Self-Generated Critiques Boost Reward Modeling for Language Models

TL;DR

This paper introduces Critic-RM, a novel framework that enhances reward modeling for language models by using self-generated critiques to improve alignment with human preferences, achieving significant accuracy gains.

Contribution

Critic-RM is the first approach to incorporate self-generated critiques into reward modeling without extra supervision, improving accuracy and reasoning in language models.

Findings

Reward modeling accuracy improved by 3.7%-7.3%.

Generated critiques help rectify flawed reasoning steps.

Demonstrates strong performance and data efficiency.

Abstract

Reward modeling is crucial for aligning large language models (LLMs) with human preferences, especially in reinforcement learning from human feedback (RLHF). However, current reward models mainly produce scalar scores and struggle to incorporate critiques in a natural language format. We hypothesize that predicting both critiques and the scalar reward would improve reward modeling ability. Motivated by this, we propose Critic-RM, a framework that improves reward models using self-generated critiques without extra supervision. Critic-RM employs a two-stage process: generating and filtering high-quality critiques, followed by joint fine-tuning on reward prediction and critique generation. Experiments across benchmarks show that Critic-RM improves reward modeling accuracy by 3.7%-7.3% compared to standard reward models and LLM judges, demonstrating strong performance and data efficiency. Additional studies further validate the effectiveness of generated critiques in rectifying flawed reasoning steps with 2.5%-3.2% gains in improving reasoning accuracy.