From Relation-Specific Heads to Shared Interventions: Contextual Entrainment in Language Models

Paper recorded by Signals 4 on 2026-10-13 in cs.CL. Abstract reproduced from arXiv; link to the original below.

Published 2026-10-13 on arXiv · recorded by Signals 4 on 2026-10-14

Category: cs.CL · 自然语言处理 · first seen 2026-10-14

Abstract

Large language models rely on prompt context to perform in-context learning, follow formats, and adapt to local evidence. This ability, however, also raises a robustness concern: models may reuse irrelevant tokens from earlier context when answering a later query. Prior work identifies this behavior as contextual entrainment and finds attention heads whose ablation reduces it, arguing that these h

Read on arXiv →