Reproduce “Towards Understanding Multimodal Fine-Tuning: Spatial Features” 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
Lachin Naghashyar, Hunar Batra, Ashkan Khakzar, Philip Torr, Ronald Clark, Christian Schroeder de Witt, Constantin Venhoff
The first mechanistic analysis of multimodal fine-tuning in Vision-Language Models from the University of Oxford details how a language backbone learns spatial reasoning capabilities. This work identified specialized, vision-grounded features primarily in mid-layers, causally linked them to a sparse set of attention heads, and demonstrated their functional importance through ablations that reduced spatial task accuracy by 9-16 percentage points.
Key insights
- Multimodal fine-tuning induces structured, interpretable changes within the language backbone, leading to the emergence of specialized, vision-grounded features for tasks like spatial reasoning.
- Adapted features, including those for spatial reasoning, are predominantly concentrated in the middle layers of the transformer, indicating localized integration and processing of visual information.
- The activation of these specialized features is causally driven by a small, sparse set of attention heads, often recurring across multiple related spatial concepts, suggesting coordinated processing mechanisms.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.