One Bias After Another: Mechanistic Reward Shaping and Persistent Biases in Language Reward Models
This addresses the issue of reward hacking in language model alignment for AI safety, but it is incremental as it builds on prior work to mitigate known biases.
The paper tackled the problem of persistent biases in language reward models (RMs), such as length, sycophancy, and overconfidence, by proposing a mechanistic reward shaping method that reduces targeted biases without degrading reward quality and using minimal labeled data.
Reward Models (RMs) are crucial for online alignment of language models (LMs) with human preferences. However, RM-based preference-tuning is vulnerable to reward hacking, whereby LM policies learn undesirable behaviors from flawed RMs. By systematically measuring biases in five high-quality RMs, including the state-of-the-art, we find that issues persist despite prior work with respect to length, sycophancy, and overconfidence. We also discover new issues related to bias toward model-specific styles and answer-order. We categorize RM failures by complexity and propose a simple post-hoc intervention to mitigate low-complexity biases that arise from spurious correlations. Our proposed mechanistic reward shaping reduces targeted biases without degrading reward quality and while using minimal labeled data. The method is extensible to new biases, model-internal, and generalizes out-of-distribution.