R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play
Original title: R^2Orch: Recursive Self-Improvement for Agent Orchestration via Self-Play within a Skill-Internalization Loop
R^2Orch adds inner self-play and outer loops using Recursive Learnability and Skill Withdrawal to enhance agent orchestration.
Agentic orchestration faces challenges like hand-designed curricula, advantage collapse, and unreliable evaluation. R^2Orch introduces recursive self-play with an inner loop improving the orchestrator on self-chosen tasks and an outer loop refining that process. It features Recursive Learnability to reward task selection based on diverse outcomes and Skill Withdrawal for removing mastered skills. Experiments show consistent outperformance on twelve datasets for question answering, math, code, and interactive tasks.
- Outperforms baselines across twelve datasets
- Introduces Recursive Learnability training objective
- Skill Withdrawal removes each skill after controlled evaluation
- Code available at https://github.com/beitao/R2Orch
From the original article
1 min read at huggingface.co
In recent years, agentic orchestration has been widely applied to solve complex human tasks. However, existing orchestration learning still faces key challenges, including hand-designed curricula, advantage collapse when a task's rollouts all reach the same outcome, and unreliable evaluation signals that judge each skill by one noisy score.
By Mingda Zhang, Yun Huang, Qiang Huang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.