RM-Distiller: Exploiting Generative LLM for Reward Model Distillation
This addresses the challenge of obtaining high-quality human preference annotations for aligning LLMs, offering a systematic approach to improve reward modeling, though it is incremental in leveraging existing teacher capabilities more fully.
The paper tackles the problem of distilling human preferences from generative LLMs for reward modeling by proposing RM-Distiller, which exploits teacher LLMs' refinement, scoring, and generation capabilities, resulting in significant outperformance over traditional methods on benchmarks and alignment tasks.
Reward models (RMs) play a pivotal role in aligning large language models (LLMs) with human preferences. Due to the difficulty of obtaining high-quality human preference annotations, distilling preferences from generative LLMs has emerged as a standard practice. However, existing approaches predominantly treat teacher models as simple binary annotators, failing to fully exploit the rich knowledge and capabilities for RM distillation. To address this, we propose RM-Distiller, a framework designed to systematically exploit the multifaceted capabilities of teacher LLMs: (1) Refinement capability, which synthesizes highly correlated response pairs to create fine-grained and contrastive signals. (2) Scoring capability, which guides the RM in capturing precise preference strength via a margin-aware optimization objective. (3) Generation capability, which incorporates the teacher's generative distribution to regularize the RM to preserve its fundamental linguistic knowledge. Extensive experiments demonstrate that RM-Distiller significantly outperforms traditional distillation methods both on RM benchmarks and reinforcement learning-based alignment, proving that exploiting multifaceted teacher capabilities is critical for effective reward modeling. To the best of our knowledge, this is the first systematic research on RM distillation from generative LLMs.