Contextualize-then-Aggregate: Circuits for In-Context Learning in Gemma-2 2B

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Bakalova, Aleksandra, Veitsman, Yana, Huang, Xinting, Hahn, Michael
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866909792216285184
author Bakalova, Aleksandra
Veitsman, Yana
Huang, Xinting
Hahn, Michael
author_facet Bakalova, Aleksandra
Veitsman, Yana
Huang, Xinting
Hahn, Michael
contents In-Context Learning (ICL) is an intriguing ability of large language models (LLMs). Despite a substantial amount of work on its behavioral aspects and how it emerges in miniature setups, it remains unclear which mechanism assembles task information from the individual examples in a fewshot prompt. We use causal interventions to identify information flow in Gemma-2 2B for five naturalistic ICL tasks. We find that the model infers task information using a two-step strategy we call contextualize-then-aggregate: In the lower layers, the model builds up representations of individual fewshot examples, which are contextualized by preceding examples through connections between fewshot input and output tokens across the sequence. In the higher layers, these representations are aggregated to identify the task and prepare prediction of the next output. The importance of the contextualization step differs between tasks, and it may become more important in the presence of ambiguous examples. Overall, by providing rigorous causal analysis, our results shed light on the mechanisms through which ICL happens in language models.
format Preprint
id arxiv_https___arxiv_org_abs_2504_00132
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Contextualize-then-Aggregate: Circuits for In-Context Learning in Gemma-2 2B
Bakalova, Aleksandra
Veitsman, Yana
Huang, Xinting
Hahn, Michael
Computation and Language
Machine Learning
In-Context Learning (ICL) is an intriguing ability of large language models (LLMs). Despite a substantial amount of work on its behavioral aspects and how it emerges in miniature setups, it remains unclear which mechanism assembles task information from the individual examples in a fewshot prompt. We use causal interventions to identify information flow in Gemma-2 2B for five naturalistic ICL tasks. We find that the model infers task information using a two-step strategy we call contextualize-then-aggregate: In the lower layers, the model builds up representations of individual fewshot examples, which are contextualized by preceding examples through connections between fewshot input and output tokens across the sequence. In the higher layers, these representations are aggregated to identify the task and prepare prediction of the next output. The importance of the contextualization step differs between tasks, and it may become more important in the presence of ambiguous examples. Overall, by providing rigorous causal analysis, our results shed light on the mechanisms through which ICL happens in language models.
title Contextualize-then-Aggregate: Circuits for In-Context Learning in Gemma-2 2B
topic Computation and Language
Machine Learning
url https://arxiv.org/abs/2504.00132