Reproduce “SlowLab: Benchmarking Autonomous AI Agents on a Year of Experiments They Cannot Undo” 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

Ningpei Ding, Zai Wen Feng

SlowLab tests language-model agents in a one-year greenhouse simulation, where they choose planting times and climate policies across four compartments. Crops occupy space for 180 days, while agents receive partial process measurements. Across 48 unseen sites, DeepSeek V4.1 Flash’s Full-feedback policy scored 5.60 EUR/m² below a domain-informed Bayesian-optimisation baseline. This simulation comparison does not establish greenhouse performance.

Summary by alphaXiv.

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