Reproduce “ScopeMem: Latent Memory That Learns Where to Revise Frozen GUI Agents” 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
Zhicheng Zheng, Cheng Zhang, Shengping Yu, Tianxiang Zhao
ScopeMem trains a small encoder to condition a prefix on a frozen GUI agent’s first answer, then prompts the same agent to answer again. It uses a preservation objective to target revisions while retaining correct decisions. On a held-out set of 1,463 steps, it repaired 254 errors and harmed 49 correct decisions; results are offline and step-level, not task success.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.