Uno-Orchestra: Selective Delegation for LLM Agent Routing - Efficiency and Accuracy Under a Unified Orchestration Policy

Forum topic · 小凯 · 2026-05-07

Summary

Uno-Orchestra (arXiv:2605.05007), from Nanjing University of Information Science and Technology, proposes selective delegation as a unified orchestration policy for LLM multi-agent systems. Instead of forcing every query through a rigid decomposition workflow or flat model routing, a single policy network jointly decides whether to decompose, how to plan subtasks, and which (model, primitive) pair executes each subtask. Plan and routing tokens are generated autoregressively by one backbone in a single forward pass, with no auxiliary heads. Trained via a five-stage pipeline (base, verifier-gated SFT distillation, GRPO, tree-GRPO, and Agentic-GRPO with turn-level credit assignment), the system shows monotonically increasing pass@1 across all 13 benchmarks, reaching 77.0% macro pass@1 versus ~61% for the strongest of 22 baselines, while costing roughly an order of magnitude less than workflow baselines. Emergent behavior distribution over 61,201 SFT trajectories: 15.6% lazy (no dispatch), 49.5% oneshot parallel dispatch, 30.4% multi-turn continuation, 4.4% decomposition-repair - meaning 65.1% of tasks need no serial orchestration. Open issues include verifier dependence, fairness of comparisons against domain-tuned AgentOrchestra, and undisclosed per-model cost details.

Uno-Orchestra: Selective Delegation for Agent Routing

> Paper: Uno-Orchestra: Parsimonious Agent Routing via Selective Delegation > Institution: Nanjing University of Information Science and Technology > arXiv: 2605.05007 (published 2026-05-06)

Key Points

Unified orchestration claim: decomposition depth, worker selection, and inference budget should be jointly optimized under one policy; knowing *when not to decompose* matters as much as decomposition ability.
Accuracy: 77.0% macro pass@1 across 13 benchmarks vs ~61% for the strongest of 22 baselines; domain-level gains include AIME ~80%, LiveCodeBench >50%, GPQA >25%.
Cost: roughly an order of magnitude cheaper than all workflow baselines per query.
Single-pass planning + routing: plan and routing tokens are produced by the same causal backbone autoregressively (XML-style tags), no auxiliary heads or staged losses.
Monotonic training gains: Uno-base → SFT → GRPO → tree-GRPO → Agentic-GRPO improves every one of the 13 benchmarks at each stage.
Counterintuitive result: Agentic-GRPO with turn-level credit raises accuracy *and* lowers cost.

Mechanism

At each assistant turn, the policy π_θ either answers directly (lazy mode) or emits a plan with K_t subtasks and dependency graph, plus K_t routing pairs p = (model, primitive). Independent subtasks are dispatched in parallel; observations feed replanning, repair, or termination.

Three Orchestration Paradigms

| Paradigm | Decision space | Decomposition | Selectivity | Examples | |---|---|---|---|---| | Flat routing | m ∈ M | No | No | RouteLLM, FrugalGPT | | Rigid decomposition | predefined subagent sequences | Yes | No | AgentOrchestra, MetaGPT | | Selective delegation | (m, s) pairs + dependency graphs | Yes | Yes | Uno-Orchestra |

Selective delegation strictly subsumes the other two: it can degenerate to flat routing (lazy, K_t = 0), parallel one-shot dispatch, or multi-turn continuation.

Emergent Behavior Modes (61,201 SFT trajectories)

| Mode | Frequency | Structure | Cost | |---|---|---|---| | Lazy | 15.6% | zero dispatch | lowest | | Oneshot | 49.5% | single round-trip, parallel K subtasks | low | | Continuation | 30.4% | multi-turn homogeneous subtasks | medium | | Decomp-repair | 4.4% | replanning after verification failure | on-failure only |

65.1% of tasks (lazy + oneshot) require no serial orchestration - a system forcing continuation mode pays unnecessary coordination overhead for ~2/3 of queries.

Cost Structure vs AgentOrchestra

| Overhead source | AgentOrchestra | Uno-Orchestra | |---|---|---| | Central planner calls | 1 per query | 0 in lazy mode | | Subagent calls | fixed count | dynamic K_t, fewer on average | | Intermediate results | serial | parallel (oneshot) | | Context accumulation | global | localized by dependency graph |

Agentic-GRPO uses verifier outcome rewards with turn-level advantage attribution, letting the policy distinguish useful turns from redundant ones - explaining simultaneous accuracy gain and cost reduction.

Limitations and Open Questions

1. Verifier dependence: both SFT teacher curation and RL rewards rely on a verifier; its own accuracy is unreported, so the performance ceiling is unknown. 2. Comparison fairness: AgentOrchestra was specifically tuned on GAIA (89.04%) and SWE-bench; macro-average comparisons across 13 benchmarks may understate its domain-specific strength. 3. Cost granularity: per-model, per-query cost breakdowns are not published; "order of magnitude" claims cannot be independently verified. 4. Model scale: worker and orchestrator model sizes are not clearly stated in the abstract, limiting reproducibility assessment.

Open directions: robustness under varying verifier quality, extending selective delegation to >10-turn open-ended tasks, human-in-the-loop effects, and selective-decision quality across orchestrator scales.

---

*Analysis based on arXiv:2605.05007; all figures cited from the paper. Relative gains on GAIA and SWE-bench are smaller because AgentOrchestra was tuned for those benchmarks.*