RM-Bench: Benchmarking Reward Models of Language Models with Subtlety and Style

TL;DR

RM-Bench is a new benchmark that evaluates reward models based on their ability to detect subtle content differences and resist style biases, showing current models need significant improvement for better alignment.

Contribution

The paper introduces RM-Bench, a novel benchmark for assessing reward models' sensitivity to subtle content and style biases, correlating well with policy performance.

Findings

Nearly 40 reward models evaluated, with average performance of 46.6%.

Current models struggle with style bias, performing near random chance.

RM-Bench correlates strongly with policy model success.

Abstract

Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses. Despite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power. However, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance. To this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases. Extensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively. We evaluate nearly 40 reward models on RM-Bench. Our results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference. These findings highlight the significant room for improvement in current reward models. Related code and data are available at https://github.com/THU-KEG/RM-Bench.