Uno-Orchestra: AI System Optimizes Multi-Agent Task Delegation
A new orchestration policy called Uno-Orchestra has been developed by researchers for multi-agent systems utilizing large language models (LLMs). This innovative approach selectively breaks down tasks and assigns subtasks to suitable (model, primitive) pairs. In contrast to conventional rigid orchestration techniques, Uno-Orchestra employs reinforcement learning based on curated trajectories that reflect actual worker interactions, optimizing decomposition depth, worker selection, and inference budget simultaneously. When evaluated against 22 baseline methods across 13 benchmarks—including math, code, knowledge, long-context, and agentic tool-use—Uno-Orchestra achieved a macro pass@1 score of 77.0%, surpassing the best workflow baseline by approximately 16% while maintaining a significantly lower per-query cost. This represents a significant advancement in the accuracy-efficiency balance of selective delegation in multi-agent systems.
Key facts
- Uno-Orchestra is a unified orchestration policy for LLM multi-agent systems.
- It selectively decomposes tasks and dispatches subtasks to admissible (model, primitive) pairs.
- Decisions are learned via RL from curated trajectories grounded in real worker interactions.
- Tested against 22 baselines on a 13-benchmark suite.
- Achieved 77.0% macro pass@1, roughly 16% above the strongest baseline.
- Per-query cost is roughly an order of magnitude lower.
- Benchmarks include math, code, knowledge, long-context, and agentic tool-use.
- Advances the accuracy-efficiency frontier of selective delegation.
Entities
Institutions
- arXiv