Reproduce “MicroMouse: Masking Harmful Supervision in Long CoT via Relative Potential” 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
Xiaowei Zhu, Hengzhu Tang, Zhenyu Zhang, Shuohuan Wang, Yubing Ren, Yanan Cao
MicroMouse masks loss on selected steps in reasoning traces while retaining the full trace as context. It compares scores for a correct answer and a format- and length-matched wrong answer, masking decreases and steps after the score peaks. Across three models, it reports average relative gains of 5.5% in greedy accuracy and 2.6% in Pass@1 over full-trace fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.