Reproduce “VLMProbe: An Open-source Library for VLM Probing and Interpretation” 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
Gexin Huang, Haisheng Fu, Yanting Yang, Chen Zhao, Ruinan Jin, Jiaqi Liu, Xiaoxiao Li
The library combines representation readouts, attention explanations, causal interventions, and sparse autoencoder analysis for vision-language models, using model-specific adapters to align inputs and internal sites. Matched comparisons show these methods answer different questions. On a GQA shirt-color case, replacing LLaVA-1.5-7B’s visual-token states reduced Yellow’s answer log probability by 1.646, while a local estimate predicted an increase; this is a case-specific result.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.