From Outcomes to Processes: Guiding PRM Learning from ORM for Inference-Time Alignment
This addresses a critical bottleneck in aligning LLMs with human preferences more effectively, though it is an incremental improvement over existing reward-guided search methods.
The paper tackles the granularity mismatch in inference-time alignment of large language models by introducing process reward models (PRMs) to replace outcome reward models (ORMs) in reward-guided search, resulting in a 3.6%-10.3% improvement in GPT-4 evaluation scores across dialogue, summarization, and reasoning tasks.
Inference-time alignment methods have gained significant attention for their efficiency and effectiveness in aligning large language models (LLMs) with human preferences. However, existing dominant approaches using reward-guided search (RGS) primarily rely on outcome reward models (ORMs), which suffer from a critical granularity mismatch: ORMs are designed to provide outcome rewards for complete responses, while RGS methods rely on process rewards to guide the policy, leading to inconsistent scoring and suboptimal alignment. To address this challenge, we introduce process reward models (PRMs) into RGS and argue that an ideal PRM should satisfy two objectives: Score Consistency, ensuring coherent evaluation across partial and complete responses, and Preference Consistency, aligning partial sequence assessments with human preferences. Based on these, we propose SP-PRM, a novel dual-consistency framework integrating score consistency-based and preference consistency-based partial evaluation modules without relying on human annotation. Extensive experiments on dialogue, summarization, and reasoning tasks demonstrate that SP-PRM substantially enhances existing RGS methods, achieving a 3.6%-10.3% improvement in GPT-4 evaluation scores across all tasks.