Reproduce “From Perception to Integration: Revisiting the Internal Dynamics of Reasoning in Vision-Language Models” 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
Rong Yu Xu, Prayag Tiwari, Shaolei Zhang
The authors study when vision-language models can use answers to composite visual questions, rather than merely represent components. They test primitive judgments, matched scenes, readouts, interventions, and truncated traces, then train a readiness detector to stop generation. Across MMStar and RealWorldQA, accuracy rises 3.13 and 3.30 points while reasoning tokens fall 79.1% and 74.5%, though accuracy falls for one model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.