We evolved an agent on one benchmark, froze it, and tested it on five it had never seen. Other self-improvement methods barely transfer, and some get worse. RRSI improved on every one, up to 24% over…

🔥 Key Highlights Across 8 Benchmarks / 3 Domains: 📈 Up to +14.1 pts improvement on the evolve split. 🌐 Robust out-of-distribution transfer, gaining up to +4.7 pts across unseen benchmarks. ⚡ Uses…

一个很强的 RSI 新信号可以关注一波： Google Research 开始研究 Agent Harness 的 Recursive Self-Improvement。 不重新训练模型，直接让 prompt、工具、memory、control flow 自动迭代。 这条路线我很看好。 因为相比之前 glm 发布的 model-level RSI 更清晰了 ： 现在很明显在 Harness…

Google's RRSI makes self-improving agents that transfer, not overfit RRSI regularizes the harness evolution - annealing edit budgets, leakage critics, noise floors, and pruning - so gains on the…