Reproduce “BaseRSI: Recursive Self-Improvement of a Base Vision-Language Model via Training-Free Harness Internalization” 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
Mingda Zhang, Yuanhao Dong, Qiang Huang, Rui Qian, Yanjin Li, Zhenyu Liu, Xiao Huang, Zihao Yang and 3 more
BaseRSI turns request-level visual-reasoning help into edits to a vision-language model, allowing it to answer original requests without that help. It measures block-level differences between harness-assisted and unedited responses, then fits low-rank edits using forward passes and linear solves. On five ID and five held-out OOD benchmarks, a 27B model gains 11.7 and 6.1 percentage points, respectively; results cover one model and ten tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.