ASMI measures uncertainty by masking attention heads, halving retained error on confident-but-fragile predictions.

The argument is that token uncertainty shows up not only in output-distribution breadth but in whether a confident prediction is fragile under perturbation of its attention pathways (arXiv 2608.11138). It's training-free: mask attention heads, measure BALD mutual information among resulting subnetworks through a semantic-agreement kernel. Single-response variant ties or beats Semantic Entropy on 10 of 12 grounded benchmark-backbone settings. The honest limit the authors state plainly: on parametric QA all variants revert to or below the zero-cost MSP baseline. This helps for retrieval-grounded answers, not closed-book ones.