Reproduce “Distilling Visual Reasoning into Text Space” 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
Wenhan Yang, Nilay Naharas, Ali Payani, Baharan Mirzasoleiman
V2T distills visual reasoning from an interleaved-image teacher into a student that answers using only the original image and text. During training, the student matches ground-truth reasoning and the teacher’s next-token distributions, with optional attention distillation and GRPO. It reaches 75.5% on an 800-example held-out set, versus 61.2% for the teacher, but results are benchmark-specific.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.