Reproduce “ReSAIL: Mitigating Collapse in Iterative Agent 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
Shengjie Jin, Hengbo Xu, Zelong Sun, YuJie Guo, Zhiwu Lu
ReSAIL improves iterative self-distillation by selecting interaction steps where privileged information changes a teacher’s predictions, balancing losses across trajectories, and retaining the student’s privileged-view behavior. On ALFWorld and TextCraft, it yields an average 22.5-percentage-point final-cycle success gain over paired baselines across 12 comparisons. Evidence covers two benchmarks, two model sizes, and three cycles, with higher training cost.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.