Recursive Self-Evolving Agents via Held-Out Selection

Short summary

Researchers introduce RSEA (Recursive Self-Evolving Agent), a framework that iteratively improves agent behavior by rewriting natural-language components—strategies, skills, and playbooks—without updating model weights. Across four benchmarks, RSEA achieves strong results but reveals no universal agent method works best. Key finding: strict held-out selection prevents performance degradation and ensures safe, monotonic improvements.

• •RSEA iteratively rewrites agent prompts, strategies, and playbooks without model updates
• •Strict held-out validation prevents unsafe context evolution and performance degradation
• •No single agent method dominates—architecture choice depends on task characteristics

Generated with AI, which can make mistakes.