R²Orch trains agent orchestrators via self-play with a skill-internalization loop
CUHKSZ · hf · 2026-11-02
CUHKSZ researchers present R²Orch, which trains agent orchestrators via recursive self-play without hand-designed curricula. The outer loop performs skill withdrawal—retiring a skill once performance with it matches without it—and distills failures into new skills; the inner loop uses learnable task selection scored by recursive learnability and GRPO updates. Compared to fixed task pools and success-rate gates, it improves accuracy across reasoning and interactive tasks, with orchestration eventually holding at zero skills.
More from coding & agent
- CEPO Splits Coordination from Execution, Boosting Multi-Agent Prompt Optimization Across 11 Benchmarks — Jongha Kim · 2026-11-02
- Running Claude in an endless autonomous loop, it only shows what it thinks is cool — jaivin · 2026-11-02
- 10 real examples of people handing everyday chores to Grok Bot, from cutting bills to tracing family roots — minchoi · 2026-11-02
- Developer begs Anthropic to rein in flood of permission dialogs in Claude Code — signulll · 2026-11-02
- Can today's AI build this 12-year-old $5/hour search engine gig? — kingslyj · 2026-11-02
- Vibe-coded production services are a ticking bomb: engineer shares 4 fixes — arpit_bhayani · 2026-11-02