Mina Arzaghi

Sujets de recherche

Apprentissage profond

Systèmes de recommandation

Site web

GitHub

Publications

Beyond Fair Answers: Learning Fair and Unlearning Biased Reasoning in LLMs

Large Language Models (LLMs) have demonstrated strong reasoning capabilities on tasks with well-defined solutions, such as mathematical and … (voir plus)logical reasoning. However, reasoning capability does not necessarily imply unbiased reasoning. In socially ambiguous contexts, where the available evidence is insufficient to support one conclusion over another, models may rely on demographic stereotypes or unsupported social assumptions. Recent work shows that chain-of-thought (CoT) reasoning can contain social stereotypes and that biased reasoning steps are associated with stereotype expression and incorrect predictions. This motivates evaluating fairness in the reasoning process, not only in the final answer. We therefore ask: Can we train LLMs not only to produce fair answers, but to reason fairly? We study this question using a fairness-oriented adaptation of BBQ, where ambiguous cases test whether demographic alternatives receive comparable probability when evidence is insufficient, while disambiguated cases test evidence-grounded accuracy. Following reasoning-trace fine-tuning approaches such as STaR and ReGiFT, we fine-tune instruction-tuned LLMs on reasoning traces generated by stronger models and compare base and fine-tuned models under direct and CoT inference. We evaluate both final decisions and generated reasoning traces, and test generalization on SocioEconomicQA and CrowS-Pairs. Building on this analysis, we propose a reasoning-level fairness framework that jointly learns fair reasoning and unlearns biased reasoning. Evidence-grounded traces form the positive learning set, while traces containing unsupported demographic or stereotype-based inferences form the forget set. Our objective combines supervised fine-tuning on fair traces, a negative objective on biased traces, and KL regularization on unrelated reasoning data to preserve general reasoning ability. We evaluate whether this process improves selection of the fair answer and, more importantly, reduces the probability gap between gendered options when the context provides no evidence favoring either one. We separately measure accuracy in disambiguated contexts to ensure that the model still uses available evidence correctly. At the reasoning level, we evaluate stereotype-dependent inferences and the relative likelihood of fair versus biased reasoning traces. We compare against ReGiFT, standard CoT, and instruction-based reasoning baselines, and evaluate generalization across additional bias benchmarks while monitoring general reasoning performance for capability loss. This work moves fairness intervention from the final decision to the reasoning process that produces it. By combining fair-reasoning supervision, biased-reasoning unlearning, and capability preservation, we investigate whether LLMs can learn not only to give fair answers, but to reason without relying on unsupported demographic stereotypes.

2026-09-29

WiML @ Neural Information Processing Systems (publié)

openreview.net

Mechanics of Bias and Reasoning: Interpreting the Impact of Chain-of-Thought Prompting on Gender Bias in LLMs

Edie Pearman

Sophia Osborne

Mira Kandlikar-Bloch

Large language models (LLMs) are increasingly deployed in socially sensitive settings despite substantial documentation that they encode gen… (voir plus)der biases. Chain-of-Thought (CoT) prompting has been proposed as an approach for bias mitigation. However, existing evaluations primarily focus on changes in LLM benchmark performance, providing limited insight into whether apparent bias reductions reflect meaningful changes in a model's internal mechanisms. In this work, we present an investigation of how CoT prompting affects gender bias in LLMs, combining benchmark-based evaluation with mechanistic interpretability techniques, and qualitative analysis of reasoning outputs. Our results confirm a stereotypical bias present in LLM outputs across benchmarks, showing that CoT prompting does not consistently reduce the bias gap. While mechanistic analyses reveal clusters of attention heads whose biased behavior is lessened with CoT, gender bias information remains pervasive throughout hidden representations, indicating any improvements from CoT are superficial and fail to transform internal processing of gender bias. A closer inspection of the reasoning chains themselves shows poor quality CoT by which the models dissociate, hallucinate, and evade the present task rather than meaningfully engage with prompt material.

2026-03-01

AFAA @ International Conference on Learning Representations (présentation orale)

openreview.net

Intrinsic Meets Extrinsic Fairness: Assessing the Downstream Impact of Bias Mitigation in Large Language Models

Alireza Dehghanpour Farashah

Large Language Models (LLMs) are increasingly deployed in sensitive domains such as finance, where intrinsic representational biases can pro… (voir plus)pagate into extrinsic harms in downstream tasks. High-stakes applications such as credit scoring are especially vulnerable, as biased model behavior can reinforce existing inequities and result in harmful disparities across demographic groups \cite{blodgett2020language}. While prior research has questioned whether intrinsic bias truly translates into extrinsic unfairness \cite{goldfarb2020intrinsic}, this connection remains poorly understood. To address this gap, we propose a four-stage evaluation framework that systematically examines the relationship between intrinsic and extrinsic fairness. In Stage 1, we establish a baseline by training models such as logistic regression, LLM embeddings, and fine-tuned classifiers without any mitigation strategy, providing reference points for fairness and accuracy. In Stage 2, we evaluate task-level mitigation through Counterfactual Data Augmentation (CDA) \cite{gallegos2024bias}, which balances gender representation by generating counterfactual training instances, allowing us to assess improvements in extrinsic fairness. In Stage 3, we adapt concept unlearning \cite{dige2024mitigating} as an intrinsic bias mitigation method, encouraging LLMs to forget socioeconomic stereotypes while preserving fluency and predictive utility, and we evaluate how this intervention impacts downstream fairness. Finally, in Stage 4, we combine CDA with unlearning to test whether dual mitigation further enhances fairness. We conduct experiments on three datasets (Adult Census Income, ACS Employment, and German Credit) using instruction-tuned LLMs (LLaMA-3.1, Phi-3, and Gemma-2) in both frozen embedding and fine-tuned classifier settings, evaluating performance with predictive accuracy and group fairness metrics, including Demographic Parity, Accuracy Parity, and Equality of Odds. Our experiments demonstrate that intrinsic bias mitigation through unlearning is highly effective; in Phi-3, for instance, it reduces gender socioeconomic stereotype gaps by 94.9\% while maintaining language fluency. In downstream tasks, unlearning consistently improves group fairness metrics while preserving predictive accuracy, whereas CDA primarily enhances demographic parity but can introduce accuracy trade-offs. For instance, on the ACS Employment dataset, unlearned Gemma-2 improved Accuracy Parity from 0.199 to 0.104 (48\% gain), and combining CDA with unlearning on Llama-3.1 reduced Demographic Parity from 0.080 to 0.014 (82\% gain). On the Adult dataset, all three models maintained accuracy above 0.82 while showing reduced fairness gaps, and on German Credit, unlearning consistently outperformed CDA by improving group fairness metrics without sacrificing predictive performance. Overall, CDA and unlearning exhibit complementary effects, with their combination yielding the strongest fairness improvements across models and datasets. This work contributes to bias mitigation and fairness in LLMs in two ways. First, we adapt concept unlearning to mitigate socioeconomic stereotyping, showing that intrinsic bias reduction improves both representational and downstream fairness. Second, we introduce a unified evaluation framework that links intrinsic and extrinsic fairness, enabling systematic comparison of mitigation strategies. The framework is flexible, applying to both fine-tuned and frozen LLMs, and offers actionable guidance for deploying fairer models in finance and other high-stakes domains.

2025-09-21

NeurIPS.cc/2025/Workshop/WiML (publié)

openreview.net

Understanding Intrinsic Socioeconomic Biases in Large Language Models