Generative RLHF-V: Learning Principles from Multi-modal Human Preference

Jiayi Zhou, Jiaming Ji, Boyuan Chen, Jiapeng Sun, Wenqi Chen, Donghai Hong, Sirui Han, Yike Guo, Yaodong Yang

Keywords: Alignment, Safety, RLHF, Preference Learning, Multi-modal LLMs

TL;DR: A novel alignment framework that integrates generative reward models with multi-modal RLHF.

Abstract: Training multi-modal large language models (MLLMs) that align with human intentions is a long-term challenge. Traditional score-only reward models for alignment suffer from low accuracy, weak generalization, and poor interpretability, blocking the progress of alignment methods, \textit{e.g.,} reinforcement learning from human feedback (RLHF). Generative reward models (GRMs) leverage MLLMs' intrinsic reasoning capabilities to discriminate pair-wise responses, but their pair-wise paradigm makes it hard to generalize to learnable rewards. We introduce Generative RLHF-V, a novel alignment framework that integrates GRMs with multi-modal RLHF. We propose a two-stage pipeline: \textbf{multi-modal generative reward modeling from RL}, where RL guides GRMs to actively capture human intention, then predict the correct pair-wise scores; and \textbf{RL optimization from grouped comparison}, which enhances multi-modal RL scoring precision by grouped responses comparison. Experimental results demonstrate that, besides out-of-distribution generalization of RM discrimination, our framework improves 4 MLLMs' performance across 7 benchmarks by 18.1\%, while the baseline RLHF is only 5.3\%. We further validate that Generative RLHF-V achieves a near-linear improvement with an increasing number of candidate responses.

Primary Area: Social and economic aspects of machine learning (e.g., fairness, interpretability, human-AI interaction, privacy, safety, strategic behavior)

Submission Number: 12568