Reproduce “Training LLMs to Follow Reasoning Instructions with Minimal Edits” 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

Alexandra Yakovleva, Luat Gia Khoi Nguyen, Juho Kannala, Harri Valpola, Alexander Ilin

The authors introduce a method that repairs the first violation of a reasoning instruction in a model’s response, resumes generation, and trains only on corrected spans. On Qwen3-32B, Offline reaches 87 ± 1% in-domain accuracy, compared with 40.3 ± 0.4% for same-trajectory SFT. Offline scores 68.4 ± 0.4% on unseen instructions, above Online’s 34 ± 9%.

Summary by alphaXiv.

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