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

2026-08-11

AI Summary

This work addresses the challenge that large language models often produce highly confident yet incorrect predictions, making reliable uncertainty estimation difficult. The authors propose Sem-ASMI, a training-agnostic uncertainty quantification method that, for the first time, links the fragility of attention subnetworks to prediction reliability. By perturbing attention heads and integrating BALD-based mutual information estimation with a semantic consistency kernel, Sem-ASMI identifies “high-confidence but fragile” errors within a single greedy decoding pass—eliminating the need for costly repeated sampling and automatically adapting to in-domain contexts. Evaluated across 12 grounded question-answering tasks, Sem-ASMI matches or outperforms the strongest baseline on 10 tasks and achieves significant gains on 3. Notably, it naturally degenerates to Maximum Softmax Probability (MSP) in parametric QA settings, confirming its domain adaptability.

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.

Problem

• uncertainty estimation
• attention fragility
• large language models
• confident-but-fragile predictions
• grounded QA

Innovation

• attention-path fragility
• uncertainty estimation
• training-free
• mutual information
• semantic agreement