Reproduce “D&A: Decoupling Description and Assessment Enables VLMs as Zero-Shot Robotic Reward 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
Marius Memmel, Shawn Yin, Urvi Rutia, Angad Wadhwa, Eldon Li, Gavin Angeloff, Ankit Goyal, Jesse Zhang and 4 more
D&A is a prompt-based robot reward method that describes video events before judging them against an explicit task-success criterion. Its vision-language model separates description from assessment, aiming to reduce rewards based on assumed success. On D&A-Bench, Gemini 3 Flash achieves Kendall τₐ of 0.672, versus 0.438 for direct prompting; this shows improved trajectory ranking on that benchmark, not universal performance.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.