RECLAIM is a of 100 NeurIPS 2025 papers designed to test whether AI agents can reproduce published results. Each paper has a fixed target result, a success criterion, and a GPU-hour budget. Papers are classified into three difficulty tiers based on what authors released: Run (code, data, weights), Retrain (no weights), and Reimplement (no code). A separate language model grades agents using logs and outputs, not the agents' own reports.
Four agents were each run once per paper. The best agent reproduced 41% of Run-tier papers, 27% of Retrain-tier papers, and 15% of Reimplement-tier papers. Failed attempts used only 29% of their budget on average, meaning most stopped early. The most common error was writing or running code without ever checking the intermediate or final outputs against the paper's reported numbers, occurring in 63 of 400 runs.
These results are interesting because they show that state-of-the-art agents still struggle with end-to-end reproduction, even when all required artifacts are present. The primary failure mode is a lack of , not insufficient compute. This highlights a gap in how agents reason about and validate their own work, and suggests directions for improving agent design.