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

The study investigates coherence illusions in Dutch language models, revealing that models exhibit similar behavior to humans when presented with incoherent texts that contain matching distractors. Key findings include the use of surprisal, attention entropy, and energy metrics to assess coherence, with models demonstrating reduced surprisal for incoherent continuations when distractors are present. This research highlights the potential for these metrics to inform understanding of coherence mechanisms in LLMs, providing insights for practitioners in model evaluation and fine-tuning.