Reproduce “TTCS: Test-Time Curriculum Synthesis for Self-Evolving” 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

Chengyi Yang, Zhishang Xiang, Yunbo Tang, Zongpei Teng, Chengsong Huang, Fei Long, Yuhan Liu, Jinsong Su

TTCS addresses test-time learning on difficult, unlabeled reasoning questions by using each test question to anchor related synthetic problems. It trains a Synthesizer and Solver together, adapting question difficulty as the solver changes and using sampled-answer agreement as a pseudo-label reward. On Qwen2.5-Math-1.5B, TTCS scores 19.79 mean@32 on AIME24, versus 13.23 for TTRL; this is one model and benchmark.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.