Reproduce “Thinking Inertia: LLMs Keep Thinking When Told Not To” 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
Dianqiao Lei, Kevin Qinghong Lin, Pan Lu, Philip Torr, James Zou
The study measures whether models show visible inference before answering when instructed not to think. It separates answer-only compliance, question relevance, and explicit inference, then compares answer formats on matched questions. For Qwen3-32B, open-ended responses had 0% answer-only compliance and 90.14% visible inference, versus 62.5% and 36.90% for Boolean verification; this concerns output, not hidden computation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.