On the Additive Compositionality of Task Vectors in Vision–Language Models

Read full text Summarise with VIRAT
Authors: Yuting Shi, Houjing Wei, Naoya Inoue
Cited by: 0
References: 20

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

In-context learning (ICL) in large language models (LLMs) has been shown to operate through task vectors-the representation that summarizes the mapping induced by in-context demonstrations and can be composed by simple arithmetic operations.While this phenomenon is well studied in LLMs, its extension to visionlanguage models (VLMs) remains underexplored.In this work, we systematically examine the additive compositionality of in-context task vectors in VLMs, extracted from textside hidden representations.Specifically, we construct compositional visual reasoning tasks with clearly defined subtasks and extract task vectors from few-shot demonstrations.Empirical experiments show that the vector for a complex task can be approximated by adding the vectors of its constituent subtasks.Beyond this, we analyze token-level contextual embeddings and show that additive composition arises because complex-task representations emerge as the superposition of atomic subtask components.

Fields of Study & Subjects

Multimodal Machine Learning Applications Domain Adaptation and Few-Shot Learning Topic Modeling

Publication details

Journal / source
Conference of the European Chapter of the Association for Computational Linguistics
Pages
513–521
Publication date
1 January 2026
Language
EN