Enregistré dans:
Détails bibliographiques
Auteurs principaux: Wang, Qidong, Hu, Junjie, Jiang, Ming
Format: Preprint
Publié: 2025
Sujets:
Accès en ligne:https://arxiv.org/abs/2509.14837
Tags: Ajouter un tag
Pas de tags, Soyez le premier à ajouter un tag!
_version_ 1866910168016486400
author Wang, Qidong
Hu, Junjie
Jiang, Ming
author_facet Wang, Qidong
Hu, Junjie
Jiang, Ming
contents Recent advances in causal interpretability have extended from language models to vision-language models (VLMs), seeking to reveal their internal mechanisms through input interventions. While textual interventions often target semantics, visual interventions typically rely on coarse pixel-level perturbations, limiting semantic insights on multimodal integration. In this study, we introduce V-SEAM, a novel framework that combines Visual Semantic Editing and Attention Modulating for causal interpretation of VLMs. V-SEAM enables concept-level visual manipulations and identifies attention heads with positive or negative contributions to predictions across three semantic levels: objects, attributes, and relationships. We observe that positive heads are often shared within the same semantic level but vary across levels, while negative heads tend to generalize broadly. Finally, we introduce an automatic method to modulate key head embeddings, demonstrating enhanced performance for both LLaVA and InstructBLIP across three diverse VQA benchmarks. Our data and code are released at: https://github.com/petergit1/V-SEAM.
format Preprint
id arxiv_https___arxiv_org_abs_2509_14837
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models
Wang, Qidong
Hu, Junjie
Jiang, Ming
Computation and Language
Recent advances in causal interpretability have extended from language models to vision-language models (VLMs), seeking to reveal their internal mechanisms through input interventions. While textual interventions often target semantics, visual interventions typically rely on coarse pixel-level perturbations, limiting semantic insights on multimodal integration. In this study, we introduce V-SEAM, a novel framework that combines Visual Semantic Editing and Attention Modulating for causal interpretation of VLMs. V-SEAM enables concept-level visual manipulations and identifies attention heads with positive or negative contributions to predictions across three semantic levels: objects, attributes, and relationships. We observe that positive heads are often shared within the same semantic level but vary across levels, while negative heads tend to generalize broadly. Finally, we introduce an automatic method to modulate key head embeddings, demonstrating enhanced performance for both LLaVA and InstructBLIP across three diverse VQA benchmarks. Our data and code are released at: https://github.com/petergit1/V-SEAM.
title V-SEAM: Visual Semantic Editing and Attention Modulating for Causal Interpretability of Vision-Language Models
topic Computation and Language
url https://arxiv.org/abs/2509.14837