Reproduce “Turning Multi-Agent Communication into Token-Level Credit with On Policy Self Distillation” 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
Zixu Wang, Yu Zhao, Xing Zheng, Bingbing Xu, Tian Xueyun, JiaTang Luo, Weiqiang Wang, Bowen Song and 2 more
OPSD4MAS uses future messages in multi-agent mathematical reasoning to provide token-specific learning signals alongside outcome-based reinforcement learning. An external model extracts facts, which are filtered for their relationship to outcomes and their influence on a target response’s token probabilities. Starting from Qwen3-4B, FW4 reaches 83.33% AIME24 pass@4 versus 63.33% for DrMAS-GRPO; evidence is limited to mathematical tasks and model scale.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.