Reproduce “Consolidating Reasoning with Test-Time Training” 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
Haotian Wu, Zeming Chen, Hao Zhao, Antoine Bosselut
TTT-Consolidate uses test-time gradient updates to adapt a small LoRA module on several solution trajectories, then solves the question anew while keeping the base model frozen. This lets trajectories interact in adapted parameters rather than only in a long prompt. For Qwen3-1.7B, in the None tier where all eight trajectories end with wrong answers, it averages 23.58% accuracy across ten math and STEM benchmarks, versus 13.02% for the strongest baseline.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.