Reproduce “The First Few Tokens Matter: Understanding Chain-of-Thought Controllability” 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

Hugh Van Deventer, Dmitrii Troitskii, Laura Ruis

The authors test whether habitual reasoning openings limit models’ compliance with instructions about their reasoning traces. In Qwen3-32B and OLMo-3.1-32B-Think, banning common first tokens and combining the ban with a compliant example raised mean compliance across eight tasks from 0.078 to 0.490 and from 0.029 to 0.374, respectively. This supports an initiation bottleneck in these proxy tasks, but not monitor evasion.

Summary by alphaXiv.

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