Reproduce “Beyond Reward Scoring: Online Multi-Agent Preference Learning for Multimodal Response Generation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ziyun Zou, Jiwoo Hong, Yifan Lu, Shao Tang, Weiyi Wu, Aman Gupta, Jin Tian, Imran Razzak
PrefMind trains a Generator and Critic to discuss answer preferences, correct mistaken judgments, and refine responses. The Generator evaluates the Critic’s judgment; a human-label-verified gate determines whether refinement proceeds. Refinement starts from the human-preferred answer. At 4B, the Qwen3-VL PrefMind Generator improves the eight-metric generation average by 2.4 points over Base, leading at 4B and 8B, but not 2B.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.