CE-RM: A Pointwise Generative Reward Model Optimized via Two-Stage Rollout and Unified Criteria

TL;DR

This paper introduces CE-RM-4B, a pointwise generative reward model optimized through a two-stage rollout and unified criteria, significantly improving evaluation and reinforcement learning in open-ended language generation.

Contribution

The paper presents a novel pointwise generative reward model with a two-stage rollout and unified criteria, addressing limitations of pairwise evaluation and enhancing RL effectiveness.

Findings

Outperforms existing reward models on diverse benchmarks.

Achieves superior results in Best-of-N scenarios.

Provides more effective improvements in downstream RL tasks.

Abstract

Automatic evaluation is crucial yet challenging for open-ended natural language generation, especially when rule-based metrics are infeasible. Compared with traditional methods, the recent LLM-as-a-Judge paradigms enable better and more flexible evaluation, and show promise as generative reward models for reinforcement learning. However, prior work has revealed a notable gap between their seemingly impressive benchmark performance and actual effectiveness in RL practice. We attribute this issue to some limitations in existing studies, including the dominance of pairwise evaluation and inadequate optimization of evaluation criteria. Therefore, we propose CE-RM-4B, a pointwise generative reward model trained with a dedicated two-stage rollout method, and adopting unified query-based criteria. Using only about 5.7K high-quality data curated from the open-source preference dataset, our CE-RM-4B achieves superior performance on diverse reward model benchmarks, especially in Best-of-N scenarios, and delivers more effective improvements in downstream RL practice.