Reproduce “Teach-to-Reason: Competition-Guided Reasoning with a Self-Improving Teacher” 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
Xiao Han, Hao Liu, Zhimin Bao, Jile Jiao, Yue Wang, Hui Guo, Xiaofeng Mou, Yi Xu
T2R trains a chest X-ray question-answering model to improve written reasoning by comparing Reasoner explanations with examples from a Teacher that competes against frozen copies. It preserves task-reward distinctions when responses differ, and uses comparison scores when all responses have identical rewards. On Covid19_heywhale, 2B T2R-R3 reports 40.37 accuracy versus 31.88 for Frozen Teacher; this does not establish clinical safety.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.