Sentence-Level Contextual Entrainment in Large Language Models

2026-06-22

AI Summary

This study investigates sentence-level context anchoring in large language models—the phenomenon whereby sentences in prompts, including counterfactual statements, are assigned anomalously high likelihoods during inference. Extending prior token-level analyses to the sentence level, the authors examine average log-probabilities across tokens in two task types across 26 models. Through cross-model comparisons, attention head ablation, and statistical analysis, they demonstrate that this behavior is widespread yet diminishes with increasing model scale. Crucially, they identify that only 2%–4% of attention heads predominantly drive this effect; deactivating these specific heads effectively mitigates the issue without degrading overall model performance, enabling precise and lossless intervention.

Abstract

Contextual entrainment, which is a newly discovered phenomenon in large language models (LLMs), refers to the tendency of a model to assign higher probabilities to tokens that appear in its context. In this work, we extend this phenomenon from the token level to the sentence level by examining the per-token mean log-probability of a sentence instead of the probabilities of individual tokens. We investigate sentence-level contextual entrainment across 26 LLMs from seven families and two datasets, which cover both subjective and objective tasks. We find that sentence-level contextual entrainment exists. This means that the sentences in the prompt (even if they are counterfactual statements) can significantly increase their probability during model inference time. As the model size increases, contextual entrainment gradually decreases. We also find that contextual entrainment is controlled by 2% to 4% of the attention heads. Turning off these attention heads can effectively mitigate contextual entrainment without hurting the model's performance.

Problem

• contextual entrainment
• large language models
• sentence-level probability
• counterfactual statements
• attention heads

Innovation

• sentence-level contextual entrainment
• large language models
• attention heads
• log-probability
• model inference