Reproduce “Self-Improving Training & Thinking” 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
Ellen Xiaoqing Tan, Jack Lanchantin, Zichun Yu, Polina Kirichenko, Shehzaad Dhuliawala, Danwei Li, Thao Nguyen, Jing Xu and 6 more
The authors present three methods that guide language models earlier in training or add self-judgment after generation. The methods use model-generated rewrites and judgments, teacher-inserted thoughts, or post-answer critiques. On Llama 3-8B, thinking mid-training followed by reinforcement learning post-training reaches 0.3837 average accuracy across five reasoning benchmarks, versus 0.1197 for direct post-training; this benchmark result does not establish deployment reliability.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.