SteerRM: Debiasing Reward Models via Sparse Autoencoders

TL;DR

SteerRM introduces a training-free, SAE-based method to debias reward models by suppressing stylistic biases at inference, improving accuracy without retraining across multiple architectures.

Contribution

This paper presents SteerRM, a novel approach that uses Sparse Autoencoders for bias mitigation in reward models without retraining or architectural changes.

Findings

Improves Hard-split accuracy by 7.3 points on average across six RMs.

Effectively generalizes across different bias types and architectures.

Identifies shared bias encoding patterns in shallow layers.

Abstract

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.