Hacker News
"Don't Ask an LLM for a Confidence Score": Models Cluster 78% of Answers on Just Three Values
A July 27 post argues self-reported confidence scores are structurally useless, citing research that models don't treat 0–100 as a continuous spectrum — 78% of responses land on three values — and that verbalized confidence is badly overconfident (Xiong et al.) and only calibrates after task-specific fine-tuning (Lin et al.). The recursion problem is the core claim: if you can't trust the answer, you can't trust the model's confidence in its confidence. The author recommends investing in retrieval quality and model self-correction instead, and points to semantic entropy (Farquhar et al.) as the credible alternative — the post presents no original experiments, only a synthesis.
↳ Follow the thread