Reproduce “MS-TTRL: Breaking the Frequency Conflict in Multiple-Solution Test-Time Reinforcement Learning” 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
Zhixia Zhang, Binqian Xu, Huaiyang Wang, Yikun Ban, Dianbo Liu
MS-TTRL addresses multiple-answer reasoning with a label-free training method. It uses the same model to verify outcomes, then applies inverse-frequency weighting only to outcomes judged correct, using frequency to encourage diversity. On 3D Structure Reconstruction with Qwen3-1.7B, it reaches 28.39 ± 0.04% Recovery Rate versus 9.07 ± 0.00% for TTRL and TTRL-CoCoV, a five-seed result not uniform across benchmarks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.