LLMs Use Confidence to Decide When to Answer or Abstain, Study Finds

Key Info

A new open-access paper in Nature Machine Intelligence from Google DeepMind and Princeton provides causal evidence that LLMs use confidence to guide their behavior—boosting or suppressing confidence changes whether the model answers or abstains.

Highlights

• Shows LLMs do not just passively report confidence; it actively influences their decision to respond or abstain.
• Causal manipulation of confidence directly alters model behavior, offering a clearer path to controlling reliability.
• Open-access publication makes the findings and methods freely available to researchers and developers.