Reproduce “Residual Visual Credit Optimization: Conserved Evidence Routing for Multimodal Reinforcement Learning” 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
Lin Qiu, Yao Liu, Diyi Hu, Hanqing Zeng, Onur Gungor, Chujie Chen, Jiayi Liu, Jianyu Wang and 1 more
RVCO routes sequence-level reward across visual reasoning tokens by comparing sampled-token likelihoods under original and patch-suppressed images, then allocating a conserved credit budget with positive support at every valid position. On Qwen3-VL-8B, it reaches a seven-benchmark macro-average accuracy@8 of 66.8%, versus 65.7% for VPPO, under matched settings; this result is limited to that backbone and evaluation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.