Large language models (LLMs) increasingly participate in conversational systems where responses emerge through multi-turn interactions rather than single prompts. We introduce a multi-agent evaluation framework that probes LLMs for bias through multi-turn conversations. We present a framework that orchestrates three roles (target, adversary, and scorer) under diverse normative fairness definitions. Experiments on five widely used LLMs using 100 adversarial prompts show that adversarial multi-turn probing reliably surfaces bias-related responses, with baseline violation rates exceeding 80% across models, while human validation with the scorer (~85% agreement) confirms the scoring reliability. Preventive prompt-conditioning strategies provide limited reduction in violations, while our iterative post-hoc mitigation pipeline substantially reduces scorer-detected violations under the active fairness profile, from over 80% to near-zero. We frame this as scorer-relative violation reduction rather than universal debiasing.

Large Language Models (LLMs) used in Retrieval-Augmented Generation (RAG) can amplify demographic bias: retrievers may surface skewed context and generators can propagate that skew into decisions. Prior work typically treats fairness in retrieval or generation in isolation, leaving end-to-end fairness in RAG underexplored. We propose a post-hoc pipeline that jointly controls both stages: (i) a Fair Greedy Reranker (FGR) that builds prefix-balanced slates toward a target group mix; (ii) a Residual Slate Bias Estimator (RSBE) using signed, prefix-sensitive NDKL to quantify remaining skew; and (iii) Confidence-Gated Logit Calibration (CGLC) that converts the residual signal into small and margin-focused logit corrections without retraining. On an occupation classification task, our approach reduces retriever-side skew (lowest NDKL among baselines for both dense and sparse retrievers) and achieves the lowest generator-side disparity (e.g., Risk Difference) while largely preserving utility. The same calibration can be tuned to alternative fairness criteria (e.g., Equal Opportunity) with minimal utility loss.

Large Language Models (LLMs) excel in Natural Language Processing (NLP) tasks but often propagate societal biases from their training data, leading to discriminatory outputs. These biases are amplified by the models’ self-attention mechanisms, which disproportionately emphasize biased correlations with sensitive tokens, like “he” or “she”, reflecting the sensitive attributes such as gender and race. To address this issue, we propose a novel fine-tuning method, called Cross-Attention-based Weight Decay (CrAWD), which modifies the LLM architecture to mitigate bias. CrAWD introduces a cross-attention mechanism between an input sequence and a sensitive token sequence, enabling the model to identify and selectively decay the attention weights of tokens associated with sensitive tokens. This reduces the influence of biased association on the model’s generation while maintaining task performance. Evaluations on real-world datasets demonstrate the effectiveness of our proposed CrAWD method. Notably, our method can handle multiple sensitive attributes by adjusting the sensitive token sequence, and it does not require full knowledge of sensitive tokens presented in the dataset, underscoring CrAWD’s versatility in promoting fair LLMs across various applications.