arXiv subjects

Hadi Mohammadi

Publications and source records attributed to Hadi Mohammadi.

13 recordsLinked to original sources

arXiv2026

trajectory-judge: What Outcome-Only LLM Judges Miss on Agent Trajectories

A direct test of an LLM judge of agent trajectories injects faults into correct runs and reports recall, per fault type or by whether the fault broke the environment outcome (loud) or not (silent). Such recall can credit a judge with detection it does not have; paired discrimination, its flag rate on the faults minus its rate on the clean runs they came from, exposes this. Our testbed, a deterministic support desk with a scripted oracle and a one-step fault injector, labels all 400 trajectories exactly. A 14B judge shown only the request and final reply scores 34% to 76% recall on four fault types that leave the reply unchanged. There its input is the clean run's, so its paired discrimination is zero and that recall is its flag rate on clean runs. Splitting by outcome survival does not fix this: its loud recall of 84% is a paired +0.393 and its silent recall of 45% a paired +0.048, all from the two fault types that change the reply. Told to check each step, the same model flags every fault of those four types and 0 of 100 clean runs (95% CI up to 3.6%). It does not reliably check the reply: of four invented promises it flags one every time and the other three once in 42 faults. Shown every step but asked only about the reply, it still reaches a paired +0.69 on reply-unchanged faults, against +1.00 when told to check each step. We recommend reporting paired discrimination against clean parents, split by whether the fault reaches the judge's input and by outcome survival, and release the testbed, raw verdicts and analysis pipeline.

cs.CL↗

arXiv2026

Explainability in Practice: A Survey of Explainable NLP Across Various Domains

Natural Language Processing (NLP) is now embedded in critical sectors including healthcare, finance, and customer relationship management, where models such as GPT-4o, Gemini, and BERT increasingly inform decisions. The black-box nature of these models has created an urgent need for transparency. This review examines explainable NLP (XNLP) as it is actually deployed, working through seven application domains: medicine, finance, systematic reviews, customer relationship management, chatbots, social and behavioral science, and human resources. For each domain, we ask what kind of explanation the setting needs, which methods are used there, and how they are evaluated. A structured cross-domain synthesis then contrasts how those requirements diverge. We compare the main explanation method families on scope, evidence of faithfulness, and computational cost. We also propose a two-tier evaluation protocol that separates a shared technical core of metrics from the domain-specific validation layer through which those metrics have to be read. The review also addresses areas that remain underrepresented in the XNLP literature, including real-world applicability, the gap between fidelity and faithfulness, and the role of human judgment in assessing explanations. It closes with research directions, among them personalized explanations, human-in-the-loop evaluation, and mechanistic interpretability for large language models.

cs.CL↗

arXiv2026

Learning Sexism Detection Using Multi-Agent Perspectivist Preference Optimization

When people label text for sexism, they often disagree, and not because some of them are wrong: they genuinely perceive sexism differently. Most NLP systems discard this disagreement by collapsing it into a majority vote. We propose the Multi-Agent Perspectivist Preference Optimization (MAP-PO) framework to keep these different perspectives. On the EXIST 2024 dataset of labeled English and Spanish tweets, we first cluster annotators by their labeling behavior rather than their demographic attributes. We then fine-tune one Large Language Model agent per cluster to reproduce that cluster's annotation behavior, and coordinate the agents with preference optimization that combines individual and team-level rewards. We evaluate MAP-PO in four settings defined by two languages and two backbone language models, asking whether each agent reproduces the annotations of its own cluster and whether the agents together reproduce the majority label. Two findings hold in all four settings. First, without fine-tuning the agents behave almost identically, so cluster-specific training is necessary. Second, we show that training each agent only on the labels of its own cluster pushes the agents far beyond the clusters they should represent, while adding a shared team-level training signal consistently keeps each agent calibrated to its cluster.

cs.CL↗

arXiv2026

Reliability-Aware Sexism Detection: Combining DPO with Annotator Agreement and Token-Level Confidence Scoring

The detection of online sexism remains an open problem. Sexism detection is inherently subjective, yet most existing systems reduce multi-annotator labels to a single majority decision and treat all instances uniformly. This ignores two informative signals: annotator agreement and model uncertainty. We propose RA-DPO (Reliability-Aware Direct Preference Optimization), which integrates annotator agreement, model confidence, and a token-level uncertainty signal into a single reliability score. RA-DPO uses this score to select high-value preference pairs during training and to support inference-time abstention, which allows the model to trade coverage for accuracy. We evaluate RA-DPO on 6,920 multilingual posts from EXIST 2023, fine-tune OpenAI gpt-4o base via DPO, and validate on two open-weight 3B models (Llama, Qwen). Results show that training on the top 30% most reliable pairs matches full-data DPO, which indicates that reliability-aware selection can reduce training cost without sacrificing performance. At inference, selective prediction reaches 96.2% accuracy at 50% coverage in the true-agreement setting and 88.7% in the deployable predicted-agreement setting, both exceeding the 85.3% no-agreement baseline. These results suggest that accounting for annotation uncertainty is beneficial for both efficient training and reliable deployment in subjective classification.

cs.CL↗

arXiv2026

EvalMORAAL: Interpretable Chain-of-Thought and LLM-as-Judge Evaluation for Moral Alignment in Large Language Models

We present EvalMORAAL, a transparent chain-of-thought (CoT) framework that uses two scoring methods (log-probabilities and direct ratings) plus a model-as-judge peer review to evaluate moral alignment in 20 large language models. We assess models on the World Values Survey (55 countries, 19 topics) and the PEW Global Attitudes Survey (39 countries, 8 topics). With EvalMORAAL, top models align closely with survey responses (Pearson's $r \approx 0.90$ on WVS). Yet we find a clear regional difference: Western regions average $r=0.82$ while non-Western regions average $r=0.61$ (a 0.21 absolute gap), indicating a persistent regional alignment gap. Our framework adds three parts: (1) two scoring methods for all models to enable fair comparison, (2) a structured CoT protocol with self-consistency checks, and (3) a model-as-judge peer review that flags 348 conflicts using a data-driven threshold. Peer agreement relates to WVS survey alignment ($r=0.74$, $p<.001$; PEW $r=0.39$, n.s.), supporting automated quality checks. These results show real progress toward culture-aware AI while highlighting open challenges for use across regions.

cs.CL↗

arXiv2026

Explainability-Based Token Replacement on LLM-Generated Text

Generative models, especially large language models (LLMs), have shown remarkable progress in producing text that appears human-like. However, they often exhibit patterns that make their output easier to detect than text written by humans. In this paper, we investigate how explainable AI (XAI) methods can be used to reduce the detectability of AI-generated text (AIGT) while also introducing a robust ensemble-based detection approach. We begin by training an ensemble classifier to distinguish AIGT from human-written text, then apply SHAP and LIME to identify tokens that most strongly influence its predictions. We propose four explainability-based token replacement strategies to modify these influential tokens. Our findings show that these token replacement approaches can significantly diminish a single classifier's ability to detect AIGT. However, our ensemble classifier maintains strong performance across multiple languages and domains, showing that a multi-model approach can mitigate the impact of token-level manipulations. These results show that XAI methods can make AIGT harder to detect by focusing on the most influential tokens. At the same time, they highlight the need for robust, ensemble-based detection strategies that can adapt to evolving approaches for hiding AIGT.

cs.CL↗

arXiv2026

Exploring Cultural Variations in Moral Judgments with Large Language Models

Large Language Models (LLMs) have shown strong performance across many tasks, but their ability to capture culturally diverse moral values remains unclear. In this paper, we examine whether LLMs mirror variations in moral attitudes reported by the World Values Survey (WVS) and the Pew Research Center's Global Attitudes Survey (PEW). We compare smaller monolingual and multilingual models (GPT-2, OPT, BLOOMZ, and Qwen) with recent instruction-tuned models (GPT-4o, GPT-4o-mini, Gemma-2-9b-it, and Llama-3.3-70B-Instruct). Using log-probability-based \emph{moral justifiability} scores, we correlate each model's outputs with survey data covering a broad set of ethical topics. Our results show that many earlier or smaller models often produce near-zero or negative correlations with human judgments. In contrast, advanced instruction-tuned models achieve substantially higher positive correlations, suggesting they better reflect real-world moral attitudes. We provide a detailed regional analysis revealing that models align better with Western, Educated, Industrialized, Rich, and Democratic (W.E.I.R.D.) nations than with other regions. While scaling model size and using instruction tuning improves alignment with cross-cultural moral norms, challenges remain for certain topics and regions. We discuss these findings in relation to bias analysis, training data diversity, information retrieval implications, and strategies for improving the cultural sensitivity of LLMs.

cs.CL↗

arXiv2025

Evaluating GRPO and DPO for Faithful Chain-of-Thought Reasoning in LLMs

Chain-of-thought (CoT) reasoning has emerged as a powerful technique for improving the problem-solving capabilities of large language models (LLMs), particularly for tasks requiring multi-step reasoning. However, recent studies show that CoT explanations often fail to reflect the model's actual reasoning process, as models may produce coherent yet misleading justifications or modify answers without acknowledging external cues. Such discrepancies undermine the reliability of CoT-based methods for safety supervision and alignment monitoring, as models can generate plausible but deceptive rationales for incorrect answers. To better understand this limitation, we evaluate two optimization methods, Group Relative Policy Optimization (GRPO) and Direct Preference Optimization (DPO), in their ability to improve CoT faithfulness. Our experiments show that GRPO achieves higher performance than DPO in larger models, with the Qwen2.5-14B-Instruct model attaining the best results across all evaluation metrics. Both approaches exhibit positive correlations between model size and performance, but GRPO shows greater potential for improving faithfulness metrics, albeit with less stable behavior at smaller scales. These results suggest that GRPO offers a promising direction for developing more transparent and trustworthy reasoning in LLMs.

cs.CL↗

arXiv2025

Do Large Language Models Understand Morality Across Cultures?

Recent advancements in large language models (LLMs) have established them as powerful tools across numerous domains. However, persistent concerns about embedded biases, such as gender, racial, and cultural biases arising from their training data, raise significant questions about the ethical use and societal consequences of these technologies. This study investigates the extent to which LLMs capture cross-cultural differences and similarities in moral perspectives. Specifically, we examine whether LLM outputs align with patterns observed in international survey data on moral attitudes. To this end, we employ three complementary methods: (1) comparing variances in moral scores produced by models versus those reported in surveys, (2) conducting cluster alignment analyses to assess correspondence between country groupings derived from LLM outputs and survey data, and (3) directly probing models with comparative prompts using systematically chosen token pairs. Our results reveal that current LLMs often fail to reproduce the full spectrum of cross-cultural moral variation, tending to compress differences and exhibit low alignment with empirical survey patterns. These findings highlight a pressing need for more robust approaches to mitigate biases and improve cultural representativeness in LLMs. We conclude by discussing the implications for the responsible development and global deployment of LLMs, emphasizing fairness and ethical alignment.

cs.CL↗

arXiv2025

Assessing the Reliability of LLMs Annotations in the Context of Demographic Bias and Model Explanation

Understanding the sources of variability in annotations is crucial for developing fair NLP systems, especially for tasks like sexism detection where demographic bias is a concern. This study investigates the extent to which annotator demographic features influence labeling decisions compared to text content. Using a Generalized Linear Mixed Model, we quantify this inf luence, finding that while statistically present, demographic factors account for a minor fraction ( 8%) of the observed variance, with tweet content being the dominant factor. We then assess the reliability of Generative AI (GenAI) models as annotators, specifically evaluating if guiding them with demographic personas improves alignment with human judgments. Our results indicate that simplistic persona prompting often fails to enhance, and sometimes degrades, performance compared to baseline models. Furthermore, explainable AI (XAI) techniques reveal that model predictions rely heavily on content-specific tokens related to sexism, rather than correlates of demographic characteristics. We argue that focusing on content-driven explanations and robust annotation protocols offers a more reliable path towards fairness than potentially persona simulation.

cs.CL↗

arXiv2025

Anchor-based oversampling for imbalanced tabular data via contrastive and adversarial learning

Imbalanced data represent a distribution with more frequencies of one class (majority) than the other (minority). This phenomenon occurs across various domains, such as security, medical care and human activity. In imbalanced learning, classification algorithms are typically inclined to classify the majority class accurately, resulting in artificially high accuracy rates. As a result, many minority samples are mistakenly labelled as majority-class instances, resulting in a bias that benefits the majority class. This study presents a framework based on boundary anchor samples to tackle the imbalance learning challenge. First, we select and use anchor samples to train a multilayer perceptron (MLP) classifier, which acts as a prior knowledge model and aids the adversarial and contrastive learning procedures. Then, we designed a novel deep generative model called Anchor Stabilized Conditional Generative Adversarial Network or Anch-SCGAN in short. Anch-SCGAN is supported with two generators for the minority and majority classes and a discriminator incorporating additional class-specific information from the pre-trained feature extractor MLP. In addition, we facilitate the generator's training procedure in two ways. First, we define a new generator loss function based on reprocessed anchor samples and contrastive learning. Second, we apply a scoring strategy to stabilize the adversarial training part in generators. We train Anch-SCGAN and further finetune it with anchor samples to improve the precision of the generated samples. Our experiments on 16 real-world imbalanced datasets illustrate that Anch-SCGAN outperforms the renowned methods in imbalanced learning.

cs.LG↗

arXiv2024

Large Language Models as Mirrors of Societal Moral Standards

Prior research has demonstrated that language models can, to a limited extent, represent moral norms in a variety of cultural contexts. This research aims to replicate these findings and further explore their validity, concentrating on issues like 'homosexuality' and 'divorce'. This study evaluates the effectiveness of these models using information from two surveys, the WVS and the PEW, that encompass moral perspectives from over 40 countries. The results show that biases exist in both monolingual and multilingual models, and they typically fall short of accurately capturing the moral intricacies of diverse cultures. However, the BLOOM model shows the best performance, exhibiting some positive correlations, but still does not achieve a comprehensive moral understanding. This research underscores the limitations of current PLMs in processing cross-cultural differences in values and highlights the importance of developing culturally aware AI systems that better align with universal human values.

cs.AI↗

arXiv2024

LLMs as mirrors of societal moral standards: reflection of cultural divergence and agreement across ethical topics

Large language models (LLMs) have become increasingly pivotal in various domains due the recent advancements in their performance capabilities. However, concerns persist regarding biases in LLMs, including gender, racial, and cultural biases derived from their training data. These biases raise critical questions about the ethical deployment and societal impact of LLMs. Acknowledging these concerns, this study investigates whether LLMs accurately reflect cross-cultural variations and similarities in moral perspectives. In assessing whether the chosen LLMs capture patterns of divergence and agreement on moral topics across cultures, three main methods are employed: (1) comparison of model-generated and survey-based moral score variances, (2) cluster alignment analysis to evaluate the correspondence between country clusters derived from model-generated moral scores and those derived from survey data, and (3) probing LLMs with direct comparative prompts. All three methods involve the use of systematic prompts and token pairs designed to assess how well LLMs understand and reflect cultural variations in moral attitudes. The findings of this study indicate overall variable and low performance in reflecting cross-cultural differences and similarities in moral values across the models tested, highlighting the necessity for improving models' accuracy in capturing these nuances effectively. The insights gained from this study aim to inform discussions on the ethical development and deployment of LLMs in global contexts, emphasizing the importance of mitigating biases and promoting fair representation across diverse cultural perspectives.

cs.AI↗