New study finds longer chain-of-thought can make LLMs overconfident

A new arXiv paper finds that increasing chain-of-thought reasoning budgets can sometimes make LLMs more overconfident, not more reliable.

The authors call the effect Calibration Drift Under Reasoning. In their framing, longer reasoning can first correct errors, then push the model into internally consistent but wrong explanations that receive high confidence.

The result is important for deployment because many systems treat longer reasoning as a safety or accuracy upgrade. Calibration may need to be measured separately from answer quality.