Head Pursuit: Probing Attention Specialization in Multimodal Transformers

TL;DR

This paper investigates how individual attention heads in multimodal transformers specialize in specific semantic or visual attributes, providing tools for understanding and editing model behavior.

Contribution

It introduces a signal processing-based interpretability method to identify and manipulate attention heads responsible for specific concepts in multimodal models.

Findings

Attention heads show consistent specialization patterns across tasks.

Editing 1% of heads can control targeted concepts in outputs.

The approach applies to language and vision-language tasks.

Abstract

Language and vision-language models have shown impressive performance across a wide range of tasks, but their internal mechanisms remain only partly understood. In this work, we study how individual attention heads in text-generative models specialize in specific semantic or visual attributes. Building on an established interpretability method, we reinterpret the practice of probing intermediate activations with the final decoding layer through the lens of signal processing. This lets us analyze multiple samples in a principled way and rank attention heads based on their relevance to target concepts. Our results show consistent patterns of specialization at the head level across both unimodal and multimodal transformers. Remarkably, we find that editing as few as 1% of the heads, selected using our method, can reliably suppress or enhance targeted concepts in the model output. We validate our approach on language tasks such as question answering and toxicity mitigation, as well as vision-language tasks including image classification and captioning. Our findings highlight an interpretable and controllable structure within attention layers, offering simple tools for understanding and editing large-scale generative models.