Reproduce “Understanding Multimodal LLMs: the Mechanistic Interpretability of Llava in Visual Question Answering” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zeping Yu, Sophia Ananiadou
Researchers at The University of Manchester developed a mechanistic interpretability framework to dissect how Multimodal Large Language Models (MLLMs) like Llava process visual information in Visual Question Answering (VQA) tasks. The work reveals striking similarities between VQA and Textual Question Answering (TQA) mechanisms in deep layers and demonstrates that visual instruction tuning enhances existing textual reasoning abilities.
Key insights
- Deep-layer VQA mechanisms in MLLMs are fundamentally analogous to TQA mechanisms, indicating a transferability of core textual reasoning pathways to multimodal contexts.
- Visual instruction tuning primarily enhances and specializes the base textual LLM's pre-existing abilities for multimodal tasks rather than creating entirely new computational mechanisms.
- Initial visual embeddings, even at the projection stage, contain semantically interpretable information about objects and their attributes (e.g., animals and colors).
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.