Reproduce “OptionWorld: A Benchmark for Diagnosing Failures of LLMs as Long-Term Personal Assistants” 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
Ruoyu Zhang, Zhaocen Liu, Weiyi Tian, Melissa Feng, Shilong Liu
OptionWorld evaluates assistants’ follow-through on hidden tasks across eight authored, 60-day simulations. It links evidence, judgments, actions, and feedback, counting success only when feedback confirms an option was preserved or an issue resolved. Tested systems’ option-preservation rates range from 23.6% to 61.6%. In five API systems, primary facts remained in the last valid input for 35.5% of failures, which does not establish attention or understanding.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.