When Context Misleads: Surprisal, Energy and Attention Entropy as Metrics of Coherence Illusions in LLMs

2026-06-19

AI Summary

This study investigates whether Dutch large language models are susceptible to the "illusion of coherence"—erroneously judging incoherent texts as coherent when distractor words match subsequent content. By analyzing monolingual and multilingual models in contexts containing discourse connectives (e.g., “again,” “too”), and integrating human acceptability judgments with eye-tracking data, the work introduces “energy” as a novel quantitative metric for discourse coherence and employs attention entropy to identify coherence-sensitive attention heads. The findings reveal that model surprisal aligns with human judgments, with distractors attenuating surprise at incoherence; specific attention heads prove critical for coherence assessment, and their ablation yields cross-task transfer effects; furthermore, the energy metric effectively captures discourse coherence, highlighting a convergence between model behavior and human cognitive mechanisms.

Abstract

Psycholinguistics studies show that human readers fall for coherence illusions: an incoherent discourse can seem coherent simply because a distractor matches what comes next. We investigate whether Dutch language models (6 monolingual and 4 multilingual) show the same behavior on texts that link back to earlier context with words such as 'again' and 'too'. First, we find that surprisal at the critical word tracks human acceptability judgments and eye-tracking data. Models are more surprised by incoherent continuations, but a matching distractor in the prior context reduces this surprisal. Second, attention entropy at the critical position identifies heads that behave differently under coherence vs. incoherence. We find that ablating these heads shows transfer effects across experiments, suggesting a shared mechanism. Third, we introduce energy from the associative-memory literature as a metric to quantify discourse coherence. Taken together, our results show that coherence illusions arise in Dutch LLMs, with entropy and energy exposing mechanisms that operate across settings.

Problem

• coherence illusions
• language models
• discourse coherence
• context effects
• psycholinguistics

Innovation

• surprisal
• attention entropy
• energy
• coherence illusions
• language models