Reproduce “VC-OPD: Visual Counterfactual On-Policy Distillation for Grounded Vision-Language Reasoning” 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
Jiahe Feng
VC-OPD weights tokens in student-generated responses by estimated visual dependence. A teacher scores the same response with original and visually weakened images; the score difference sets token weight, while the original-image teacher sets update direction. With a Qwen2.5-VL-3B student and Qwen2.5-VL-7B teacher, it reaches 48.87 average exact-match accuracy versus 46.64 for RKL across seven benchmarks, in @8 evaluations at temperature 1.0.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.