debiasing

This paper investigates whether label-free strategies for multiple-choice benchmarks can remove option-order sensitivity in large language models, finding that neither two-stage prompting nor independent hypothesis scoring reliably improves accuracy.

This paper proposes a novel method to mitigate scoring bias in LLM-as-a-Judge by having LLMs randomly generate numbers to measure their latent numerical bias, then rectifying token generation probabilities accordingly. Experiments across four tasks show the method outperforms baselines and reveals that scoring bias varies across models, tasks, and score ranges.

This paper proposes HEIMAT, a heuristic-style automatic debiasing framework for language models that uses heuristic prompts to reveal biases and fine-tunes the model to reduce bias while preserving NLU performance.

This paper introduces Narrative Anchoring, a failure mode where clinical language models produce divergent diagnoses when identical clinical facts are expressed in different sociolinguistic registers. The authors release a USMLE-derived dataset and propose NarrativeShield, a three-agent pipeline that reduces the anchoring gap to near-zero.

This paper investigates how alignment tuning introduces cue-induced biases such as sycophancy in LLMs, finding that biases are installed by alignment rather than pretraining and can be decoded and steered via hidden state directions.

This paper investigates preprocessing-based stereotype mitigation methods in NLP and finds that while they reduce targeted stereotypes, they can inadvertently increase stereotyping or counter-stereotyping for other demographic groups, including across unrelated categories. The authors demonstrate these side effects across model families and preprocessing strategies, and discuss implications for evaluation and mitigation practices.

Introduces Reward-Gated Test-Time Adaptation (RG-TTA), a reinforcement learning framework that selectively applies debiasing to CLIP models based on input bias sensitivity, resolving the fairness-utility trade-off.

This paper audits multilingual clinical ASR systems on psychiatric interviews in Indian languages and proposes SamaVaani, a unified debiasing technique to improve performance and fairness across demographic groups.

A post-hoc method reduces spurious correlations in fine-tuned LLMs by truncating the tail of the SVD of the weight update matrix. It reduces the spurious-group gap by up to 5x with less than 2pp accuracy loss, without retraining or group labels.

PEARL introduces a contrastive percentile approximation framework to mitigate behavioral intensity imbalance in recommender systems, achieving significant gains in engagement metrics in a production livestream platform serving billions of users.

DebiasRAG proposes a tuning-free, query-specific debiasing framework using retrieval-augmented generation to reduce social biases in LLMs without degrading their original capabilities.

Researchers from MIT, WPI, and Google propose WRING, a novel post-processing debiasing method for Vision-Language Models that avoids the 'Whac-a-mole dilemma' of amplifying other biases when removing specific ones.

This paper proposes Product-of-Experts (PoE) training to reduce dataset artifacts in Natural Language Inference, downweighting examples where biased models are overconfident. PoE nearly preserves accuracy on SNLI (89.10% vs. 89.30%) while reducing bias reliance by ~4.85 percentage points.