation

17.81

General Knowledge

23.60

Problem Solving

(B2) MarBERT root-match scores across cognitive dimension

Score (%)

11.86

Personal & Social Life

2.46

Animals & Living Beings

4.55

Transportation & Movement

9.38

Colors, Shapes & Forms

11.43

Home & Living Environment

4.98

Nature & Ecosystems

5.87

Food & Nutrition

13.64

Clothing & Accessories

15.99

Materials & Substances

10.11

Human Actions & Behaviors

9.86

Physical & Mental Health

7.98

Arts, Music & Creativity

14.58

Games, Sports & Recreation

10.36

Fiction & Entertainment

12.02

Work & Productivity

(C1) AraBERT-large root-match scores across life domains

Score (%)

8.51

Everyday Situation

9.56

General Knowledge

21.74

Problem Solving

(C2) AraBERT-large root-match scores across cognitive dimension

Figure 4. Strengths and weaknesses of AraGPT-base, MarBERT, and AraBERT-large.

FANAR-27B consolidates the gains of FANAR-9B, outper-

forming it across nearly all functional and cognitive categories.

Notable improvements are seen in visually grounded domains

such as Colors, Shapes & Forms (+35%) , as well as in Animals

& Living Beings (+21%), and Home & Living Environment

(+23%). Cognitive categories also beneﬁt, with Everyday Sit-

uations (+14%), General Knowledge (+15%), and Problem

Solving (+11%) showing clear gains.

FANAR-27B achieves strong performance across functional

categories and shows improved contextual integration compared

to smaller models, with fewer extreme or irrelevant predictions.

Its performance is closer to human cloze predictability in several

visually grounded and everyday domains, suggesting improved

semantic precision with increased model capacity. Despite these

advances, the model still exhibits a tendency toward overgener-

alization in some cases, and persistent gaps in problem solving

indicate that challenges in higher-order reasoning remain.

4.3.2. Performance analysis by cultural grounding

Building on the previous analysis, we examine model perfor-

mance across the three cultural grounding categories: Universal

(N = 1,615 items; knowledge shared across cultures, such as

physical laws and daily routines), Western/Global (N = 78

items; mainstream concepts common in global media, such as

karaoke or Monopoly), and Arabic-speciﬁc (N = 131 items;

knowledge grounded in Arabic cultural contexts). This distri-

bution reﬂects the natural skew in commonsense knowledge,

where universal concepts are inherently more frequent than lo-

calized ones, thereby preserving realistic real-world frequencies

rather than enforcing an artiﬁcially uniform design. The results,

reported in Table 6 show that:

• Across all models, universal items consistently achieve

higher performance than Arabic-speciﬁc items. For exam-

ple, AraGPT-base achieves 14.49% on universal vs. 10.69%

on Arabic, MarBERT 17.28% vs. 12.21%, AraBERT-large

4.95% vs. 2.29%, FANAR-9B 35.60% vs. 30.53%, and

Basma Sayah et al.

Score (%)

60.28

Personal & Social Life

47.36

Animals & Living Beings

67.24

Transportation & Movement

45.71

Colors, Shapes & Forms

54.28

Home & Living Environment

55.62

Nature & Ecosystems

68.62

Food & Nutrition

58.66

Clothing & Accessories

75.64

Materials & Substances

68.12

Human Actions & Behaviors

63.78

Physical & Mental Health

48.19

Arts, Music & Creativity

51.39

Games, Sports & Recreation

47.82

Fiction & Entertainment

66.48

Work & Productivity

(A1) Fanar-9B root-match scores across life domains

Score (%)

58.84

Everyday Situation

58.12

General Knowledge

53.42

Problem Solving

(A2) Fanar-9B cognitive dimension scores

Score (%)

77.28

Personal & Social Life

67.88

Animals & Living Beings

68.46

Transportation & Movement

80.54

Colors, Shapes & Forms

77.63

Home & Living Environment

72.90

Nature & Ecosystems

85.30

Food & Nutrition

71.02

Clothing & Accessories

85.65

Materials & Substances

82.63

Human Actions & Behaviors

71.80

Physical & Mental Health

61.91

Arts, Music & Creativity

68.11

Games, Sports & Recreation

64.17

Fiction & Entertainment

79.67

Work & Productivity

(B1) Fanar-27B root-match scores across life domains

Score (%)

72.59

Everyday Situation

73.31

General Knowledge

63.98

Problem Solving

(B2) Fanar-27B cognitive dimension scores

Figure 5. Strengths and weaknesses of Fanar-9B and Fanar-27B.

FANAR-27B 64.00% vs. 52.56%. This indicates a persistent

performance gap between universal and culturally speciﬁc

knowledge across all models.

• Performance on western/global items varies across mod-

els. AraGPT-base (7.69%) and MarBERT (3.85%) score

lower on western items than on either Universal or Arabic-

speciﬁc categories. In contrast, AraBERT-large (5.13%) and

FANAR-9B (37.17%) achieve slightly higher scores on west-

ern items than on Universal items. FANAR-27B (54.96%)

shows intermediate performance, with western scores lower

than universal (64.00%) but higher than Arabic-speciﬁc

items (52.56%). This variation may reﬂect diﬀerences in

model capacity and exposure to culturally diverse content

during training.

• Scaling from FANAR-9B to FANAR-27B yields higher ab-

solute performance across all categories, with FANAR-27B

reaching 64.00% (universal), 54.96% (western), and 52.56%

(Arabic). However, the Universal–Arabic gap remains sub-

stantial, suggesting that while increased model capacity

improves overall performance, disparities across cultural

categories persist.

Overall, all models consistently show lower performance on

Arabic-speciﬁc items than on universal items. This gap may

reﬂect the comparatively limited representation of culturally

speciﬁc knowledge in large-scale pretraining data rather than

diﬀerences in item quality or ambiguity, particularly given the

high human validation performance on these items.

Multidimensional Arabic Commonsense Benchmark

Model Overall Universal (U) Western (W) Arabic (A) Gap (U-A) Gap (U-W)

AraGPT-base 13.93% 14.49% 7.69% 10.69% 3.80% 6.80%

MarBERT 16.34% 17.28% 3.85% 12.21% 5.06% 13.43%

AraBERT-large 4.77% 4.95% 5.13% 2.29% 2.66% -0.17%

FANAR-9B 35.31% 35.60% 37.17% 30.53% 5.07% -1.58%

FANAR-27B 62.88% 64.00% 54.96% 52.56% 9.06% 11.46%

Table 6. Performance comparison across cultural grounding categories. Overall accuracy is reported in the Overall column.

5. Conclusion

This study introduces JAMAL, a language-agnostic frame-

work for evaluating commonsense reasoning in language models,

alongside its Arabic instantiation as a benchmark. JAMAL

adopts a three-axis taxonomy spanning a functional dimension

(56 life-domain categories), a cognitive dimension (everyday

situations, general knowledge, and problem-solving), and a

cultural grounding axis (universal, western/global, and Arabic-

speciﬁc knowledge). Together, these axes enable ﬁne-grained,

multi-dimensional evaluation that diagnoses model common-

sense reasoning capabilities.

JAMAL is constructed through a multi-stage pipeline com-

bining manual curation and LLM-assisted generation with

human reﬁnement. This hybrid design leverages the scalability

of LLMs while maintaining quality through human veriﬁcation.

Its items follow the CPARG format [44], which imposes strict

structural constraints and relies on subtle contextual cues to

elicit commonsense inference.

The empirical evaluation of ﬁve Arabic language models

reveals consistent performance gaps across functional, cogni-

tive, and cultural dimensions, with FANAR-27B achieving the

strongest overall results. Overall, JAMAL provides a struc-

tured and extensible benchmark for interpretable evaluation of

commonsense reasoning, addressing a key gap in Arabic NLP

and supporting future cross-lingual and culturally aware model

assessment.

6. Limitations and Future Perspectives

While JAMAL provides a structured framework for evaluating

contextual commonsense reasoning in Arabic language mod-

els, it is not exhaustive. Commonsense knowledge is broad,

dynamic, and culturally situated. Future work could expand

the benchmark with additional examples across functional life-

domain categories and the problem-solving dimension, as well

as include a wider range of culturally speciﬁc cases reﬂecting

the diversity of Arabic-speaking communities.

The current benchmark primarily uses Modern Standard

Arabic (MSA). Extending this work to Arabic dialects and

more diverse linguistic settings is a natural direction for fu-

ture research and would allow for a broader evaluation of model

robustness across diﬀerent forms of Arabic usage.

Although JAMAL has undergone manual veriﬁcation, a

full human performance baseline over the entire dataset re-

mains an important direction for future work. Such a baseline

would enable a more precise estimation of the human–model

performance gap.

Finally, the CPARG format is a high-constraint cloze-based

design in which each item presents a controlled two-sentence

context with a single masked target. This reduces prompt

variability and enables consistent assessment of contextual in-

ference. However, it reﬂects a constrained form of language use

rather than open-ended interaction, and thus captures only a

subset of naturalistic language behavior.

Ethical Statement

No ethical approval was required for this study, as it did not

involve human or animal subjects.

Funding

This research received no speciﬁc grant from any funding

agency in the public, commercial, or not-for-proﬁt sectors.

Declaration of competing interests

The authors declare that they have no known competing ﬁnan-

cial interests or personal relationships that could have appeared

to inﬂuence the work reported in this paper.

Data Availability Statements

The data supporting the ﬁndings of this study are openly avail-

able in https://github.com/BasmaSayah/An- ICF- guided- commons

ense-benchmark.

Credit authorship contribution statement

Basma Sayah: Conceptualization, Methodology, Data cura-

tion, Validation, Investigation, Writing – Original Draft. Attia

Nehar: Methodology, Investigation, Writing – Review & Edit-

ing. Hadda Cherroun: Methodology, Investigation, Writing –

Review & Editing. Slimane Bellaouar: Methodology, Investiga-

tion, Writing – Review & Editing. Firoj Alam: Methodology,

Investigation, Resources, Writing – Review & Editing.

Basma Sayah et al.

A. JAMAL Functional Dimension Categories Adapted from the International Classiﬁcation of

Functioning (ICF)

Benchmark Domain Benchmark Subcategories Mapped ICF Code and Component

Personal & Social Life Family & Friends, Celebrations, Emotion, Com-

munication

d7 “Interpersonal interactions and relationships”

Animals & Living Beings Pets, Wildlife, Farming Animals, Insects & Small

Creatures

e245 “Animals”

Transportation & Movement Land Transport, Water Transport, Air Transport,

Walking & Running & Movement

d4 “Mobility”

Colors, Shapes, and Forms Basic Colors, Geometrical Shapes, Patterns &

Designs

b156 “Perception of visual stimuli”; b160

“Thought functions”

Home & Living Environment Buildings & Structures, Rooms & Spaces, House-

hold Objects, Household Tools & Appliances

e1 “Products and technology”

Nature & Ecosystems Plants & Trees, Bodies of Water, Weather & Sea-

sons, Landscapes

e2 “Natural environment and human-made

changes to environment”

Food & Nutrition Types of Food, Cooking & Eating, Farming &

Gathering

d550 “Eating”; d630 “Preparing meals”

Clothing & Personal Accessories Fabrics & Textiles, Types of Clothes, Footwear,

Accessories

d540 “Dressing”; e1 “Products and technology”

Materials & Substances Natural Materials, Manufactured Materials,

Chemical Substances

e1 “Products and technology”; e2 “Natural en-

vironment and human-made changes to environ-

ment”

Human Actions & Behaviors Learning & Education, Work Actions, Play Ac-

tions, Helping & Caring

d1 “Learning and applying knowledge”; d8 “Major

life areas”

Physical & Mental Health Physical Abilities, Emotional Well-being, Ill-

nesses & Conditions, Health Maintenance

b1 “Mental functions”; b7 “Neuromusculoskeletal

and movement-related functions”

Arts, Music, and Creativity Visual Arts, Performing Arts, Music, Crafts &

Hobbies

d920 “Recreation and leisure”; b160 “Thought

functions”

Games, Sports, and Recreation Team Sports, Individual Sports, Board Games,

Recreational Activities (e.g., hide and seek, tag,

playground games)

d920 “Recreation and leisure”

Fiction & Entertainment Stories & Books, Films & TV Shows, Fantasy &

Myths

d920 “Recreation and leisure”; b160 “Thought

functions”

Work & Productivity Jobs & Occupations, Workplaces, Workplace

Tools & Machinery, Business

d850 “Remunerative employment”; e1 “Products

and technology”

Table 7. ICF-based functional dimension categories used in the JAMAL benchmark construction.

Multidimensional Arabic Commonsense Benchmark

B. Distribution of JAMAL items across the 56 functional categories

0 5 10 15 20 25 30 35 40 45 50 55 60

Board Games

Individual Sports

Team Sports

Play and Recre-

ational Activities

Crafts and Hobbies

Music

Performing Arts

Visual Arts

Comics and Animation

Fantasy and Myths

Films and TV Shows

Stories and Books

Health Maintenance

Illnesses and Conditions

Physical Abilities

Patterns and Designs

Geometrical Shapes

Basic Colors

Learning and Education

Work Actions

Helping and Caring

Chemical Substances

Manufactured Materials

Natural Materials

Footwear

Accessories

Types of Clothes

Fabrics and Textiles

Farming and Gathering

Cooking and Eating

Types of Food

Business and Commerce

Tools and Technology

Workplaces and

Environments

Jobs and Occupations

Landscapes

Weather and Seasons

Bodies of Water

Plants and Trees

Rooms and Spaces

Household Objects

Buildings and Structures

Walking

Running

and Movement

Air Transport

Water Transport

Land Transport

Insects and Small Creatures

Farming Animals

Wildlife

Pets

Communication

Emotions

Number of Items

Games/Sports/Recreation Arts/Music/Creativity Fiction/Entertainment

Physical/Mental Health Colors/Shapes/Forms Human Actions/Behaviors

Materials/Substances Clothing/Accessories Food/Nutrition

Work/Productivity Nature/Ecosystems Home/Living Environment

Transportation/Movement Animals/Living Beings Personal/Social Life

Figure 6. Distribution of JAMAL items across functional categories.

Basma Sayah et al.

C. Benchmark Veriﬁcation Guidelines

The reviewer was tasked with examining the entire benchmark.

If a context failed to satisfy any of the following conditions, the

reviewer was instructed to mark it with a ✗ in the designated

box.

Veriﬁcation Criteria

• Each context must consist of two sentences.

• The target word must be the most natural completion of

the context.

• No alternative completion should be reasonably plausible

upon reading the context.

• The target word should not be explicitly mentioned in the

context; however, the ﬁrst sentence should provide a clear

hint toward it.

• The correct completion should not be inferable from the

second sentence alone.

• The target word must be suﬃciently clear and predictable

so that a reader can infer it from the context.

Example of a Valid Context

✓ Context: Rosalyn announced happily, “Checkmate.” She

was going to become really good at playing. . .

Target word: chess

Justiﬁcation: The ﬁrst sentence provides a clear cue

(“checkmate”), making “chess” the only plausible comple-

tion.

Examples of Invalid Contexts

1. Context: Ahmad went to the market. He bought some.. .

Target word: apples

✗ Justiﬁcation: Multiple completions are possible (e.g.,

bread, milk, meat, fruit).

2. Context: Layla decided to bake fresh bread in the oven.

She enjoyed the smell of. . .

Target word: bread

✗ Justiﬁcation: The target word (“bread”) is explicitly

mentioned in the ﬁrst sentence.

3. Context: The class time ended. The teacher started wiping

the. . .

Target word: board

✗ Justiﬁcation: The correct completion can be inferred

from the second sentence alone.

D. Benchmark Validation Guidelines

The reviewers were tasked with completing sentence contexts

and assessing multiple criteria for each benchmark item. The

following guidelines outline the validation procedure.

Validation Objectives

• Assess cloze predictability: whether reviewer predictions

match the target words.

• Assess reference agreement: whether the expected word

represents the only natural completion.

• Determine inferential consistency: which sentence provides

contextual clues for prediction.

• Verify category alignment : whether contexts properly be-

long to their assigned categories.

Validation Procedure

Step 1: Read the full context

Each item consists of two consecutive sentences. Read both

sentences before making any judgments.

Step 2: Predict the ﬁnal word

Write a single word that most naturally completes the con-

text. Choose the most coherent completion based on linguistic

intuition.

Step 3: Mark inference sources

✓ Inferable from Sentence 1: Mark if the ﬁrst sentence

contains the hint to the target word.

✓ Inferable from Sentence 2: Mark if the target word can

be predicted from the second sentence alone.

Step 4: Validate category assignment

✓ Matches Category 1: Mark if the context clearly aligns

with Category 1.

✓ Matches Category 2: Mark if the context clearly aligns

with Category 2.

Step 5: Compare the predicted word to the expected word

✓ Mark True if the predicted word matches the expected word

exactly.

✓ Mark False if the predicted word does not match the

expected word.

Step 6: Assess the expected word

✓ Mark True if the expected word is the only plausible

completion.

✓ Mark False if there exist other plausible completions that

are not synonymous with the expected word.

Step 7: Document issues

Use the notes column to record:

• Unclear sentences.

• Mismatches between context and assigned categories.

• Diﬃculties in predicting the target word.

• Any other relevant observations.

Example of Proper Annotation

✓ Context: He put the letter in the envelope and dropped it

in the mailbox. The worker noticed it was missing a . . .

Target word: stamp

Category 1 (Cognitive): Everyday Situation

Category 2 (Functional): Communication

Validator prediction: stamp

Annotation: Same word: ✓, Inferable from S1: ✓, Inferable

from S2 alone: ✗, Matches Category 1: ✓, Matches Category

2: ✓, Uniqueness of the expected completion: ✓

General Instructions

• Rely on natural linguistic intuition; no external resources

are permitted.

• Work independently without discussing items with the other

reviewer.

• Leave cells blank when uncertain rather than guessing.

• Apply check marks only in designated columns.

Submission Requirements

Before ﬁnal submission, reviewers must:

Multidimensional Arabic Commonsense Benchmark

1. Ensure each prediction ﬁeld contains exactly one word.

2. Verify check marks are placed only where appropriate.

3. Conﬁrm all notes provide clear explanations where needed.

4. Save the completed ﬁle and send it to the researcher.

References

1. T. B. Brown et al., “Language models are few-shot

learners,” 2020, doi: https://doi.org/10.48550/arXiv.2005.

14165.

2. OpenAI, J. Achiam, S. Adler, S. Agarwal, L. Ahmad et al.,

“Gpt-4 technical report,” arXiv preprint arXiv:2303.08774,

2023, doi: https://doi.org/10.48550 /arXiv .2303.08774.

[Online]. Available: https://arxiv.org/abs/2303.08774

3. H. Touvron et al., “Llama 2: Open foundation and ﬁne-

tuned chat models,” arXiv preprint arXiv:2307.09288,

2023. [Online]. Available: https://arxiv.org/abs/2307.09288

4. J. Bai et al., “Qwen technical report,” arXiv preprint

arXiv:2309.16609, 2023, doi: https://doi.org/10.48550/a

rXiv.2309.16609.

5. G. Team et al., “Gemini: A family of highly capable

multimodal models,” 2025. [Online]. Available: https:

//arxiv.org/abs/2312.11805

6. R. Zellers, A. Holtzman, Y. Bisk, A. Farhadi, and Y. Choi,

“HellaSwag: Can a machine really ﬁnish your sentence?” in

Proceedings of the 57th Annual Meeting of the Association

for Computational Linguistics, A. Korhonen, D. Traum,

and L. Màrquez, Eds. Florence, Italy: Association for

Computational Linguistics, Jul. 2019, pp. 4791–4800, doi:

https://doi.org/10.18653/v1/P19-1472. [Online]. Available:

https://aclanthology.org/P19-1472/

7. D. Hendrycks et al., “Measuring massive multitask

language understanding,” 2021. [Online]. Available:

https://arxiv.org/abs/2009.03300

8. G. Lai, Q. Xie, H. Liu, Y. Yang, and E. Hovy,

“RACE: Large-scale ReAding comprehension dataset from

examinations,” in Proceedings of the 2017 Conference

on Empirical Methods in Natural Language Processing,

M. Palmer, R. Hwa, and S. Riedel, Eds. Copenhagen,

Denmark: Association for Computational Linguistics, Sep.

2017, pp. 785–794, doi: https://doi.org/10.18653/v1/D1

7-1082. [Online]. Available: https://aclanthology.org/D17

-1082/

9. A. Abdelali et al., “ LAraBench: Benchmarking Arabic

AI with large language models,” in Proceedings of

the 18th Conference of the European Chapter of

the Association for Computational Linguistics (Volume

1: Long Papers), Y. Graham and M. Purver, Eds.

St. Julian’s, Malta: Association for Computational

Linguistics, Mar. 2024, pp. 487–520, doi: https://do

i.org/10.18653/v1/2024.eacl- long.30. [Online]. Available:

https://aclanthology.org/2024.eacl-long.30/

10. M. T. I. Khondaker, A. Waheed, E. M. B. Nagoudi,

and M. Abdul-Mageed, “ GPTAraEval: A comprehensive

evaluation of ChatGPT on Arabic NLP,” in Proceedings

of the 2023 Conference on Empirical Methods in

Natural Language Processing, H. Bouamor, J. Pino, and

K. Bali, Eds. Singapore: Association for Computational

Linguistics, Dec. 2023, pp. 220–247, doi: https:

//d oi .org/10.18653 /v 1/ 20 23 .emnlp-main.16. [Online].

Available: https://aclanthology.org/2023.emnlp-main.16/

11. F. Koto et al., “ ArabicMMLU: Assessing massive multitask

language understanding in Arabic,” in Findings of

the Association for Computational Linguistics: ACL

2024, L.-W. Ku, A. Martins, and V. Srikumar,

Eds. Bangkok, Thailand: Association for Computational

Linguistics, Aug. 2024, pp. 5622–5640, doi: htt ps:

//doi.org /10.18653/v 1/202 4.find ings- ac l.334. [Online].

Available: https://aclanthology.org/2024.ﬁndings- acl.334/

12. B. Mousi et al., “AraDiCE: Benchmarks for dialectal

and cultural capabilities in LLMs,” in Proceedings of

the 31st International Conference on Computational

Linguistics, O. Rambow, L. Wanner, M. Apidianaki,

H. Al-Khalifa, B. D. Eugenio, and S. Schockaert,

Eds. Abu Dhabi, UAE: Association for Computational

Linguistics, Jan. 2025, pp. 4186–4218. [Online]. Available:

https://aclanthology.org/2025.coling-main.283/

13. R. N. Almatham et al., “BALSAM: A platform for bench-

marking Arabic large language models,” in Proceedings of

The Third Arabic Natural Language Processing Confer-

ence, K. Darwish et al., Eds. Suzhou, China: Association

for Computational Linguistics, Nov. 2025, pp. 258–277,

doi: https://doi.org/10.18653/v1/2025.arabicnlp- main.21.

[Online]. Available: https://aclanthology.org/2025.arabicnl

p-main.21/

14. S. Al-Khalifa, N. Durrani, H. Al-Khalifa, and F. Alam,

“The landscape of arabic large language models,” Commu-

nications of the ACM, vol. 68, no. 10, pp. 54–61, 2025,

doi: https://doi.org/10.1145/3737453.

15. K. Smith et al., “The origins of common sense in humans

and machines,” in Proceedings of the Annual Meeting of the

Cognitive Science Society, vol. 42, 2020, pp. 3–4. [Online].

Available: https://escholarship.org/uc/item/2367w9c4

16. S. K. Tawalbeh and M. Al-Smadi, “Is this sentence

valid? an arabic dataset for commonsense validation,”

CoRR, vol. abs/2008.10873, 2020. [Online]. Available:

https://arxiv.org/abs/2008.10873

17. N. Sengupta et al., “Jais and jais-chat: Arabic-centric

foundation and instruction-tuned open generative large

language models,” 2023. [Online]. Available: ht tp s:

//arxiv.org/abs/2308.16149

18. M. S. Bari et al., “ALLam: Large language models for

arabic and english,” in The Thirteenth International

Conference on Learning Representations, 2025. [Online].

Available: https://openreview.net/forum?id=MscdsFVZrN

19. F. Team et al., “Fanar: An arabic-centric multimodal

generative ai platform,” 2025. [Online]. Available:

https://arxiv.org/abs/2501.13944

20. World Health Organization, International Classiﬁcation

of Functioning, Disability and Health (ICF). Geneva:

World Health Organization, 2001. [Online]. Available:

https://www.who.int/classiﬁcations/icf/en/

21. R. Zellers, Y. Bisk, R. Schwartz, and Y. Choi,

“SWAG: A large-scale adversarial dataset for grounded

commonsense inference,” in Proceedings of the 2018

Conference on Empirical Methods in Natural Language

Processing, E. Riloﬀ, D. Chiang, J. Hockenmaier, and

J. Tsujii, Eds. Brussels, Belgium: Association for

Computational Linguistics, Oct.-Nov. 2018, pp. 93–104,

doi: https: //doi .org/ 10.18 653/v 1/D18-100 9. [Online].

Available: https://aclanthology.org/D18-1009/

22. A. Talmor, J. Herzig, N. Lourie, and J. Berant, “ Com-

monsenseQA: A question answering challenge targeting

commonsense knowledge,” in Proceedings of the 2019 Con-

ference of the North American Chapter of the Association

for Computational Linguistics: Human Language Tech-

nologies, Volume 1 (Long and Short Papers), J. Burstein,

Basma Sayah et al.

C. Doran, and T. Solorio, Eds. Minneapolis, Minnesota:

Association for Computational Linguistics, Jun. 2019, pp.

4149–4158, doi: https://doi.org/10.18653/v1/N19- 1421.

[Online]. Available: https://aclanthology.org/N19-1421/

23. M. Sap, H. Rashkin, D. Chen, R. Le Bras, and

Y. Choi, “Social IQa: Commonsense reasoning about social

interactions,” in Proceedings of the 2019 Conference on

Empirical Methods in Natural Language Processing and

the 9th International Joint Conference on Natural Lan-

guage Processing (EMNLP-IJCNLP), K. Inui, J. Jiang,

V. Ng, and X. Wan, Eds. Hong Kong, China: Association

for Computational Linguistics, Nov. 2019, pp. 4463–4473,

doi: https: //doi .org/ 10.18 653/v 1/D19-145 4. [Online].

Available: https://aclanthology.org/D19-1454/

24. Y. Bisk, R. Zellers, R. Le Bras, J. Gao, and Y. Choi,

“Piqa: Reasoning about physical commonsense in natural

language,” in Proceedings of the AAAI Conference on

Artiﬁcial Intel ligence, vol. 34, no. 05, 2020, pp. 7432–7439,

doi: https://doi.org/10.1609/aaai.v34i05.6239. [Online].

Available: https://ojs.aaai.org/index.php/AAAI/article/vi

ew/6239

25. M. Chen, M. D’Arcy, A. Liu, J. Fernandez, and D. Downey,

“CODAH: An adversarially-authored question answering

dataset for common sense,” in Proceedings of the 3rd

Workshop on Evaluating Vector Space Representations

for NLP, A. Rogers, A. Drozd, A. Rumshisky, and

Y. Goldberg, Eds. Minneapolis, USA: Association

for Computational Linguistics, Jun. 2019, pp. 63–69,

doi: https ://doi.org/ 10. 18653/v1 /W19- 2008. [Online].

Available: https://aclanthology.org/W19- 2008/

26. B. Y. Lin, S. Lee, X. Qiao, and X. Ren, “Common

sense beyond English: Evaluating and improving mul-

tilingual language models for commonsense reasoning,”

in Proceedings of the 59th Annual Meeting of the

Association for Computational Linguistics and the 11th

International Joint Conference on Natural Language

Processing (Volume 1: Long Papers), C. Zong, F. Xia,

W. Li, and R. Navigli, Eds. Online: Association for

Computational Linguistics, Aug. 2021, pp. 1274–1287, doi:

https://doi.org/10.18653/v1/2021.acl- long.102. [Online].

Available: https://aclanthology.org/2021.acl-long.102/

27. H. Santos, A. M. Mulvehill, K. Shen, M. Kejriwal,

and D. L. McGuinness, “Tg-csr: A human-labeled

dataset grounded in nine formal commonsense categories,”

Data in Brief, vol. 51, p. 109666, 2023, doi:

htt ps:// doi.o rg /10 .1 016 /j .dib. 2023. 10 966 6. [Online].

Available: https://www.sciencedirect.com/science/article/

pii/S2352340923007515

28. S. Saha, P. Yadav, L. Bauer, and M. Bansal, “ ExplaGraphs:

An explanation graph generation task for structured

commonsense reasoning,” in Proceedings of the 2021

Conference on Empirical Methods in Natural Language

Processing, M.-F. Moens, X. Huang, L. Specia, and

S. W.-t. Yih, Eds. Online and Punta Cana, Dominican

Republic: Association for Computational Linguistics,

Nov. 2021, pp. 7716–7740, doi: https: //doi.o rg/10

.18 65 3/ v1/2021.emnl p- main.609. [Online]. Available:

https://aclanthology.org/2021.emnlp-main.609/

29. W. Zhan et al., “Score: Benchmarking long-chain reasoning

in commonsense scenarios,” 2025. [Online]. Available:

https://arxiv.org/abs/2503.06218

30. H. Mozannar, E. Maamary, K. El Ha jal, and H. Hajj,

“Neural Arabic question answering,” in Proceedings of the

Fourth Arabic Natural Language Processing Workshop,

W. El-Ha jj et al., Eds. Florence, Italy: Association

for Computational Linguistics, Aug. 2019, pp. 108–118,

doi: ht tps ://doi.org/ 10. 18653/v1 /W19- 4612. [Online].

Available: https://aclanthology.org/W19- 4612/

31. J. L. Lee et al., “Massively multilingual pronunciation

modeling with WikiPron,” in Proceedings of the Twelfth

Language Resources and Evaluation Conference, N. Cal-

zolari et al., Eds. Marseille, France: European Language

Resources Association, May 2020, pp. 4223–4228. [Online].

Available: https://aclanthology.org/2020.lrec-1.521/

32. S. Lamsiyah et al., “Arabicsense: A benchmark for

evaluating commonsense reasoning in arabic with large

language models,” in Proceedings of the 4th Workshop on

Arabic Corpus Linguistics (WACL-4), 2025, pp. 1–11.

[Online]. Available: https://aclanthology.org/2025.wacl-1.1

33. A. Sadallah et al., “Commonsense reasoning in arab

culture,” in Proceedings of the 63rd Annual Meeting of

the Association for Computational Linguistics (Volume

1: Long Papers), 2025, pp. 7695–7710, doi: https :

//doi.org/10.18650/ACL.2025.380.

34. A. Hornby and J. Turnbull, Oxford Advanced Learner’s

Dictionary of Current English. Oxford University Press,

2015.

35. F. Ilievski, A. Oltramari, K. Ma, B. Zhang, D. L.

McGuinness, and P. Szekely, “Dimensions of commonsense

knowledge,” Knowledge-Based Systems, vol. 229, p.

107347, 2021, doi: https://doi.org/10.1016/j.knosys.2021.

107347. [Online]. Available: https://www.sciencedirect.co

m/science/article/pii/S0950705121006092

36. D. B. Lenat, “Cyc: a large-scale investment in knowledge

infrastructure,” Commun. ACM, vol. 38, no. 11, p. 33–38,

Nov. 1995, doi: https://doi.org/10.1145/219717.219745.

37. M. E. Whiting and D. J. Watts, “A framework for

quantifying individual and collective common sense,”

Proceedings of the National Academy of Sciences,

vol. 121, no. 4, p. e2309535121, 2024, doi: htt ps:

//doi.org/10.1073/pnas.2309535121.

38. H. A. Simon and A. Newell, Human problem solving:

The state of the theory in 1970. American Psychological

Association, 1971, vol. 26, no. 2, doi: https://doi.org/10.1

037/h0030806.

39. R. C. Schank and R. P. Abelson, Scripts, plans, goals,

and understanding: An inquiry into human know ledge

structures. Psychology press, 2013.

40. P. Schuyler, Common Sense. Los Angeles, CA: Higher

Shelf Publishing, 2003. [Online]. Available: https :

//www.amazon.com/Common- Sense- Peter- Schuyler/dp/1

932636021

41. R. Pesonen, “Casual reasoning : A social ecological look

at human cognition and common sense,” 2019. [Online].

Available: htt ps :/ /a pi .s em anticscho la r. or g/ Co rp us ID:

202257933

42. J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “Bert:

Pre-training of deep bidirectional transformers for language

understanding,” in North American Chapter of the

Association for Computational Linguistics, 2019. [Online].

Available: htt ps :/ /a pi .s em anticscho la r. or g/ Co rp us ID:

52967399

43. A. Radford, “Improving language understanding with

unsupervised learning,” OpenAI Res, 2018. [Online].

Available: https://cdn.openai.com/research-covers/langua

ge-unsupervised/language_understanding_paper.pdf

Multidimensional Arabic Commonsense Benchmark

44. A. Ettinger, “What BERT is not: Lessons from a new

suite of psycholinguistic diagnostics for language models,”

Transactions of the Association for Computational

Linguistics, vol. 8, pp. 34–48, 2020, doi: htt ps :

//d oi.org/ 10.1162 /tacl_a _00298. [Online]. Available:

https://aclanthology.org/2020.tacl-1.3/

45. K. D. Federmeier and M. Kutas, “A rose by any other name:

Long-term memory structure and sentence processing,”

Journal of Memory and Language, vol. 41, no. 4, pp. 469–

495, 1999, doi: https://doi.org/10.1006/jmla.1999.2660.

[Online]. Available: https://www.sciencedirect.com/scienc

e/article/pii/S0749596X99926608

46. K. Sakaguchi, R. L. Bras, C. Bhagavatula, and Y. Choi,

“Winogrande: an adversarial winograd schema challenge at

scale,” Commun. ACM, vol. 64, no. 9, p. 99–106, Aug.

2021, doi: https://doi.org/10.1145/3474381.

47. R. H. Ennis, W. L. Gardiner, R. Morrow, D. Paulus, and

L. Ringel, The cornell class-reasoning test, Form X, 1964.

48. S. Ouyang, J. M. Zhang, M. Harman, and M. Wang, “An

empirical study of the non-determinism of chatgpt in code

generation,” ACM Trans. Softw. Eng. Methodol., vol. 34,

no. 2, Jan. 2025, doi: https://doi.org/10.1145/3697010.

49. K. Ethayarajh, “How contextual are contextualized word

representations? Comparing the geometry of BERT, ELMo,

and GPT-2 embeddings,” in Proceedings of the 2019

Conference on Empirical Methods in Natural Language

Processing and the 9th International Joint Conference on

Natural Language Processing (EMNLP-IJCNLP), K. Inui,

J. Jiang, V. Ng, and X. Wan, Eds. Hong Kong, China:

Association for Computational Linguistics, Nov. 2019,

pp. 55–65, doi: https://doi.org/10.18653/v1/D19- 1006.

[Online]. Available: https://aclanthology.org/D19-1006/

50. W. Gurnee, N. Nanda, M. Pauly, K. Harvey, D. Troitskii,

and D. Bertsimas, “Finding neurons in a haystack:

Case studies with sparse probing,” Transactions on

Machine Learning Research, 2023. [Online]. Available:

https://openreview.net/forum?id=JYs1R9IMJr

51. A. Holtzman, J. Buys, L. Du, M. Forbes, and Y. Choi, “The

curious case of neural text degeneration,” in International

Conference on Learning Representations, 2020. [Online].

Available: https://openreview.net/forum?id=rygGQyrFvH

52. G. K. Zipf, Human Behavior and the Principle of Least

Eﬀort: An Introduction to Human Ecology. Cambridge,

MA: Addison-Wesley, 1949.

53. O. Levy, Y. Goldberg, and I. Dagan, “Improving dis-

tributional similarity with lessons learned from word

embeddings,” Transactions of the Association for Com-

putational Linguistics, vol. 3, pp. 211–225, 2015,

doi: https://doi.o rg/10 .1162 /tacl _a_00 134. [Online].

Available: https://aclanthology.org/Q15-1016/

54. A. Srivastava et al., “Beyond the imitation game:

Quantifying and extrapolating the capabilities of language

models,” Transactions on Machine Learning Research,

2023, featured Certiﬁcation. [Online]. Available: https:

//openreview.net/forum?id=uyTL5Bvosj

55. E. M. Bender, T. Gebru, A. McMillan-Major, and

S. Shmitchell, “On the dangers of stochastic parrots: Can

language models be too big? ,” in Proceedings of the

2021 ACM Conference on Fairness, Accountability, and

Transparency, ser. FAccT ’21. New York, NY, USA:

Association for Computing Machinery, 2021, p. 610–623,

doi: https://doi.org/10.1145/3442188.3445922.

56. W. Antoun, F. Baly, and H. Hajj, “AraGPT2: Pre-

trained transformer for Arabic language generation,”

in Proceedings of the Sixth Arabic Natural Language

Processing Workshop, N. Habash et al., Eds. Kyiv,

Ukraine (Virtual): Association for Computational Lin-

guistics, Apr. 2021, pp. 196–207. [Online]. Available:

https://aclanthology.org/2021.wanlp-1.21/

BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/798kjf21

Research Article

RESEARCH ARTICLE

Design and Evaluation of an Interpretable Multimodal

Deep Learning Framework for Early Alzheimer’s

Disease Detection

Shehu Mohammed,

1,∗

Neha Malhotra

and Anmol Singh Rai

School of Computer Applications, Lovely Professional University, Phagwara, India and

Shrimann Supersp eciality Hospitals, Jalandhar,

India

∗

Corresponding author. mohammedshehumafara@gmail.com

Received on 22 December 2025; Accepted on 24 June 2026

Abstract

Alzheimer’s disease is a progressive neurodegenerative disorder that signiﬁcantly impairs memory and cognitive functions

and aﬀects over 55 million people worldwide. The successful management and planning require early and accurate diagno-

sis. Conventional radiological assessment is often subjective and time-consuming, which highlights the need for automated

and reliable diagnostic solutions. Most deep learning models show promise for classifying neuroimaging data, but they tend

to be less computationally eﬃcient and less interpretable, and they cannot be integrated into patient-centric processes.

The gap between developing diagnostic algorithms with high accuracy and implementing them in a supportive framework

that includes patients and caregivers is large. This paper introduces a comprehensive, hybrid framework that addresses

these gaps. We present a dual-modality diagnostic system: a deep learning pipeline using EﬃcientNetV2-S for CT scan

classiﬁcation, complemented by a Feedforward Neural Network (FNN) that analyses structured clinical data for holistic

patient assessment. This diagnostic core is integrated into a user-friendly graphical user interface (GUI) and supplemented

by ”NeuroBot,” an AI-powered chatbot that provides domain-speciﬁc information and support. The two models have

been trained using the transfer learning method on a curated dataset of 30,000 brain CT slices. The EﬃcientNetV2-S

model achieved an accuracy of 98.19%. After hyperparameter tuning, the FNN model achieved an optimised accuracy

of 87.21%. The importance of the features addressed by the models was proved with the help of the statistical t-tests of

the corresponding clinical data. The integrated system enables a scalable, translatable, and patient-centered system to

improve the early analysis and treatment of Alzheimer’s disease.

Key words: Alzheimer’s Disease, CT Scan Classiﬁcation, EﬃcientNetV2-S, Deep Learning, AI Chatbot, Medical

Imaging, Transfer Learning

1. Introduction

Alzheimer’s disease (AD) is the most prevalent form of de-

mentia and primarily aﬀects the aging population, accounting

for approximately 60-70% of dementia cases worldwide. It is

a neurodegenerative, advanced, and progressive disorder that

slowly aﬀects the cognitive functions, starting with the mem-

ory and then with the dysfunction of judgment, language, and

behaviour. The disease has a signiﬁcant social impact, placing

substantial emotional and ﬁnancial burdens on patients, fam-

ilies, and healthcare systems worldwide. Early and accurate

diagnosis is therefore critical important. The timely diagnosis

will enable patients and the individuals who provide care to in-

vestigate the treatment interventions, develop useful long-term

care plans, and eventually improve the quality of life of patients

as the disease advances[1, 2].

The neuroimaging procedures are the core of the AD diag-

nostics process, and they allow clinicians to observe structural

changes of the brain typical of the disease, e.g., cortical at-

rophy and ventricular enlargement. Among these techniques,

Computed Tomography (CT) scans serve as a vital tool. CT

imaging is particularly valuable due to its rapid acquisition

time, widespread availability, and relative cost-eﬀectiveness

compared to other modalities like Magnetic Resonance Imaging

(MRI), making it a cornerstone of initial neurological workups,

especially in resource-limited settings[3]. However, the conven-

tional analysis of these scans is not without its challenges. The

diagnostic process is a subjective one and depends on the radiol-

ogist’s interpretation, and this cannot be adequately performed

in a short time, besides being prone to inter-rater inconsistency

and human error, especially in detecting the changes at the

initial stages of AD.

Shehu Mohammed et al.

These diagnostic limitations have been a primary driver for

the development of computer-aided diagnostic (CAD) systems,

a ﬁeld that has been revolutionized by the advent of deep learn-

ing. As extensively documented in systematic reviews, both

machine learning and deep learning techniques have demon-

strated profound success in the automated analysis of medical

images[4]. CNNs have been exceptionally skilled at this task,

in particular. Their architecture allows them to automatically

learn and identify complex, hierarchical patterns within visual

data, enabling the detection of subtle pathological indicators

in MRI and brain CT slices that may be imperceptible to the

human eye[5, 6]. This capability has spurred the creation of

numerous advanced mo dels, including various hybridized deep

learning approaches, all aimed at pushing the boundaries of

diagnostic accuracy and reliability[7].

Despite this remarkable progress, a signiﬁcant gap per-

sists between the development of high-accuracy algorithms in

a research setting and their practical deployment in clinical

workﬂows. Many state-of-the-art CNN mo dels face consider-

able hurdles, including the need for substantial computational

resources, which limits their real-time application. Moreover,

the inherent ”black-box” nature of many deep learning mod-

els can be a barrier to clinical adoption; for a diagnosis to be

trusted, clinicians must be able to understand and verify the

reasoning behind it, a critical factor in diﬀerential diagnosis[8].

Beyond the technical challenges, most existing research has fo-

cused almost exclusively on the diagnostic algorithm itself. This

narrow focus often neglects the broader clinical ecosystem and

the critical need for an integrated, user-friendly system that

not only provides a diagnosis but also supports patients and

caregivers with accessible, contextual information[9].

To overcome these complex issues, this paper proposes a

new and unique hybrid framework that combines the high-

performance diagnostic engine with an interactive patient sup-

port system that is based and comprehensive. Our primary

contribution is the development of a dual-model deep learning

pipeline that leverages a powerful and eﬃcient EﬃcientNetV2-S

architecture for the classiﬁcation of brain CT slices with ex-

ceptional accuracy and computational eﬃciency[10]. The core

novelty of our work lies in embedding these performant and

interpretable models, which utilize Grad-CAM for visual ex-

planations, within a complete, end-to-end ecosystem. This

system includes an intuitive graphical user interface (GUI) for

seamless interaction by clinicians and an AI-powered chatbot,

NeuroBot, designed to answer AD-related queries from pa-

tients and caregivers. By moving beyond mere classiﬁcation,

this holistic approach creates a practical, scalable, and sup-

portive tool designed to enhance early Alzheimer’s detection

and improve the overall standard of patient care[11]. Further-

more, to capture non-visual risk factors and cognitive metrics

that are crucial for a comprehensive diagnosis, our framework

integrates a secondary pathway that leverages a robust FNN to

analyze structured patient clinical records.

Although many deep learning models have achieved high

diagnostic accuracy for Alzheimer’s disease detection, most

studies focus primarily on algorithm development and eval-

uation using experimental datasets. These models are rarely

integrated into user-friendly systems that support clinical work-

ﬂows, patient interaction, or caregiver guidance. As a result,

there remains a gap between high-performance diagnostic al-

gorithms and practical systems that can be deployed in real

clinical environments. The proposed framework addresses this

gap by integrating the diagnostic models within an interac-

tive ecosystem that includes a graphical user interface and

an AI-based assistant to support clinicians, patients, and

caregivers.

The main contributions of this study can be listed as follows:

• Development of a dual-modality deep learning framework,

which uses EﬃcientNetV2-S for CT image classiﬁcation,

and a Feedforward Neural Network for structured clinical

data analysis.

• The application of explainable AI, which uses Grad-CAM

for region highlighting on brain CT slices, thereby improv-

ing their interpretability.

• Development of a user-centric deployment platform, which

includes a GUI and an AI-based chatbot, referred to as

NeuroBot.

• The validation of clinical features through statistical anal-

ysis, which includes independent sample t-tests and corre-

lation analysis to ascertain the signiﬁcance of cognitive and

lifestyle factors, such as MMSE and ADLs.

2. Literature Review

The ML and DL applied to the neurology practice have com-

pletely changed the Alzheimer Disease (AD) diagnostic process

and provided meaningful and data-intuitive answers to the cur-

rent practices. The evolution of these techniques has been

rapid, moving from foundational models to highly sophisti-

cated, specialized architectures. Early research in automated

AD detection was primarily centered on classical machine

learning algorithms. Models such as Support Vector Machines

(SVMs), Random Forests, and Decision Trees were applied to

neuroimaging data, but their eﬃcacy was often constrained by

a reliance on handcrafted feature extraction[12]. This process,

which required extensive domain knowledge to manually deﬁne

and extract relevant features like hippocampal volume or corti-

cal thickness, was not only labor-intensive but also limited the

models’ ability to discover novel, complex patterns within the

data. Despite the need to take such measures, the subsequent

emergence of deep learning, and, particularly, the Convolu-

tional Neural Networks (CNNs) became a game-changer in the

matter. The CNNs have revolutionised the study of medical

images; they enable end-to-end learning of hierarchical fea-

tures beneath the raw pixel data. This capability has led to

a proliferation of studies demonstrating the accurate predic-

tion and diagnosis of AD using a variety of deep learning

models, which consistently outperform their traditional ML

counterparts[13, 14].

The current research in the area, however, has predom-

inantly been interested in the application of deep learning

to analyze neuroimaging images, and, more speciﬁcally, in

Magnetic Resonance Imaging (MRI) and, more recently, in

Computed Tomography (CT) scans. The authors have experi-

mentally shown that applying advanced image processing and

enhancement techniques before model training improves AD de-

tection accuracy[15]. Recognizing that a single data source may

not capture the complexity of AD pathology, numerous stud-

ies have investigated multimodal deep learning methods. These

models integrate data from diﬀerent neuroimaging mo dalities

(e.g., structural MRI, functional MRI, and PET scans) to create

a more comprehensive and robust diagnostic picture[16]. The

sheer volume of research has produced a rich body of literature

that systematically reviews the diverse array of AD detection

techniques, highlighting the consistent and rapid progress in di-

agnostic accuracy and model sophistication[17]. Furthermore,

the application of these powerful models has expanded beyond

AD to include the diagnosis of related neurodegenerative condi-

tions, such as tauopathies, thereby demonstrating their broader

versatility and clinical potential[18].

Since the ﬁeld has become mature, the quest to achieve

increasingly high accuracy has given rise to the emergence

of new and complicated model architectures. Hybrid systems

and ensemble-based systems are emerging to the fore. These

advanced approaches combine the predictive strengths of mul-

tiple deep learning models to create more robust, reliable, and

generalizable identiﬁcation systems that are less prone to the

biases of a single model[19, 20]. While neuroimaging remains

the primary data source, some innovative models have been de-

signed to diagnose AD based exclusively on structured patient

clinical records, providing a valuable alternative or comple-

mentary diagnostic pathway that does not require imaging[21].

This has contributed to the development of sophisticated ar-

chitectures capable not only of binary classiﬁcation but also of

detecting and diﬀerentiating between the various stages of AD,

from early-stage Mild Cognitive Impairment (MCI) to advanced

dementia, which is crucial for tailoring patient care[22, 23].

For these powerful deep learning models to transition from

research laboratories to real-world clinical environments, two

practical factors have become paramount: computational eﬃ-

ciency and model interpretability. Clinically viable diagnostic

tools have been developed to continue to focus on the trans-

fer learning application. The ﬁne-tuning of large models, which

have been trained on general image datasets, on smaller medical

imaging datasets is a very eﬀective way to achieve state-of-the-

art performance with small datasets[24]. This strategy has been

a central theme in the broader investigation of deep learning

for enhancing early detection and supporting clinical decision-

making[25, 26]. Consequently, a signiﬁcant portion of modern

research is focused on developing models for early detection, as

this is the stage at which interventions are most likely to be

eﬀective[27]. To further improve performance, some researchers

have proposed integrative models that combine the strengths of

traditional ML with advanced deep learning architectures[28].

Finally, to address the ”black-box” problem, there is a growing

emphasis on creating explainable AI (XAI). The development of

specialized, attention-based explainable networks and custom

models, such as ADD-Net, underscores the ﬁeld’s commit-

ment to creating diagnostic tools that are not only accurate

and eﬃcient but also transparent and trustworthy for clinical

use[29, 30].

Although many researchers have reported positive results

in detecting Alzheimer’s disease using deep learning mo dels

in their literature, a wide range of methodological variations

can be seen in their approaches. Most of the existing liter-

ature has focused on MRI-based deep learning models using

CNN-based architectures and has achieved good accuracy us-

ing ADNI databases, ranging from 85% to 96%. However, a

few researchers have also used multimodal data for the ac-

curate prediction of Alzheimer’s disease using MRI and PET

scans. However, these techniques require expensive hardware.

In addition, fewer researchers have focused on CT-based deep

learning models for detecting Alzheimer’s disease. Moreover,

few researchers have focused only on prediction models without

using interpretability mechanisms or deployment frameworks.

Although various researchers have used various models, such as

ADD-Net and attention-based models, to explain their models,

these models are not used in deployment frameworks. However,

in the proposed model, a CT-based deep learning model was

used in conjunction with a structured clinical data model, an

explainable AI model using a Grad-CAM mechanism, and a

user-centric deployment model using a GUI and an AI chatbot

support system[31].

3. Methodology and Experiment

The proposed framework integrates a dual-model deep learning

pipeline for AD classiﬁcation with an interactive user interface

and an AI chatbot. The architecture is designed for accuracy,

interpretability, and user engagement.

3.1. Dataset and System Architecture

The general pro cedure of our system is illustrated in Fig-

ure 1. The training and evaluation dataset consists of 10,240

brain CT slices, evenly distributed b etween Alzheimer’s and

non-demented cases. In addition, a clinical dataset containing

2,149 patient records was used for structured data analysis.

Figure 2 presents the distribution of diagnoses in the clini-

cal dataset, with 64.6% classiﬁed as Non-Demented and 35.4%

as Demented. All the images were resized to an average of

224 × 224 pixels and normalised. Random rotations, horizontal

ﬂips, and zooming were used as data augmentation techniques

to improve the model’s robustness.

Figure 1. Alzheimer’s Detection System Architecture.

A stringent, automated preprocessing pipeline was applied

to the clinical data to prepare it for the FNN model. The ﬁrst

split of the features was into numerical (e.g., Age, MMSE) and

categorical ones (e.g., Gender, Smoking status). One-hot en-

coding was then applied to all categorical features in order to

put them in numerical form, and all numerical features were

rescaled to a normalized range. This ensures that all features

of the corresponding role in the prediction of the model are not

aﬀected by diﬀerent scales. All this transformation process was

stored as a single ob ject of pipeline in order to make sure that

Shehu Mohammed et al.

Figure 2. Diagnosis Distribution in the Clinical Dataset.

whenever new information is keyed in the prediction process, it

would be processed similarly to the training information.

The CT imaging dataset used in this study consisted of ap-

proximately 30,000 brain CT slices derived from the Shrimaan

Super Specialty Hospital repository and clinical datasets for

Alzheimer’s disease research. Of these, 10,240 CT slices were

selected for training and evaluation purposes. The images were

divided into training, validation, and testing subsets using a

70–10–20 split. In addition, an extended testing p ool was used

during the ﬁnal evaluation, resulting in approximately 4,500

CT slices used for performance assessment. Each CT slice in the

dataset was linked to a sub ject-level diagnosis label: Demented

or Non-Demented.

To avoid data leakage in the model, the clinical dataset was

split at the patient level instead of the slice level. This implies

that brain CT slices for a speciﬁc patient were only included

in either the training or testing dataset. Finally, the clinical

dataset was split into 70% for training, 10% for validation, and

20% for testing.

The clinical dataset included 2,149 patient records, which

included various physiological measures, cognitive measures

such as Mini-Mental State Examination (MMSE), and lifestyle

factors. For handling missing values in the clinical dataset, me-

dian imputation was applied for numerical features and most

frequent imputation for categorical features. This particular

process guaranteed the proper transformation of both training

and testing data.

3.2. System Deployment Architecture

The framework for deployment is modular, designed to function

in real-world environments. It has four components. The ﬁrst

is the user interface, the second is the inference engine, the

third is the data management system, and the fourth is the AI

chatbot module.

The primary interface with the system is the Graphical User

Interface (GUI). In the web-based system, the user can input

the CT scan of the brain or the clinical parameters. This is

then sent to the model inference server, which then uses the

EﬃcientNetV2-S and the FNN models to perform the predic-

tion. The EﬃcientNetV2-S is used to perform the prediction

with the CT scan, while the FNN is used with the clinical

parameters.

There is also an AI chatbot mo dule that interacts with the

user in natural language to ask questions and educate the user

about Alzheimer’s disease. This module is known as the Neu-

roBot. The backend is designed to support an anonymized

clinical dataset and to perform inference requests with the

models.

The System deployment architecture of the proposed

Alzheimer’s detection framework is illustrated in Figure 3.

The system integrates a web-based user interface with back-

end APIs, deep learning inference models (EﬃcientNetV2-S for

brain CT slices and FNN for clinical data), and a chatbot mod-

ule to support clinical decision-making and patient interaction.

Figure 3. System Deployment Architecture Diagram.

3.3. Diagnostic Model Architectures

To create a comprehensive diagnostic tool, we employ two types

of neural nets: Convolutional Neural Networks (CNNs) to pro-

cess image data and Feedforward Neural Networks (FNNs) to

process data based on clinical features.

3.3.1. Architectures for Image-Based Analysis (CNNs)

The fundamental operation in a CNN is the two-dimensional

convolution. Convolution is a method of feature extraction, in

which a kernel is applied to a given input image or feature map.

This is arithmetically measured as:

O(i, j) = (I ∗ K )(i, j) =

I(i − m, j − n) · K(m, n) (1)

O(i, j) characterises the output feature map of location (i, j),

I is the input, and K is the convolutional kernel. Activation

functionality is applied to add non-linearity. Our application is

the ReLU, which is:

f(x) = max(0, x) (2)

For the image classiﬁcation task, this study utilizes the

EﬃcientNetV2-S architecture, a powerful and computationally

eﬃcient model. This model achieves its eﬃciency through the

use of Fused-MBConv blocks in its early layers.

Fused-MBConv blo ck simpliﬁes the normal inverted resid-

ual block by fusing the ﬁrst 3 × 3 depthwise convolution of the

block with the second 1 × 1 projection convolution into a stan-

dard 3 × 3 convolution block. This minimizes the production

of memory access overhead increases and enhances the training

speed on advanced accelerators. The work of a Fused-MBConv

block could be explained by the equation:

y = BN (Conv

3×3

(BN (Conv

1×1

(x)))) + x (3)

In this case, x is a signal sent into the block, and the equa-

tion shows the series of a 1 × 1 expansion convolution, a 3 × 3

standard convolution (a replacement of the individual depth-

wise and projection steps) and Batch Normalization (BN). The

end result y is gotten by summing the original input x via a

residual (skip) connection, which makes this a distinguishing

feature that makes it likely to train very deep networks.

3.3.2. Architecture for Clinical Data Analysis (FNN)

A Feedforward Neural Network (FNN) was developed to ana-

lyze structured clinical data. This kind of network was trained

using a hyperparameter optimization strategy that entailed the

optimization of the predictive performance using the Optuna

framework.

The FNN is made up of a series of dense (fully connected)

layers. The general unit is its dense layer, which is a linear

transform of its input vector x, which can be expressed as:

y = Wx + b (4)

The input is represented by y , the learnable weight matrix is

W , and the learnable bias is b.

The result of each of the dense layers is then run through

the activation function of the Rectiﬁed Linear Unit (ReLU) to

introduce non-linearity, as well as permit the model to learn

more intricate patterns, and is deﬁned as:

f(x) = max(0, x) (5)

Where γ and β are learnt scale and shift parameters, and the

constant ϵ is a small number to maintain numerical stability.

This is followed by the normalized output going through the

(ReLU) activation function.

To avoid overﬁtting, a ReLU activation will be followed by

a Dropout layer. Dropout is random and assigns a part of the

input units to 0 at every update time throughout the training

period, which assists in augmenting the strength of the network.

The optimized architecture that was discovered by hy-

perparameterOptimization is an input layer and two hidden

layers:

1. The initial hidden layer is a dense layer that contains 57

units, an activation of ReLU, and a Dropout with a rate

of 0.38.

2. The second hidden layer is a dense layer with 129 units,

and the next layer consists of a ReLU activation and a

Dropout layer at a rate of 0.49.

The concluding component is a single output neuron that

produces a raw logit, z. This logit is then transformed into

a probability using the sigmoid activation function. For im-

proved numerical stability during training, this function is

integrated directly into the loss function (BCEWithLogitsLoss).

The sigmoid function is deﬁned as:

σ(z) =

1 + e

−z

(6)

3.4. Interactive Framework and Training Algorithm

The project’s workﬂow, executed within a Jupyter Noteb ook,

provides two distinct diagnostic pathways: one for image-based

analysis and another for clinical feature-based analysis. The

classiﬁcation pipeline for the image-based pathway is detailed

in Algorithm 1, while the corresponding process for the clinical

data FNN model is outlined in Algorithm 2.

Algorithm 1 AD Classiﬁcation Pipeline (Image-Based CNN)

1: Input: Raw brain CT image I

raw

2: Data Loading: The image dataset is loaded from the

data/image directory and split into training and validation

sets.

3: Preprocessing & Augmentation:

4: I

resized

⇐ RandomResizedCrop(I

raw

, (224, 224))

5: I

flipped

⇐ RandomHorizontalFlip(I

resized

)

6: I

tensor

⇐ ToTensor(I

flipped

)

7: I

norm

⇐ Normalize(I

tensor

)

8: Model Loading: Load pre-trained mo del M

(EﬃcientNetV2-S) and adapt its ﬁnal layer for binary

classiﬁcation.

9: Training: The model is ﬁne-tuned on the training set using

an Adam optimizer and Binary Cross-Entropy with Logits

loss (BCEWithLogitsLoss).

10: Prediction:

11: Z ⇐ M(I

norm

) // Get raw logit output from the CNN

12: Classiﬁcation:

13: P ⇐ σ(Z) // Apply sigmoid function to get probability

14: If P > 0.5, then C ⇐ ’Demented’

15: Else C ⇐ ’Non-Demented’

16: Output: Return class label C.

Shehu Mohammed et al.

Algorithm 2 Clinical Data Classiﬁcation Pipeline (FNN)

1: Input: Raw clinical feature set D

raw

from

alzheimers disease data.csv.

2: Data Splitting: The dataset is split into training and testing

sets (80/20 split).

3: Preprocessing:

4: A ColumnTransformer pipeline is ﬁtted on the training

set:

5: Numerical features are imputed (median) and stan-

dardized (StandardScaler).

6: Categorical features are imputed (most frequent) and

one-hot encoded (OneHotEncoder).

7: D

train processed

⇐ Apply ﬁtted pipeline to training data.

8: D

test processed

⇐ Apply ﬁtted pip eline to test data.

9: Class Imbalance Handling:

10: D

train

resampled

⇐ Apply SMOTE (Synthetic Minority

Over-sampling Technique) to D

train

processed

to balance the

classes.

11: Model Loading: Load the pre-trained Feedforward Neural

Network (FNN).

12: Tensor Conversion:

13: T

input

⇐ ToTensor(D

test

processed

)

14: Prediction:

15: Z ⇐ M

fnn

input

) // Get raw logit output from the FNN

16: Classiﬁcation:

17: P ⇐ σ(Z) // Apply sigmoid function to get probability

18: If P > 0.5, then C ⇐ ’Demented’

19: Else C ⇐ ’Non-Demented’

20: Output: Return class label C.

4. Results and Discussion

The eﬃcacy of the hybrid framework that was suggested was

critically evaluated using the aid of an integrated approach.

This was accompanied by a quantitative measure of the classiﬁ-

cation performance of the deep learning models using a held-out

test set of 4,500 brain CT slices, a large-scale measure of the

associated clinical data to validate the validity of the features of

the models, as well as a qualitative measure of the user interface

and the interactivity aspects.

4.1. Performance Metrics

We used quantitative measures of the models using a sam-

ple set of standard classiﬁcation measures that is decided by

the elements of the confusion matrix: True Positives (TP),

True Negatives (TN), False Positives (FP), and False Negatives

(FN).

• Accuracy: The percentage of all the correct predictions.

Accuracy =

T P + T N

T P + T N + FP + F N

(7)

• Precision: The proportion of positive predictions that were

actually correct.

P recision =

T P

T P + F P

(8)

• Recall (Sensitivity): The proportion of actual positives

that were identiﬁed correctly.

Recall =

T P

T P + F N

(9)

• F1-Score: The harmonic mean of Precision and Recall,

providing a single score that balances both.

F 1-Score =

2 · P recision · Recall

P recision + Recall

(10)

4.2. Model Interpretability and Performance

Evaluation

Besides the model’s accuracy, interpretability is another critical

requirement for clinical decision-support system. In this study,

the interpretability of the model was facilitated by the use of a

technique called Gradient-weighted Class Activation Mapping

(Grad-CAM) in the EﬃcientNetV2-S CNN model. This tech-

nique enables the production of visual maps that highlight areas

of the images used in the model that are being used to make a

prediction, as illustrated in Figure 4. These maps allow veriﬁ-

cation that the model focuses on anatomically relevant regions

associated with Alzheimer’s disease, such as cortical atrophy

and ventricular enlargement.

Although Grad-CAM visualizations, statistical validation,

and feature-imp ortance analysis provide valuable insights into

the model’s decision-making process. Future work will investi-

gate quantitative explainability metrics, including localization-

based evaluation and clinician-assisted validation of explana-

tion maps, to provide a more rigorous assessment of model

interpretability and clinical trustworthiness.

Moreover, the interpretability of the clinical data pathway

was facilitated by the use of statistical tests, such as indepen-

dent sample t-tests and correlation tests, which showed that

clinical variables, such as MMSE, Activities of Daily Living

(ADL), and functional assessment, signiﬁcantly vary across the

diﬀerent diagnostic groups.

Figure 4. Grad-CAM visualization highlighting imp ortant regions in

brain CT slices used by the EﬃcientNetV2-S model for Alzheimer’s

disease classiﬁcation. The highlighted areas indicate regions contributing

most strongly to the mo del’s prediction.

The EﬃcientNetV2-S model achieved the b est validation

performance with an accuracy of 98.19% compared to other

architectures, compared with other evaluated architectures,

including EﬃcientNetV2-B0, ResNet50V2, and VGG16, as in-

dicated in Table 1. Figure 5 shows the visualization of the

performance of these models in comparison with each other, and

it is evident that the EﬃcientNetV2-S model is more accurate.

Figure 6 displays the training and validation history of the

EﬃcientNetV2-S model and depicts the change in the accu-

racy and loss after 10 epochs. This trend is convergently stable,

and the trend has small overﬁtting, which is an attribute that

indicates good ﬁne-tuning and regularization.

Model Accuracy (%)

EﬃcientNetV2-S 98.19%

EﬃcientNetV2-B0 95.8%

ResNet50V2 94.9%

VGG16 91.2%

DenseNet121 90.7%

MobileNetV2 89.3%

Table 1. Performance Comparison of Deep Learning Models

Figure 5. Visualization of Performance Comparison of All Models.

Figure 6. Training and Validation History for the EﬃcientNetV2-S

Model, showing Loss and Accuracy over 10 epo chs.

As a method of measuring the reliability of classiﬁcations,

a confusion matrix of the EﬃcientNetV2-S model was con-

structed and presented in Figure 7. The matrix shows the

accuracy of the model and recall of both the Demented and

Non-Demented classes, that provide us with insight into the

discriminatory ability of the model.

The FNN model, after undergoing hyperparameter tuning

with Optuna, achieved a ﬁnal test accuracy of 87.21% on the

clinical dataset. This result validates the eﬀectiveness of using

structured clinical data for prediction and highlights the beneﬁt

of automated hyperparameter optimization.

In the dual-modality framework, there is the incorporation

of medical images and clinical data, which makes the diagno-

sis even better. The CNN examines the images, digging deeper

into the changes seen on the CT scan, such as cortical atrophy

and enlarged ventricles, which indicate neurodegeneration. On

the other hand, the FNN examines the clinical data, which in-

cludes cognitive tests such as MMSE, daily living activities, and

other relevant demographic factors that may be associated with

the condition. Statistical analysis reveals that MMSE, ADL,

and other functional measures vary signiﬁcantly across diﬀer-

ent diagnostic groups. The dual-modality framework, therefore,

examines both visible changes on the brain images and the pa-

tient’s individual risk factors, making the diagnosis even better

and more reliable.

Figure 7. Confusion Matrix for EﬃcientNetV2-S on the Test Set.

To further evaluate the eﬀectiveness of the proposed frame-

work, its performance was compared with several recent state-

of-the-art approaches for Alzheimer’s detection reported in the

literature. These studies employed various deep learning archi-

tectures and imaging modalities, including MRI-based convolu-

tional neural networks and multimodal models. The comparison

result presented in Table 2 demonstrates that the proposed

framework achieves competitive performance while integrating

CT imaging, structured clinical data, and an interactive clinical

support system.

It should be noted that direct comparison of classiﬁcation

accuracies across studies should be interpreted with caution be-

cause the evaluated datasets, imaging modalities, sample sizes,

and experimental protocols diﬀer substantially. Most of the

compared studies employed MRI-based datasets, particularly

ADNI, whereas the prop osed framework was developed using

brain CT images and structured clinical data obtained from a

hospital-based cohort. MRI generally provides higher soft-tissue

contrast than CT imaging, which may inﬂuence diagnostic

performance, Furthermore, variations in dataset composition,

preprocessing procedures, and evaluation strategies can aﬀect

reported accuracies. Despite these diﬀerences, the proposed

framework achieved competitive performance while simulta-

neously providing explainability through Grad-CAM and a

deployment-oriented clinical support platform.

4.3. User Interface and Syste m Interaction

The practical utility of the diagnostic models is realized through

an intuitive and user-centric graphical user interface (GUI). The

system’s main dashboard, shown in Figure 8, provides a clean

and accessible entry point for users. It presents two primary

options for analysis, feature-based (clinical data) and image-

based, alongside the integrated ”NeuroBot” AI assistant.

For a feature-based analysis, the user navigates to a com-

prehensive data input form (Table 3), where they can enter

demographic and clinical parameters. After submission, the sys-

tem presents a detailed results page (Figure 9) that not only

displays the diagnosis (’Demented’ or ’Non-Demented’) but also

provides actionable suggestions and precautions tailored to the

outcome. This aspect turns the system into a supportive system

rather than a basic classiﬁer. For an image-based diagnosis, the

Shehu Mohammed et al.

Study Data Modality Model / Method Dataset Accuracy (%)

Pradhan et al., 2024 MRI DenseNet + ResNet50 ADNI 95.3

Helaly et al., 2022 MRI CNN-based Deep Learning ADNI 93.6

Saleem et al., 2022 MRI Transfer Learning CNN ADNI 94.7

Ayus & Gupta, 2024 MRI Hybrid Ensemble DL ADNI 96.2

Alwakid et al., 2024 MRI Image Processing + CNN ADNI 94.5

Avila-Jim´enez et al., 2024 Clinical Records Deep Learning Mo del Clinical Dataset 85.0

Proposed Method CT + Clinical Data EﬃcientNetV2-S + FNN Hospital + Clinical Dataset 98.19

Table 2. Comparison of the proposed method with recent state-of-the-art Alzheimer’s disease detection approaches.

Figure 8. Main Dashboard Interface with Chatbot Panel.

user interacts with a simple drag-and-drop interface to upload

a brain CT scan (Figure 10). The system processes the image

and returns a clear results page displaying the diagnosis, a con-

ﬁdence score, and preventive measures or daily activities that

may be beneﬁcial (Figure 11). This consistent, informative clin-

ical workﬂow is formulated to be easily ventured by clinicians,

patients, and even caregivers, and this gap b etween a sophis-

ticated AI model and a practical and usable tool is addressed.

4.4. Statistical Validation and Feature Analysis

To oﬀer a sound clinical basis to our models, the struc-

tured clinical dataset underwent a complete statistical analysis.

This validation guarantees that the characteristics learnt by

the models are associated with accepted clinical markers of

Alzheimer’s disease.

This was done by the use of an independent samples t-test

to identify the clinical characteristics that signiﬁcantly diﬀered

between the Demented and Non-Demented groups. Figure 12

gives the general ﬁndings of the t-test in a manner that sum-

marizes the comparative statistics of the two diagnostic groups.

In the analysis, it was found that the signiﬁcant indicators that

constitute an overwhelming proportion of the signiﬁcant indi-

cators were the MMSE score, and that the p-value was 0.0000.

These diﬀerences in distribution between the Age and MMSE

age scores, as represented graphically in Figure 13 and Fig-

ure 14 of the boxplot is a very sharp and signiﬁcant drop in

MMSE scores in the demented group compared with Age.

The correlation heatmap in Figure 15 also shows correlations

between major clinical characteristics and the ﬁnal diagnosis,

indicating that the MMSE score is most strongly negatively

correlated with dementia diagnosis.

Most importantly, all these statistics are directly related to

the feature-importance analysis of the FNN model presented

in Figure 16. These characteristics, the FunctionalAssessment,

ADL (Activities of Daily Living), MemoryComplaints and

MMSE, were evaluated as the most eﬀective predictors for the

model. Such a high concordance rate assures that the FNN

model evidently learns to focus on the clinically signiﬁcant

variables associated with Alzheimer’s disease.

In addition to the visual explanations provided by Grad-

CAM, quantitative evidence supporting model interpretabil-

ity was obtained through statistical validation of the clinical

variables. Independent sample t-tests demonstrated signiﬁ-

cant diﬀerences between demented and Non-demented groups,

with MMSE exhibiting a highly signiﬁcant association (p <

0.001). Correlation analysis further conﬁrmed strong rela-

tionships between important clinical features and dementia

Feature Value Feature Value

Age 68 SystolicBP 118

Gender 0 DiastolicBP 78

Ethnicity 0 CholesterolTotal 180

EducationLevel 3 CholesterolLDL 100

BMI 24.5 CholesterolHDL 65

Smoking 0 CholesterolTriglycerides 130

AlcoholConsumption 2 MMSE 29

PhysicalActivity 56 FunctionalAssessment 1.0

DietQuality 9 MemoryComplaints 0

SleepQuality 8 BehavioralProblems 0

FamilyHistoryAlzheimers 0 ADL 1.0

CardiovascularDisease 0 Confusion 0

Diabetes 0 Disorientation 0

Depression 0 Personality Changes 0

HeadInjury 0 Diﬃculty Completing Tasks 0

Hypertension 0 Forgetfulness 1

Doctor In Charge 221

Table 3. Sample Input Values for Feature-Based Analysis.

Figure 9. Analysis Results Page for Clinical Data Submission.

diagnosis. Furthermore, feature-importance analysis identiﬁed

MMSE, Functional Assessment, ADL, and Memory Complaints

as the most inﬂuential used by the FNN model. The agreement

between statistical signiﬁcance and model-derived feature im-

portance provides quantitative evidence that the model focuses

on clinically meaningful biomarkers associated with Alzheimer’s

disease.

4.5. User Engagement Framework

The ﬁnal component of our evaluation focused on the AI-

powered chatbot, NeuroBot, which serves as the primary to ol

for patient and caregiver support. The chatbot was assessed

across ﬁve key criteria: Domain Relevance, Medical Accuracy,

Politeness and Safety, Responsiveness, and its ability to refuse

out-of-scope questions. As illustrated in the radar chart in

Figure 17, NeuroBot scored perfectly or near-perfectly on all

measures. This lends credibility to its accuracy, safety, and rel-

evancy in delivering pertinent information within the ﬁeld of

Alzheimer Disease. Also, response time analysis revealed that

the response to user queries was received in time (85% of queries

were attended to within a range of 4.2 to 4.6 seconds), which

guaranteed a consistent and interactive user experience.

5. Conclusion and Future Work

This study developed and evaluated a dual-modality framework

for the early detection of Alzheimer’s disease, where the deep

learning pipeline is applied to the CT image processing and the

hyperparameter-optimized Feedforward Neural Network (FNN)

is applied to the clinical data processing.

The ﬁndings demonstrate that using a ﬁne-tuned Eﬃ-

cientNetV2-S architecture for image analysis can achieve high

diagnostic accuracy, reaching a peak validation performance

of 98.19%. Complementing this, the FNN mo del, optimized

through systematic hyperparameter tuning, achieved a ﬁnal

accuracy of 87.21% on structured clinical data. The statisti-

cal analysis of this clinical data further solidiﬁed our approach,

conﬁrming that the models are learning from features, such as

the MMSE score, that are strongly correlated with established

indicators of cognitive decline.

One of the ma jor contributions that this research makes is

the holistic approach. When putting these diagnostic models

into a broader ecosystem, which includes a user-friendly inter-

face and the AI-p owered NeuroBot, this piece of work oﬀers a

blueprint of an all-inclusive support platform. This usability-

based strategy is also essential in translating the gap between

the complicated AI technology and the daily clinical practice,

Shehu Mohammed et al.

Figure 10. Drag-and-Drop Interface for Image-Based Analysis.

Figure 11. Analysis Results Page for Image-Based Diagnosis.

and allows patients, caregivers, and clinicians to place the right

diagnostics and information right at their ﬁngertips.

The proposed framework oﬀers promising avenues for future

development and presents several opp ortunities for improve-

ment. The capabilities of the diagnostic functions will be

expanded to provide a more comprehensive portrait of the pa-

tient, with a priority on the integration of multimodal data from

MRI scans and genetic markers. We will also make the existing

binary classiﬁcation models extendable so that Alzheimer’s can

be classiﬁed in multiple steps, so that we are in a position to

further distinguish between MCI and the onset of further stages

of the disease.

The last and most important stage will be the push of the

framework into clinical validation by means of large-scale trials.

This will be essential for validating the system’s performance

across diverse populations and is a necessary step for its even-

tual integration into standard clinical workﬂows. Concurrently,

we will explore the optimization of the models for deployment

on edge devices and expand the chatbot’s capabilities to include

summarizing prediction results and oﬀering multilingual sup-

port, thereby increasing the system’s accessibility and global

impact.

It should be noted that the above evaluation was carried out

using controlled experimental data. Although the mo dels were

Figure 12. Visualization of T-Test Results Comparing Demented vs. Non-Demented Groups.

Figure 13. Boxplot of Age Distribution by Diagnosis.

validated using a held-out test set, the models were not vali-

dated using any independent clinical data sets. In the future,

the focus will be on evaluating the framework using data sets

from multiple institutions and clinical settings.

Shehu Mohammed et al.

Figure 14. MMSE Score Distribution by Diagnosis (T-Test Validation).

Figure 15. Correlation Heatmap of Top Clinical Features

Figure 16. Top 10 Most Important Features.

Figure 17. Radar Chart Evaluation of the Chatbot.

Shehu Mohammed et al.

Ethical Statement

Ethical approval of this study was obtained from the Insti-

tutional Ethics Committee of Lovely Professional University,

India (Ref: LPU/IEC-LPU/2025/1/2, dated 15 February 2025).

Permission to access clinical and imaging data was granted

by Shrimaan Superspeciality Hospital, Jalandhar, India. The

study was retrospective and involved analysis of previously col-

lected anonymised clinical and brain CT imaging data. No

direct patient contact occurred, and no personally identiﬁable

information was accessed. In accordance with the Institutional

Ethics Committee clariﬁcation, the requirement for informed

consent was waived/not applicable.

This study employed anonymized brain images from the

brain CT slices of patients at Shrimann Superspeciality Hospi-

tal. No personally identiﬁable patient information was accessed

during this study. The data were used strictly for academic

research in accordance with the hospital’s data privacy regula-

tions. The proposed framework is intended for use as a clinical

decision supp ort tool. However, it is not intended for use as

a substitute for clinical expertise. The NeuroBot chatbot pro-

vides general information guidance only and does not replace

professional advice.

Funding

The scholar was sponsored by Tertiary Education Trust Fund

(TETFUND) Nigeria for Higher education Studies only.

Declaration of competing interests

The authors declare that they have no known competing ﬁnan-

cial interests or personal relationships that could have appeared

to inﬂuence the work reported in this paper.

Data Availability Statements

The data supporting the ﬁndings of this study are available

from the corresp onding author upon reasonable request. How-

ever, the data are not publicly available due to privacy or

ethical restrictions.

Credit authorship contribution statement

Shehu Mohammed: Conceptualization; Project Administration;

Methodology; Data Curation; Software Development; Investi-

gation; Writing – Original Draft Preparation.

Neha Malhotra: Supervision; Writing – Review & Editing; For-

mal Analysis.

Anmol Singh Rai: Resources; Validation; Visualization.

References

1. S. Alinsaif, “Dca-enhanced alzheimer’s detection with

shearlet and deep learning integration,” Computers in

Biology and Medicine, vol. 185, p. 109538, 2025, doi:

https://doi.org/10.1016/j.compbiomed.2024.109538.

2. M. Srikanth and P. Bellapukonda, “The early detection of

alzheimer’s illness using machine learning and deep learning

algorithms,” Journal of Pharmaceutical Negative Results,

vol. 13, no. 9, pp. 4852–4859, 2022, doi: https://doi.org/

10.47750/pnr.2022.13.S09.603. [Online]. Available: https://

www.pnrjournal.com/index.php/home/article/view/4470

3. E. M. Mohammed, A. M. Fakhrudeen, and O. Y. Alani,

“Detection of alzheimer’s disease using deep learning

models: A systematic literature review,” Informatics

in Medicine Unlocked, vol. 50, p. 101551, 2024, doi:

https://doi.org/10.1016/j.imu.2024.101551.

4. A. D. Arya et al., “A systematic review on machine

learning and deep learning techniques in the eﬀective

diagnosis of alzheimer’s disease,” Brain Informatics,

vol. 10, no. 1, p. 17, 2023, doi: https://doi.org/10.1186/

s40708-023- 00195-7.

5. N. Pradhan, S. Sagar, and A. S. Singh, “Analysis

of mri image data for alzheimer disease detection

using deep learning techniques,” Multimedia Tools and

Applications, vol. 83, no. 6, pp. 17 729–17 752, 2024, doi:

https://doi.org/10.1007/s11042-023-16256-2.

6. P. Kishore, C. U. Kumari, M. Kumar, and T. Pavani,

“Detection and analysis of alzheimer’s disease using

various machine learning algorithms,” Materials today:

proceedings, vol. 45, pp. 1502–1508, 2021, doi: https:

//doi.org/10.1016/j.matpr.2020.07.645.

7. P. Balaji, M. A. Chaurasia, S. M. Bilfaqih, A. Muniasamy,

and L. E. G. Alsid, “Hybridized deep learning approach

for detecting alzheimer’s disease,” Biomedicines, vol. 11,

no. 1, p. 149, 2023, doi: https://doi.org/10.3390/

biomedicines11010149.

8. S. Mirabian, F. Mohammadian, Z. Ganji, H. Zare, and

E. Hasanpour Khalesi, “The potential role of machine

learning and deep learning in diﬀerential diagnosis of

alzheimer’s disease and ftd using imaging biomarkers:

A review,” The Neuroradiology Journal, vol. 38,

no. 5, pp. 571–587, 2025, doi: https://doi.org/10.1177/

19714009251313511.

9. J. Chua et al., “Utilizing deep learning to predict

alzheimer’s disease and mild cognitive impairment with

optical coherence tomography,” Alzheimer’s & Dementia:

Diagnosis, Assessment & Disease Monitoring, vol. 17,

no. 1, p. e70041, 2025, doi: https://doi.org/10.1002/dad2.

70041.

10. A. A. A. El-Latif, S. A. Chelloug, M. Alabdulhaﬁth,

and M. Hammad, “Accurate detection of alzheimer’s

disease using lightweight deep learning model on mri

data,” Diagnostics, vol. 13, no. 7, p. 1216, 2023, doi:

https://doi.org/10.3390/diagnostics13071216.

11. K. Lokesh, N. P. Challa, A. S. Satwik, J. C. Kiran, N. K.

Rao, and B. Naseeba, “Early alzheimer’s disease detection

using deep learning,” EAI Endorsed Transactions on

Pervasive Health and Technology, vol. 9, 2023, doi:

https://doi.org/10.4108/eetpht.9.3966.

12. S. Balne and A. Elumalai, “Machine learning and deep

learning algorithms used to diagnosis of alzheimer’s,”

Materials Today: Proceedings, vol. 47, pp. 5151–5156,

2021, doi: https://doi.org/10.1016/j.matpr.2021.05.499.

13. S. Baskar, M. L. Prasad, N. Sharma, T. Katale, P. C. S.

Reddy et al., “An accurate prediction and diagnosis

of alzheimer’s disease using deep learning,” in 2023

IEEE North Karnataka Subsection Flagship International

Conference (NKCon). IEEE, 2023, pp. 1–7, doi:

https://doi.org/10.1109/nkcon59507.2023.10396132.

14. A. Muydinov, “Advances in deep learning techniques for

alzheimer’s disease detection using mri images review,” in

Proceedings of the 8th International Conference on Future

Networks & Distributed Systems, 2024, pp. 266–269, doi:

https://doi.org/10.1145/3726122.3726160.

15. G. N. Alwakid, S. Tahir, M. Humayun, and W. Gouda,

“Improving alzheimer’s detection with deep learning and

image processing techniques,” IEEE Access, vol. 12,

pp. 153 445–153456, 2024, doi: https://doi.org/10.1109/

ACCESS.2024.3481238.

16. G. S. Chhabra, A. Guru, B. J. Rajput, L. Dewangan,

and S. K. Swarnkar, “Multimodal neuroimaging for

early alzheimer’s detection: a deep learning approach,”

in 2023 14th international conference on computing

communication and networking technologies (ICCCNT).

IEEE, 2023, pp. 1–5, doi: https://doi.org/10.1109/

ICCCNT56998.2023.10307780.

17. S. Afzal et al., “Alzheimer disease detection techniques and

methods: A review.” International journal of interactive

multimedia and artiﬁcial intelligence, vol. 6, no. 7, pp.

26–38, 2021, doi: https://doi.org/10.9781/ijimai.2021.04.

005.

18. S. Koga, A. Ikeda, and D. W. Dickson, “Deep learning-

based model for diagnosing alzheimer’s disease and

tauopathies,” Neuropathology and Applied Neurobiology,

vol. 48, no. 1, p. e12759, 2022, doi: https://doi.org/10.

1111/nan.12759.

19. I. Ayus and D. Gupta, “A novel hybrid ensemble

based alzheimer’s identiﬁcation system using deep learning

technique,” Biomedical Signal Processing and Control,

vol. 92, p. 106079, 2024, doi: https://doi.org/10.1016/j.

bspc.2024.106079.

20. S. Suganyadevi, A. S. Rajasekaran, N. Satheesh, R. Sug-

anthi, R. Naveenkumar et al., “Alzheimer’s disease

diagnosis using deep learning approach,” in 2023 Second

International Conference on Electronics and Renewable

Systems (ICEARS). IEEE, 2023, pp. 1205–1209, doi:

https://doi.org/10.1109/ICEARS56392.2023.10085017.

21. J. L.

Avila-Jim´enez, V. Cant´on-Habas, M. del Pilar

Carrera-Gonz´alez, M. Rich-Ruiz, and S. Ventura, “A

deep learning model for alzheimer’s disease diagnosis

based on patient clinical records,” Computers in Biology

and Medicine, vol. 169, p. 107814, 2024, doi:

https://doi.org/10.1016/j.compbiomed.2023.107814.

22. S. Sava¸s, “Detecting the stages of alzheimer’s disease with

pre-trained deep learning architectures,” Arabian Journal

for Science and Engineering, vol. 47, no. 2, pp. 2201–2218,

2022, doi: https://doi.org/10.1007/s13369-021- 06131-3.

23. W. Al Shehri, “Alzheimer’s disease diagnosis and

classiﬁcation using deep learning techniques,” PeerJ

Computer Science, vol. 8, p. e1177, 2022, doi:

https://doi.org/10.7717/peerj-cs.1177.

24. T. J. Saleem et al., “Deep learning-based diagnosis of

alzheimer’s disease,” Journal of Personalized Medicine,

vol. 12, no. 5, p. 815, 2022, doi: https://doi.org/10.3390/

jpm12050815.

25. G. Hcini, I. Jdey, and H. Dhahri, “Investigating

deep learning for early detection and decision-making in

alzheimer’s disease: A comprehensive review: G. hcini et

al.” Neural Processing Letters, vol. 56, no. 3, p. 153, 2024,

doi: https://doi.org/10.1007/s11063-024- 11600- 5.

26. S. Jain and R. Srivastava, “Enhanced eeg-based alzheimer’s

disease detection using synchrosqueezing transform and

deep transfer learning,” Neuroscience, vol. 576, pp.

105–117, 2025, doi: https://doi.org/10.1016/j.neuroscience.

2025.04.041.

27. H. A. Helaly, M. Badawy, and A. Y. Haikal, “Deep learning

approach for early detection of alzheimer’s disease,”

Cognitive computation, vol. 14, no. 5, pp. 1711–1727,

2022, doi: https://doi.org/10.1007/s12559-021- 09946-2.

28. T. Vanaja, K. Shanmugavadivel, M. Subramanian, and

C. Kanimozhiselvi, “Advancing alzheimer’s detection:

integrative approaches in mri analysis with traditional

and deep learning models,” Neural Computing and

Applications, vol. 37, no. 14, pp. 8527–8546, 2025, doi:

https://doi.org/10.1007/s00521-025-10993-1.

29. M. M. S. Fareed et al., “Add-net: an eﬀective deep

learning model for early detection of alzheimer disease in

mri scans,” IEEE Access, vol. 10, pp. 96 930–96 951, 2022,

doi: https://doi.org/10.1109/ACCESS.2022.3204395.

30. E. Kina, “Tleablcnn: Brain and alzheimer’s disease

detection using attention based explainable deep learning

and smote using imbalanced brain mri,” IEEE Access,

2025, doi: https://doi.org/10.1109/ACCESS.2025.3539550.

31. M. Taiyeb Khosroshahi et al., “Explainable artiﬁcial

intelligence in neuroimaging of alzheimer’s disease,”

Diagnostics, vol. 15, no. 5, p. 612, 2025, doi:

https://doi.org/10.3390/diagnostics15050612.

BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/ryt1gh36

Corrigendum

CORRIGENDUM

Corrigendum regarding missing Funding statements in

previously published articles

Accepted on 20 June 2026

Abstract

Funding statements were missing in the published version of the following articles that appeared in previous issues of

BenchCouncil Transactions on Benchmarks, Standards and Evaluations. The correct Funding statements, as provided

and veriﬁed by the authors, are listed below.

1.“Predicting the number of call center incoming calls us-

ing deep learning”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2025: 100213) h t t p s :

//doi.org/10.1016/j.tbench.2025.100213

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

2.“Comparative study of deep learning models for Parkinson’s

disease detection”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2025: 100219) h t t p s :

//doi.org/10.1016/j.tbench.2025.100219

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

3.“An adaptive opposite slime mold feature selection algorithm

for complex optimization problems”(BenchCouncil Transac-

tions on Benchmarks, Standards and Evaluations Journal,

2025: 100250) https://doi.org/10.1016/j.tbench.2025.100250

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

4.“Evaluation of mechanical properties of natural ﬁber based

polymer comp osite”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2024: 100183)

https://doi.org/10.1016/j.tbench.2024.100183

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

5.“MultiPoint: Enabling scalable pre-silicon p erformance eval-

uation for multi-task workloads”(BenchCouncil Transactions

on Benchmarks, Standards and Evaluations Journal, 2024:

100189) https://doi.org/10.1016/j.tbench.2025.100189

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

6.“BinCodex: A comprehensive and multi-level dataset for

evaluating binary code similarity detection techniques”(Bench-

Council Transactions on Benchmarks, Standards and Evalua-

tions Journal, 2024: 100163) https: / /doi.org/10 .1016/j.tbe n

ch.2024.100163

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

7.“Analyzing the impact of opportunistic maintenance op-

timization on manufacturing industries in Bangladesh: An

empirical study”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2024: 100172) h t t p s :

//doi.org/10.1016/j.tbench.2024.100172

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

8.“TensorTable: Extending PyTorch for mixed relational and

linear algebra pipelines”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2024: 100161)

https://doi.org/10.1016/j.tbench.2024.100161

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

9.“Evaluatology: The science and engineering of evalua-

tion”(BenchCouncil Transactions on Benchmarks, Standards

and Evaluations Journal, 2024: 100162) htt ps:// do i. org/1 0

.1016/j.tbench.2024.100162

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

10.“CloudAISim: A toolkit for modelling and simulation of

modern applications in AI-driven cloud computing environ-

ments”(BenchCouncil Transactions on Benchmarks, Standards

and Evaluations Journal, 2023: 100150) https://doi.org/10.1

016/j.tbench.2024.100150

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

Short Article Title

11.“Benchmarking, ethical alignment, and evaluation frame-

work for conversational AI: Advancing resp onsible develop-

ment of ChatGPT”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2023: 100136)

https://doi.org/10.1016/j.tbench.2023.100136

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

12.“Analyzing the potential beneﬁts and use cases of Chat-

GPT as a tool for improving the eﬃciency and eﬀectiveness

of business operations”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2023: 100140)

https://doi.org/10.1016/j.tbench.2023.100140

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

13.“MetaverseBench: Instantiating and benchmarking meta-

verse challenges”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2023: 100138) h t t p s :

//doi.org/10.1016/j.tbench.2023.100138

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

14.“Mind meets machine: Unravelling GPT-4’s cognitive psy-

chology”(BenchCouncil Transactions on Benchmarks, Stan-

dards and Evaluations Journal, 2023: 100139) htt ps :/ /d oi .

org/10.1016/j.tbench.2023.100139

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

15.“Unlocking the opportunities through ChatGPT Tool

towards ameliorating the education system”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2023: 100115) https://doi.org/10.1016/j.tbench.2023.10

0115

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

16.“Benchmarking HTAP databases for performance isolation

and real-time analytics”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2023: 100122)

https://doi.org/10.1016/j.tbench.2023.100122

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

17.“CoviDetector: A transfer learning-based semi supervised

approach to detect Covid-19 using CXR images”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2023: 100119) https://doi.org/10.1016/j.tbench.2023.10

0119

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

18.“DPUBench: An application-driven scalable benchmark

suite for comprehensive DPU evaluation”(BenchCouncil Trans-

actions on Benchmarks, Standards and Evaluations Journal,

2023: 100120) https://doi.org/10.1016/j.tbench.2023.100120

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

19.“ERMDS: A obfuscation dataset for evaluating robustness

of learning-based malware detection system”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2023: 100106) https://doi.org/10.1016/j.tbench.2023.10

0106

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

20.“SNNBench: End-to-end AI-oriented spiking neural net-

work benchmarking”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2023: 100108)

https://doi.org/10.1016/j.tbench.2023.100108

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

21.“Enabling hyperscale web services”(BenchCouncil Trans-

actions on Benchmarks, Standards and Evaluations Journal,

2023: 100192) https://doi.org/10.1016/j.tbench.2023.100092

Funding: This work was supported by (1) the Center for

Applications Driving Architectures (ADA), one of six centers

of JUMP, a Semiconductor Research Corporation program co-

sponsored by DARPA; (2) NSF Grant IIS1539011; (3) gifts from

Intel and Google; and (4) a Facebook Fellowship.

22.“ChatGPT for healthcare services: An emerging stage for an

innovative perspective”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2023: 100105)

https://doi.org/10.1016/j.tbench.2023.100105

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

23.“HPC AI500 V3.0: A scalable HPC AI benchmarking frame-

work”(BenchCouncil Transactions on Benchmarks, Standards

and Evaluations Journal, 2022: 100083) https://doi.org/10.1

016/j.tbench.2022.100083

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

24.“Performance characterization and optimization of pruning

patterns for sparse DNN inference”(BenchCouncil Transactions

on Benchmarks, Standards and Evaluations Journal, 2022:

100090) https://doi.org/10.1016/j.tbench.2023.100090

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

25.“Enabling Reduced Simpoint Size Through LiveCache and

Detail Warmup”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2022: 100082) h t t p s :

//doi.org/10.1016/j.tbench.2022.100082

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

26.An era of ChatGPT as a signiﬁcant futuristic support tool:

A study on features, abilities, and challenges”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2022: 100089) https://doi.org/10.1016/j.tbench.2023.10

0089

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

27.“An extensive study on Internet of Behavior (IoB) en-

abled Healthcare-Systems: Features, facilitators, and chal-

lenges”(BenchCouncil Transactions on Benchmarks, Standards

and Evaluations Journal, 2022: 100085) https://doi.org/10.1

016/j.tbench.2023.100085

Short Article Title

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

28.“High fusion computers: The IoTs, edges, data centers,

and humans-in-the-loop as a computer”(BenchCouncil Trans-

actions on Benchmarks, Standards and Evaluations Journal,

2022: 100075) https://doi.org/10.1016/j.tbench.2022.100075

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

29.“A review of Blockchain Technology applications for ﬁ-

nancial services”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2022: 100073) h t t p s :

//doi.org/10.1016/j.tbench.2022.100073

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

30.“SAIBench: Benchmarking AI for Science”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2022: 100063) https://doi.org/10.1016/j.tbench.2022.10

0063

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

31.“Performance and energy consumption tradeoﬀ in server

consolidation”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2022: 100060) h t t p s :

//doi.org/10.1016/j.tbench.2022.100060

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

32.“Challenges and recent advances in the design of real-time

wireless Cyber-Physical Systems”(BenchCouncil Transactions

on Benchmarks, Standards and Evaluations Journal, 2022:

100036) https://doi.org/10.1016/j.tbench.2022.100036

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

33.“Benchmarking feature selection methods with diﬀer-

ent prediction mo dels on large-scale healthcare event

data”(BenchCouncil Transactions on Benchmarks, Standards

and Evaluations Journal, 2021: 100004) https://doi.org/10.1

016/j.tbench.2021.100004

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

34.“MVDI25K: A large-scale dataset of microscopic vaginal

discharge images”(BenchCouncil Transactions on Benchmarks,

Standards and Evaluations Journal, 2021: 100008) h t t p s :

//doi.org/10.1016/j.tbench.2021.100008

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

35.“Fallout: Distributed systems testing as a service”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2021: 100010) https://doi.org/10.1016/j.tbench.2021.10

0010

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

36.“Revisiting the eﬀects of the Spectre and Meltdown patches

using the top-down microarchitectural method and purchasing

power parity theory”(BenchCouncil Transactions on Bench-

marks, Standards and Evaluations Journal, 2021: 100011)

https://doi.org/10.1016/j.tbench.2021.100011

Funding: This research received no speciﬁc grant from any

funding agency in the public, commercial, or not-for-proﬁt sec-

tors.

BenchCouncil Transactions on Benchmarks, Standards and

Evaluations, 2026

DOI: https://doi.org/10.66834/t02ph695

Corrigendum

CORRIGENDUM

Corrigendum regarding incorrect Declaration of

Conﬂict-of-Interest Statements in Previously

Published Articles

Accepted on 11 June 2026

Abstract

Declaration of Competing Interest statements were incorrectly included in the published version of the following arti-

cles that appeared in previous issues of BenchCouncil Transactions on Benchmarks, Standards and Evaluations. The

appropriate Conﬂict-of-Interest Statements, provided by the authors, are included below.

1.”Evaluatology-driven artiﬁcial intelligence”(BenchCouncil

Transactions on Benchmarks, Standards and Evaluations Jour-

nal, 2025: 100245) https://doi.org/10.1016/j.tbench.2025.10

0245

Declaration of competing interest: Guoxin Kang is an Asso-

ciate Editor, Wanling Gao is an Assistant Editor-in-Chief, and

Jianfeng Zhan is the Editor-in-Chief of BenchCouncil Transac-

tions on Benchmarks, Standards and Evaluations. They were

not involved in the editorial review process or the decision to

publish this article.

2.”MultiPoint: Enabling scalable pre-silicon p erformance eval-

uation for multi-task workloads”(BenchCouncil Transactions

on Benchmarks, Standards and Evaluations Journal, 2024:

100189) https://doi.org/10.1016/j.tbench.2025.100189

The authors Yuxuan Wu and Wenxiang Wang declare the

following personal relationship that may be considered a po-

tential competing interest: they are currently employed by

Loongson Technology, Beijing, China.