← Back to all news
arXiv Computation and Language

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

The study examines whether Dutch language models exhibit coherence‑illusion effects similar to human readers, using texts that refer back to earlier context with words like ‘again’ and ‘too’. Surprisal at the critical word aligns with human acceptability and eye‑tracking data, showing that models are more surprised by incoherent continuations unless a matching distractor is present. Attention entropy and an energy metric from associative‑memory literature reveal heads that behave differently under coherence versus incoherence, and ablating these heads demonstrates transfer effects across experiments, indicating a shared underlying mechanism.

Related stories

arXiv AI
Sep 2

Energy-Based Transformers as Predictors of Reading Difficulty

The paper investigates energy-based transformers as predictors of reading difficulty, extending the use of transformer language models in psycholinguistics. It demonstrates that the energy measure from these models robustly predicts reading times across multiple corpora, outperforming traditional metrics like surprisal and attention entropy. In a controlled experiment on relative clause processing, energy captures known asymmetries, suggesting it may unify previously complementary predictors.

By Jakub Dotlacil, Ece Takmaz
llms
More like this →
arXiv Computation and Language
Sep 11

Does Linguistic Structure Enrichment Enhance Coherence Assessment? Not With Current Architectures

The paper examines whether adding syntactic and rhetorical structure to text can improve the prediction of incoherence in large language model outputs. Experiments show that plain text actually yields higher accuracy, as the added structural information conflicts with the models’ architectures. The authors also demonstrate that coherence assessment can help detect misleading content by applying zero‑shot experiments to a Brazilian disinformation dataset.

By Victor Mazzotti, Luiz Pereira, Marina Bitencourt dos Santos, Helena Maia, Carlos Caetano, N\'adia Felix, Sandra Avila
llms
More like this →
arXiv AI
Sep 7

Patterns of Priming in Production: Lexical, Semantic and Structural Alignment in Language Model Generation

The paper studies structural priming in language model production by conducting controlled sentence‑completion experiments on dative constructions. Results show that language models exhibit priming effects, especially when sentences are semantically coherent, with stronger relative increases for double‑object datives and larger absolute increases for prepositional‑object datives. The study also finds that primed completions involve more lexico‑semantic repetition, indicating that priming operates across syntactic, lexical, and semantic levels.

By Giulia Pucci, Ruizhe Li, Arabella Sinclair
llmssafety
More like this →
arXiv Machine Learning
Jun 5

Masks Can Be Distracting: On Context Comprehension in Diffusion Language Models

arXiv:2511. 21338v2 Announce Type: replace Abstract: Masked Diffusion Language Models (MDLMs) have recently emerged as a promising alternative to Autoregressive Language Models (ARLMs), leveraging a denoising objective that, in principle, should enable more uniform context utilisation.

By Julianna Piskorz, Cristina Pinneri, Alvaro Correia, Motasem Alfarra, Risheek Garrepalli, Christos Louizos
llmsdiffusionfine-tuningsafety
More like this →
arXiv Computation and Language
Aug 27

Distinct dynamics of conceptual and referential disruptions in human reading and large language model processing

The study investigates how disrupting conceptual versus referential information in short narratives affects human reading and large language model (LLM) processing. In humans, conceptual disruptions cause a strong, localized processing cost that peaks early and declines quickly, while referential disruptions produce weaker, gradually decreasing effects that are more influenced by sentence boundaries. In LLMs, both disruptions appear immediately at the manipulated word; surprisal patterns mirror human reading, whereas output-layer representations show that referential disruption initially causes a larger displacement before both types decay following a power-law.

By Rui He, Nihal Altay, Wolfram Hinzen
llms
More like this →