Attention-Path Fragility as an Uncertainty Signal in Large Language Models

A recent study introduces a technique for assessing uncertainty in large language models by analyzing the vulnerability of confident predictions when attention pathways are altered. This method, called ASMI (Attention-Subnetwork Mutual Information), functions as a training-free estimator that evaluates BALD mutual information by masking attention heads and examining the resulting subnetworks, employing a semantic-agreement kernel to mitigate surface-form discrepancies. The findings indicate that this signal provides additional error-predictive insights beyond mere confidence and entropy, especially for 'confident-but-fragile' predictions, effectively reducing the retained error of a confidence filter by approximately half. ASMI's distinctiveness is graded by regime, enabling it to forecast its own applicability, particularly when answers are guided by given context. The research can be found on arXiv under identifier 2608.11138.