Research map / Multi-agent systems and games / LLM agents

Learning to Orchestrate Agents in Natural Language with the Conductor

ICLR 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Trains a 7B Conductor LLM with GRPO to write natural-language agentic workflows (subtasks, worker assignments, access lists) over a pool of frontier LLM workers, reaching SOTA on LiveCodeBench, GPQA-Diamond and other reasoning benchmarks, with extensions to random worker pools and recursive self-calls.

Problem

Existing LLM multi-agent and routing methods are costly and constrained to fixed, human-designed topologies or model selection, leaving the potential of coordinating diverse specialized LLMs underexploited.

Approach

Train a small LLM orchestrator end-to-end with RL (GRPO, outcome reward) to output free-form natural-language workflow steps that specify subtasks, assigned worker, and context access, so it learns prompt engineering, topology design, and task-adaptive coordination; extend with randomized worker pools and self-recursive calls for test-time scaling.

Key results

• Conductor (7B) achieves 83.93 on LiveCodeBench and 87.5 on GPQA-Diamond under unconstrained setting, beating the best prior reported results (e.g., GPT 5 at 82.90 and 82.3).
• In the controlled constrained setting, Conductor averages 72.35 across MATH500, MMLU, RLPR, LiveCodeBench vs. 62.13 for MoA and 64.52 for GPT 5 with 5x self-reflection.
• Recursive Conductor reaches 63.00 average on AIME25, BigCodeBench and GPQA-D vs. 61.93 for the non-recursive Conductor and 52.34 for Gemini Pro 2.5.

Surprising findings

• Few-shot examples drawn from out-of-distribution tasks (e.g., math for coding) gave better Conductor performance than in-distribution examples.
• Qwen3-32B without thinking outperformed its thinking mode on BigCodeBench, attributed to verbosity causing formatting failures.

Limitations

• Relies on expensive frontier closed-source workers, raising cost and access barriers (acknowledged in ethics statement).
• Baseline comparisons depend on re-implementations and best-of-own-or-leaderboard reporting, with unconstrained results mixing different setups.

Experimental setup

Datasets: MATH500 MMLU RLPR LiveCodeBench v1 LiveCodeBench v6 AIME25 GPQA-Diamond BigCodeBench

Benchmarks: MATH500 MMLU RLPR LiveCodeBench v6 AIME25 BigCodeBench GPQA-Diamond

Metrics: accuracy average accuracy across tasks token usage cost per sample

Baselines: Gemini 2.5 Pro Claude Sonnet 4 GPT-5 Qwen3-32B DeepSeek-R1-Distill-Qwen-32B gemma-3-27b-it MASRouter Mixture-of-Agents RouterDC Smoothie

Builds on: GRPO DeepSeek-R1 Mixture-of-Agents MasRouter RouterDC Smoothie

Closest papers by meaning

• Reward Modeling for Multi-Agent Orchestration (NeurIPS 2026 · similarity 0.89)
• Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.88)
• Offline Training of Language Model Agents with Functions as Learnable Weights (ICML 2024 · similarity 0.88)
• Self-Challenging Language Model Agents (NeurIPS 2025 · similarity 0.88)
• Improving the Efficiency of Language Agent Teams with Adaptive Task Graphs (NeurIPS 2026 · similarity 0.88)
• AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation (ICML 2026 · similarity 0.88)