Thinking Inertia: LLMs keep reasoning even when explicitly told not to, Stanford paper finds

_akhaliq · x · 2026-10-15

A Stanford-led NeurIPS 2026 paper (authors include Philip Torr and James Zou) studies the neglected counterpart of thinking modes: no-thinking behavior.

Key findings:

Common proxies for no-thinking (disabled thinking mode, absence of long traces) are unreliable — models may still emit reasoning
The paper normalizes responses into pre-answer trace + final answer and proposes three metrics: Empty-Thinking Rate, instruction-aware Question-Pre-answer Relevance, and LLM-as-judge Explicit Inference Rate
Across six prompting interventions on six LLMs and Boolean/multiple-choice/open-ended tasks, explicit no-thinking controls cannot reliably eliminate visible inference — "Thinking Inertia" persists and grows as the answer space opens
Accuracy stays stable on Boolean/multiple-choice; open-ended tasks reveal a trade-off between answer-only compliance and task accuracy; supplying candidate answers makes answer-only output easier

Conclusion: not-thinking is a non-trivial capability that should be evaluated systematically alongside reasoning ability. HF commenters note models were never explicitly trained to recognize "don't use reasoning tokens."

Original post →

Thinking Inertia: LLMs keep reasoning even when explicitly told not to, Stanford paper finds

More from Models