Uncertainty estimation aims to detect hallucinated outputs of large language models to improve reliability (arXiv:2604.00445, 2024). UE metrics often exhibit unstable performance across configurations, which limits applicability. The authors formalise this phenomenon as proxy failure because most UE metrics originate from model behaviour rather than being explicitly grounded in the factual correctness of LLM outputs.

UE metrics become non-discriminative precisely in low-information regimes. The proposed Truth AnChoring (TAC) method is a post-hoc calibration approach that maps raw scores to truth-aligned scores (arXiv:2604.00445, 2024). TAC supports learning of well-calibrated uncertainty estimates even with noisy and few-shot supervision.

The work presents a practical calibration protocol and highlights limitations of treating heuristic UE metrics as direct indicators of truth uncertainty (arXiv:2604.00445, 2024). TAC is positioned as a necessary step toward more reliable uncertainty estimation for LLMs. Code is available at https://github.com/ponhvoan/TruthAnchor/.