Late Attention Layers Alone Can Copy Entity Tokens, but Not Without Attending to Their Context

2026-09-28

AI Summary

This study addresses the unclear roles of specific layers and contextual influence mechanisms in entity copying within large language models. Using Qwen3-8B as the testbed, this work proposes the "Genie in a Bottle" method alongside an attention ablation technique, conducting controlled experiments through layer intervention and attention masking. The research precisely localizes functional layers and information dependencies, revealing that a two-layer block in the latter half of the model is both necessary and sufficient for entity copying while confirming that contextual attention is indispensable. By establishing the pivotal role of late layers under contextual guidance and elucidating information propagation pathways, this project provides a novel paradigm for understanding the internal mechanisms of large language models.

Abstract

Large language models (LLMs) reliably perform entity copying, in which a model copies tokens referring to an entity, termed entity tokens, from the prompt into its output to answer a question. Although entity copying is straightforward for most LLMs, existing research does not provide a systematic account of which layers specialize in this fundamental task or how other tokens in the same sequence, termed context tokens, influence the model's ability to copy the entity tokens. To address these questions, we conduct experiments on Qwen3-8B using two novel methods: genie-in-a-bottle, which controls exactly which layers can participate in an entity-copying task, and attention lobotomy, which cuts off specific tokens'attention to entity tokens without affecting the remaining attention distribution. We find that two distinct groups of layers in the second half of the model are both necessary and sufficient for entity copying. Moreover, in addition to the decoding position's attention to entity tokens, context tokens'attention to entity tokens also proves necessary for copying the exact tokens, even though context tokens do not store entity information themselves unless they satisfy particular semantic properties. Our findings establish the critical role of late layers in entity copying under the guidance of context tokens, calling for future work on how models propagate and consume entity information.

Problem

• entity copying
• large language models
• attention layers
• context tokens

Innovation

• Entity Copying
• Genie-in-a-bottle
• Attention Lobotomy
• Late Attention Layers
• Context Tokens