Attention-Path Fragility as an Uncertainty Signal in Large Language Models
2026-08-11 • Computation and Language
Computation and LanguageArtificial Intelligence
AI summaryⓘ
The authors suggest that a model's uncertainty about an answer isn't just about how unsure its output looks, but also whether confident answers fall apart when parts of its attention are changed. They created a method called ASMI that checks how much different attention setups in the model agree on answers, without needing extra training. They found ASMI helps spot errors especially when the model is confident but fragile, improving error prediction beyond usual confidence measures. ASMI works best when the answer depends on given context, but less so when the model just recalls facts it has learned. Their tests show ASMI often matches or beats other methods that need more computation.
model uncertaintyattention mechanismmutual informationBALDconfidence estimationgrounded question answeringsemantic agreement kernelparametric knowledgeerror predictionsubnetwork fragility
Authors
Minsoo Kim, Sungyoung Ji, Kisung Moon, Ilyong Yoon
Abstract
We propose that a model's uncertainty about a token is reflected not only in the breadth of its output distribution but also in whether a confident prediction is \emph{fragile} under perturbation of its attention pathways. We instantiate this as ASMI (Attention-Subnetwork Mutual Information), a training-free estimator that masks attention heads and measures the BALD mutual information among the resulting subnetworks, with a semantic-agreement kernel to discount surface-form disagreement. The signal is not a restatement of output confidence: on grounded QA an out-of-fold test shows it adds error-predictive information beyond single-pass confidence and entropy, concentrated in \emph{confident-but-fragile} predictions, where acting on it roughly halves the retained error of a confidence filter. The distinctness is regime-graded, so ASMI predicts its own domain of applicability, strong where answers are routed through provided context and bounded by design where they are recalled from parametric knowledge. Sem-ASMI reads the signal from a single greedy response, without the stochastic generations the strongest baselines require, and ties or beats Semantic Entropy on ten of the twelve grounded benchmark-backbone settings. Across the same twelve settings, the best ASMI variant, typically the adaptive one reusing the ten samples already drawn for the baselines, ties or leads the strongest baseline in eight, significantly in three under a paired test. On parametric QA all variants revert to or below the zero-cost MSP baseline, exactly as predicted, and the estimates are near-deterministic across reruns. A head-level analysis shows that what tracks this boundary is not the presence of head-level fragility but whether that fragility couples to errors.