Reproduce “SheetMind: Actions Set Accuracy, Agents Set the Failure Mode” 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
Lyuhao Chen, Xi Cheng, Yanming Kang, Ruiyan Zhu, Ke Liu, Rakesh Chowdary Machineni, Yulang Fei, Brian Zhu and 6 more
SheetMind evaluates whether spreadsheet-agent performance depends on its action interface or added agents. Its Manager decomposes instructions, an Action Agent uses a high-level API of about 30 operations, and a Reflection Agent checks each saved action. On 221 SCB-221 tasks, GPT-5-mini scores 55.2% with this API versus 8.1% with eight primitive operations; the comparison concerns API packages, not granularity alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.