mlsanityData to 14 Oct 2026

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

arXiv 2606.21203 ↗First posted 19 Jun 2026cs.CL cs.AIcited by 0

Not on the map. Why?

Abstract (arXiv)

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 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.

Key facts. No citing papers in the map yet (data through 14 Oct 2026). Builds on 26 of its 53 references that are in the map. First posted on arXiv 19 Jun 2026.

Builds on

26 of 53 references.

References · 26

  1. 2017Attention Is All You Needcited by 60,361
  2. 2022Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 smallcited by 1,165
  3. 2021Probing Classifiers: Promises, Shortcomings, and Advancescited by 876
  4. 2023Large Language Models Can Be Easily Distracted by Irrelevant Contextcited by 841
  5. 2020Hopfield Networks is All You Needcited by 475
  6. 2023Making Retrieval-Augmented Language Models Robust to Irrelevant Contextcited by 369
  7. 2016Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependenciescited by 354
  8. 2016Dense Associative Memory for Pattern Recognitioncited by 236
  9. 2022Why Does Surprisal From Larger Transformer-Based Language Models Provide a Poorer Fit to Human Reading Times?cited by 113
  10. 2024EuroLLM: Multilingual Language Models for Europecited by 106
  11. 2018Under the Hood: Using Diagnostic Classifiers to Investigate and Improve how Language Models Track Agreement Informationcited by 97
  12. 2023Testing the Predictions of Surprisal Theory in 11 Languagescited by 77
14 more
  1. 2025MultiBLiMP 1.0: A Massively Multilingual Benchmark of Linguistic Minimal Pairscited by 71
  2. 2020As Good as New. How to Successfully Recycle English GPT-2 to Make Models for Other Languagescited by 54
  3. 2024CausalGym: Benchmarking causal interpretability methods on linguistic taskscited by 46
  4. 2023Quantifying Context Mixing in Transformerscited by 43
  5. 2025Large Language Models Are Human-Like Internallycited by 29
  6. 2021Accounting for Agreement Phenomena in Sentence Comprehension with Transformer Language Models: Effects of Similarity-based Interference on Surprisal and Attentioncited by 27
  7. 2025Causal Interventions Reveal Shared Structure Across English Filler-Gap Constructionscited by 21
  8. 2024Do LLMs dream of elephants (when told not to)? Latent concept association and associative memory in transformerscited by 15
  9. 2022Entropy- and Distance-Based Predictors From GPT-2 Attention Patterns Predict Reading Times Over and Above GPT-2 Surprisalcited by 14
  10. 2023Language Resources for Dutch Large Language Modellingcited by 11
  11. 2025NRGPT: An Energy-based Alternative for GPTcited by 8
  12. 2025MeMo: Towards Language Models with Associative Memory Mechanismscited by 5
  13. 2024GEITje 7B Ultra: A Conversational Model for Dutchcited by 4
  14. 2026Diverging Transformer Predictions for Human Sentence Processing: A Comprehensive Analysis of Agreement Attraction Effectscited by 2

Built on by

Not cited yet.