Today for AI
HOT RADAR
llmHEAT 0.0°

AI CLUSTERED EVENT · 10/10/2026

Tsinghua, Oxford, and Stanford Study Reveals 'Thinking Inertia': LLMs Continue Reasoning Even When Thinking Is Disabled

1 reports archived0 independent sourcesupdated 10/10/2026, 12:10:00
Synthesis & Latest Updates

A new joint study reveals that LLMs exhibit strong 'thinking inertia,' continuing to reason internally even when explicit thinking features are disabled; DeepSeek-V4-Flash retained reasoning traces in 99.9% of open-ended answers. Forcing answer-only outputs increased compliance to ~40% but reduced accuracy by ~15 points, highlighting the critical role of explicit reasoning chains for complex tasks.

LATEST/A new joint study reveals that LLMs exhibit strong 'thinking inertia,' continuing to reason internally even when explicit thinking features are disabled; DeepSeek-V4-Flash retained reasoning traces in 99.9% of open-ended answers. Forcing answer-only outputs increased compliance to ~40% but reduced accuracy by ~15 points, highlighting the critical role of explicit reasoning chains for complex tasks.

HEAT TRENDHourly heat curve

52 fully observed hours
Now
0.0
Peak
12.713:00
24h change
-100%

The line compares only sources observed throughout; gap hours are interpolated to keep the trend continuous.

TIMELINECoverage timeline

Total 1 reports · Latest first
  1. Rohan Paul (@rohanpaul_ai)T2·78 pts
    • LLMs exhibit 'thinking inertia': they tend to generate internal reasoning processes for open-ended questions even when explicit thinking modes are disabled.
    • Empirical data: DeepSeek-V4-Flash retained reasoning traces in 99.9% of open-ended answers after disabling the thinking feature.
    • Performance trade-off: Forcing models to skip reasoning and answer directly improves format compliance but causes a significant drop in accuracy (~15%).
    • Implication: Short replies or missing think tags do not prove the model skipped reasoning; it may be happening implicitly.