Recursive Self-Evolving Agents via Held-Out Selection

Researchers introduce RSEA, a Recursive Self-Evolving Agent that improves LLM agents by rewriting its strategy, skills, and playbook from its own trajectories. RSEA commits changes only if they do not regress on a held-out split, ensuring monotone-safe self-evolution. Across four benchmarks, RSEA demonstrates strong performance, particularly on ALFWorld, and highlights the importance of guarded context evolution.