AIDE² rewrote its own agent code for 8 days and kept seven improvements that transferred.

arXiv 2609.26457 proposes changes to its own code, benchmarks the modified agents on AI R&D tasks, and keeps only what wins on hidden evaluations. Seven successive improvements in an autonomous run, including a new search policy and context-compressing memory, and the gains generalized to held-out ML engineering, heuristic algorithm engineering and weather forecasting. Gating on hidden evals rather than self-reported scores is what separates this from the self-improvement claims that don't replicate.