The article argues that LLM confidence scores are fundamentally unreliable because they are generated by the same next-token prediction mechanism as all other output, making '90% confident' no more meaningful than any other word choice. It distinguishes epistemic uncertainty (reducible with more data) from aleatoric uncertainty (irreducible), and proposes replacing point estimates with full probability distributions over possible outcomes as a path toward addressing hallucination via RLCD.
Background
Calibration of LLM confidence scores has been a persistent challenge; models tend to produce overconfident-sounding outputs regardless of actual accuracy. RLCD (likely Reinforcement Learning from Confidence/Distribution data) represents a newer approach to training models to express calibrated uncertainty rather than relying on implicit next-token fluency.
- Source
- Lobsters
- Published
- Sep 28, 2026 at 11:04 PM
- Score
- 5.0 / 10