Researchers propose Truth AnChoring method to fix unreliable uncertainty detection in large language models by calibrating metrics against actual factual correctness.

Uncertainty estimation (UE) metrics in LLMs often fail to reliably detect hallucinations due to 'proxy failure'—they measure model behavior rather than actual factual accuracy

UE metrics become unreliable in low-information scenarios where they struggle to discriminate between correct and incorrect outputs

Truth AnChoring (TAC) is a post-hoc calibration method that remaps raw uncertainty scores to truth-aligned scores, improving reliability even with limited training data

The approach enables better-calibrated uncertainty estimates and provides a practical calibration protocol for improving LLM reliability