Reproduce “ForecastGym: A Live, Controlled, and Replayable Benchmark for Forecasting 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

Yicheng Tao, Yuante Li, Zipei Fan

ForecastGym records search evidence prospectively for open forecasting questions and replays the same evidence across models and memory settings. It compares four search providers while holding Gemini 3.8 Flash fixed, finding last-state Brier scores from 0.212 with Parallel to 0.255 with Serper. This one-batch comparison includes provider-specific follow-ups and formatting, so it does not isolate search ranking alone.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.