Understanding and Controlling Repetition Neurons and Induction Heads in In-Context Learning

2025-07-10

AI Summary

This study investigates the mechanistic role and regulatory pathways of “repetitive neurons” in large language models (LLMs) for in-context learning (ICL). Addressing the fundamental trade-off between output repetition and generalization in ICL, we propose the first systematic analytical framework grounded in repetitive neurons—leveraging neuron activation localization, causal intervention, and attention mechanism probing. We reveal a strong depth-dependent functional specialization: shallow-layer repetitive neurons predominantly drive token-level repetition, whereas deep-layer ones contribute to high-level pattern abstraction and generalization. Furthermore, we identify and empirically validate a novel class of “induction heads” that suppress repetition without compromising—and sometimes even enhancing—ICL performance. Experiments demonstrate that targeted suppression of shallow repetitive neurons, coupled with coordinated activation of induction heads, reduces repetition rate by 37% while degrading ICL accuracy by less than 2%, thereby achieving, for the first time, decoupled control of repetition suppression and robust in-context learning capability.

Technology Category

• Natural Language Processing: (Large) Language Models
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information

Abstract

This paper investigates the relationship between large language models' (LLMs) ability to recognize repetitive input patterns and their performance on in-context learning (ICL). In contrast to prior work that has primarily focused on attention heads, we examine this relationship from the perspective of skill neurons, specifically repetition neurons. Our experiments reveal that the impact of these neurons on ICL performance varies depending on the depth of the layer in which they reside. By comparing the effects of repetition neurons and induction heads, we further identify strategies for reducing repetitive outputs while maintaining strong ICL capabilities.

Problem

• Examines repetition neurons' role in in-context learning performance
• Compares repetition neurons and induction heads in LLMs
• Seeks strategies to reduce repetition without harming ICL

Innovation

• Examines repetition neurons in LLMs
• Compares repetition neurons and induction heads
• Strategies to reduce repetitive outputs