Tsinghua et al. Identify 'Thinking Inertia' in LLMs Despite Disabled Reasoning Modes

1 reports archived1 independent sourcesupdated 10/25/2026, 12:17:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Researchers from Tsinghua University, Oxford, and Stanford discovered that LLMs may still generate reasoning steps before the final answer even when 'thinking modes' are explicitly disabled and direct responses are requested. This phenomenon, termed 'Thinking Inertia,' is more pronounced with open-ended questions. The study introduces a three-tier evaluation framework (ETR, QRel, EIR) to quantify this behavior, highlighting limitations in current reasoning toggle mechanisms.

LATEST/Researchers from Tsinghua University, Oxford, and Stanford discovered that LLMs may still generate reasoning steps before the final answer even when 'thinking modes' are explicitly disabled and direct responses are requested. This phenomenon, termed 'Thinking Inertia,' is more pronounced with open-ended questions. The study introduces a three-tier evaluation framework (ETR, QRel, EIR) to quantify this behavior, highlighting limitations in current reasoning toggle mechanisms.

Hourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

Coverage timeline

Total 1 reports · Latest first

新智元 (微信公众号)T2Oct 25, 12:17·78 pts
Tsinghua et al. Identify 'Thinking Inertia' in LLMs Despite Disabled Reasoning Modes
- Disabling thinking switches does not stop reasoning; models may still produce implicit or explicit intermediate steps.
- Proposes a three-tier metric system: ETR (strict format), QRel (relevance), and EIR (explicit reasoning) to accurately identify thinking inertia.
- Higher question openness correlates with increased thinking inertia, posing potential risks for low-latency scenarios.
View Related Reports & Analysis Original Source