Reproduce “CT-OPD: Counterfactual Trace On-Policy Distillation for Diffusion 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
Long Qian, Bingke Zhu, Jiaqi Wei, Yu Li, Yingying Chen, Jinqiao Wang
CT-OPD trains diffusion vision-language models by using student trajectories to select unresolved positions and a completed teacher answer to provide both visible context and target words. The student learns those targets with its own categorical prediction loss. On LLaDA-V, it scores 49.79 across nine benchmarks, versus 40.00 for Base; these results establish benchmark gains, not faster training.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.