paper-with-me

Papers

SteerRM: Debiasing Reward Models via Sparse Autoencoders

2026-03-13 · Mengyuan Sun, Zhuohao Yu, Weizheng Gu, Shikun Zhang, Wei Ye arxiv

Reward models (RMs) are critical components of alignment pipelines, yet they exhibit biases toward superficial stylistic cues, preferring better-presented responses over semantically superior ones. Existing debiasing methods typically require retraining or architectural modifications, while direct activation suppression degrades performance due to representation entanglement. We propose SteerRM, the first training-free method for debiasing reward models using Sparse Autoencoder (SAE)-based interventions. SteerRM isolates stylistic effects using contrastive paired responses, identifies bias-related SAE features with a strength-stability criterion, and suppresses them at inference time. Across six reward models on RM-Bench, SteerRM improves Hard-split accuracy by 7.3 points on average while preserving overall performance. Results on a Gemma-based reward model and a controlled non-format bias further suggest generalization across RM architectures and bias types. We further find that format-related features are concentrated in shallow layers and transfer across models, revealing shared architecture-level bias encoding patterns. These results show that SAE-based interventions can mitigate reward-model biases without retraining, providing a practical and interpretable solution for alignment pipelines.

📄 PDF Abstract BibTeX arXiv:2603.12795

Code (0)

등록된 구현이 없습니다.

Similar Papers 제목 키워드 기반

debiaSAE: Benchmarking and Mitigating Vision-Language Model Bias

2024-10-17 · Kuleen Sasse, Shan Chen, Jackson Pond, Danielle Bitterman 외

As Vision Language Models (VLMs) gain widespread use, their fairness remains under-explored. In this paper, we analyze demographic biases across five models and six datasets. We find that portrait datasets like UTKFace a…

BenchmarkingBias DetectionFairnessLanguage Modeling+1

Mitigating Reward Hacking in RLHF via Bayesian Non-negative Reward Modeling

2026-02-11 · Zhibin Duan, Guowei Rong, Zhuo Li, Bo Chen 외 arxiv

Reward models learned from human preferences are central to aligning large language models (LLMs) via reinforcement learning from human feedback, yet they are often vulnerable to reward hacking due to noisy annotations a…

Reinforcement Learning

Debiasing Multilingual LLMs in Cross-lingual Latent Space

2025-08-25 · Qiwei Peng, Guimin Hu, Yekun Chai, Anders Søgaard arxiv

Debiasing techniques such as SentDebias aim to reduce bias in large language models (LLMs). Previous studies have evaluated their cross-lingual transferability by directly applying these methods to LLM representations, r…

Interpretable Debiasing of Vision-Language Models for Social Fairness

2026-02-27 · Na Min An, Yoonna Jang, Yusuke Hirota, Ryo Hachiuma 외 arxiv

The rapid advancement of Vision-Language models (VLMs) has raised growing concerns that their black-box reasoning processes could lead to unintended forms of social bias. Current debiasing approaches focus on mitigating …

Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders

2026-05-07 · Shunchang Liu, Xin Chen, Belen Martin Urcelay, Francesco Croce arxiv

Preference learning in large language models relies on reward models as proxies for human judgment. However, these models frequently exhibit preference instability, producing contradictory preference assignments in respo…