EasyDeepLearn

How well-calibrated are LLM confidences and how do you fix them?

hard

Answer

  • Base LLMs are decently calibrated (token probabilities match observed accuracy in-distribution).
  • RLHF-tuned LLMs are overconfident — asked 'how sure are you?' they say 90% but are right 60% of the time.
  • Fixes: (1) verbalized confidence prompts + calibration on a labeled dev set; (2) sampling-based confidence (fraction of K samples that agree); (3) explicit calibration training (e.g., True/False + Confidence-tuning); (4) ensembling.
  • Calibration matters for downstream decisions: threshold on the LLM's confidence before deciding to escalate to a human.
Check yourself — multiple choice
  • LLMs are perfectly calibrated
  • Base LLMs OK, RLHF-tuned are overconfident; use verbalized confidence + calibration on dev set / sampling agreement / ensembling
  • Only classification
  • Impossible to calibrate

LLM calibration: RLHF makes overconfident; verbal + sampling + calibration fix.

#reliability#evaluation#alignment

Practise LLMs & GenAI

214 interview questions in this topic.

Related questions