Do Vision-Language Models Reason Like Humans? Exploring The Functional Roles of Attention Heads

Title: Do Vision-Language Models Reason Like Humans? Exploring The Functional Roles of Attention Heads
Speaker: Yanbei Jiang
Time: 11 am, Tuesday, 28th October, 2025
Location: !290-4-4206-Edinburgh Room
Zoom: https://unimelb.zoom.us/j/9969138032?pwd=NCtxT0Z1OWp3RUREVmxxV1hBMkxDZz09
or meetingId/pwd: uomnlp/uomnlp

Abstract:
Despite excelling on multimodal benchmarks, vision–language models (VLMs) largely remain a black box. In this paper, we propose a novel interpretability framework to systematically analyze the internal mechanisms of VLMs, focusing on the functional roles of attention heads in multimodal reasoning. To this end, we introduce CogVision, a dataset that decomposes complex multimodal questions into step-by-step subquestions designed to simulate human reasoning through a chain-of-thought paradigm, with each subquestion associated with specific receptive or cognitive functions such as high-level visual reception and inference. Using a probing-based methodology, we identify attention heads that specialize in these functions and characterize them as functional heads. Our analysis across diverse VLM families reveals that these functional heads are universally sparse, vary in number and distribution across functions, and mediate interactions and hierarchical organization. Furthermore, intervention experiments demonstrate their critical role in multimodal reasoning: removing functional heads leads to performance degradation, while emphasizing them enhances accuracy. These findings provide new insights into the cognitive organization of VLMs and suggest promising directions for designing models with more human-aligned perceptual and reasoning abilities.

Bio:
I’m a second-year PhD student in the NLP group, advised by Jey Han and Kris Ehinger. My research focuses on enhancing step-by-step visual reasoning in multimodal Vision-Language Models (VLMs). This work is about the architectural introspection and interpretability of VLMs to better understand and improve their performance. I’m also exploring reinforcement learning methods with process reward models.