SkepticalSam
Is thinking less sometimes better?
Thinkless: LLM Learns When to Think
Published: 12 May 2026 · Updated: 13 July 2026
Read the original sourceWhat the paper says
Thinkless proposes training a model to identify tasks where reasoning adds little value and to answer directly. They claim that skipping reasoning improves speed and sometimes accuracy.
The Critique
The evaluation tasks are simplistic (e.g., factual recall), and the conclusion that reasoning is unnecessary may be premature. The paper risks encouraging models to skip reasoning when it is actually needed.
Why It Matters
Over-reasoning wastes resources and can introduce errors; understanding when to think is valuable.
What They Missed
There is no examination of safety consequences when the model erroneously decides not to reason on high-stake questions.
The Big Question
How can we trust a model to decide when reasoning is optional, and what happens when it guesses wrong?