Automatically Finding Reward Model Biases

TL;DR

This paper presents an automated method using large language models to identify biases in reward models for language models, revealing both known and novel biases to improve interpretability.

Contribution

It introduces an iterative LLM-based approach for discovering reward model biases, outperforming traditional search methods and validated on synthetic biases.

Findings

Recovered known biases in reward models

Discovered novel biases such as redundant spacing and hallucinations

Evolutionary iteration outperforms flat search methods

Abstract

Reward models are central to large language model (LLM) post-training. However, past work has shown that they can reward spurious or undesirable attributes such as length, format, hallucinations, and sycophancy. In this work, we introduce and study the research problem of automatically finding reward model biases in natural language. We offer a simple approach of using an LLM to iteratively propose and refine candidate biases. Our method can recover known biases and surface novel ones: for example, we found that Skywork-V2-8B, a leading open-weight reward model, often mistakenly favors responses with redundant spacing and responses with hallucinated content. In addition, we show evidence that evolutionary iteration outperforms flat best-of-N search, and we validate the recall of our pipeline using synthetically injected biases. We hope our work contributes to further research on improving RMs through automated interpretability methods.