Circuit Tracing in Vision-Language Models: Understanding the Internal Mechanisms of Multimodal Thinking

TL;DR

This paper introduces a novel framework for transparent circuit tracing in vision-language models, revealing how they hierarchically integrate visual and semantic information, and demonstrating the causal and controllable nature of specific circuits.

Contribution

It presents the first systematic approach for analyzing internal circuits in VLMs, enabling understanding and control of multimodal reasoning processes.

Findings

Visual feature circuits handle mathematical reasoning

Circuits support cross-modal associations

Circuits are shown to be causal and controllable

Abstract

Vision-language models (VLMs) are powerful but remain opaque black boxes. We introduce the first framework for transparent circuit tracing in VLMs to systematically analyze multimodal reasoning. By utilizing transcoders, attribution graphs, and attention-based methods, we uncover how VLMs hierarchically integrate visual and semantic concepts. We reveal that distinct visual feature circuits can handle mathematical reasoning and support cross-modal associations. Validated through feature steering and circuit patching, our framework proves these circuits are causal and controllable, laying the groundwork for more explainable and reliable VLMs.