“Fragility of Attention Paths”: a new metric reveals uncertainty of language models in context-guided answers

A team of researchers, including Minsu Kim, Sungyoung Ji, and Kiseon Moon, and Eilyeung Yoon, presented a new study aimed at revealing the uncertainty in large language models by observing the fragility of internal attention paths. The paper, titled Fragility of Attention Path as a Signal of Uncertainty in Large Language Models, deposited in the arXiv archive under the computer science categories of language and artificial intelligence on August 11, 2026, is based on the hypothesis that the model's uncertainty about tokens is not limited to the breadth of the final output distribution, but also extends to the stability of confident predictions or their fragility when disturbances are introduced to the attention paths.

To measure this phenomenon, the researchers developed an estimation tool that does not require retraining, known as the mutual information of sub-attention networks. This tool masks attention heads and measures the mutual information according to the Bald criterion between the resulting sub-networks. The tool also relies on a semantic consistency kernel aimed at excluding differences in the surface formulation of texts. Independent tests in context-based question-answering tasks showed that this signal provides predictive information about errors that exceeds the confidence obtained from a single pass and entropy, and this effectiveness is concentrated in predictions that are apparently confident but fragile in their internal structure, where the adoption of this signal led to a reduction of the remaining error of the confidence filter by approximately half.

The study shows that the distinctiveness of this signal varies according to the mode of operation, making the tool capable of predicting its own scope of application, where it works strongly when answers depend on guidance through the attached context, while the limits of its effectiveness decline by design when answers are retrieved from the procedural knowledge stored within the model. The semantic sequence of the tool was able to read the signal based on a single response and without the need for repeated random generations required by strong reference methods, achieving a tie or superiority over the semantic entropy method in ten out of twelve reference settings based on context.

Across the same twelve settings, the best adaptive pattern of the tool, which reuses the ten samples drawn for the reference methods, succeeded in tying with the strongest reference method or surpassing it in eight settings, including three settings that showed statistically significant superiority according to a paired test.When moving to question-answering tasks based on procedural knowledge, all variations returned to the level of the zero-cost reference point or slightly below it, just as the researchers expected, with nearly deterministic estimates appearing across repeated rounds of operation. The extended analysis at the level of attention heads showed that the boundary between the two patterns is not related to the mere presence of fragility in the heads, but rather to the association of that fragility with actual errors, according to the paper, which is nineteen pages long and under scientific review.