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

AgentConductor: Topology Evolution for Multi-Agent Competition-Level Code Generation

ICML 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 Qwen2.5-3B orchestrator with SFT then GRPO to generate difficulty-aware, density-controlled layered DAG agent topologies in YAML, refined over turns using execution feedback, for competition-level code generation.

Problem

Existing LLM multi-agent systems use fixed or query-level static interaction topologies that do not scale density with task difficulty or refine topology using execution feedback within an instance, causing redundant communication and cost on easy problems.

Approach

An RL-trained LLM orchestrator emits a YAML layered-DAG topology per query, conditioned on inferred difficulty, with a graph density reward (node, edge, depth terms) bounded by difficulty-specific node caps, and regenerates the topology over turns from test feedback.

Key results

• AgentConductor (3B) reaches 58.8% pass@1 on APPS, 14.6 points above the strongest baseline (FlowReasoner/MetaGPT-level second best at 51.3 on APPS under MetaGPT per table; paper claims 14.6 margin).
• Average over five datasets is 72.1 versus 68.7 for the next best (MetaGPT), with 97.5% on HumanEval and 95.1% on MBPP.
• On APPS, prompt tokens drop to 277,600 versus 320,550 to 531,450 for listed baselines (paper reports up to 68% token cost reduction).

Surprising findings

• Removing the SFT stage yields almost no valid topologies (15% on APPS), so RL alone from a small backbone fails to produce usable topologies.
• Topology optimization baselines underperform fixed multi-agent frameworks like MetaGPT on average, which the authors attribute to rigid learned topologies.

Limitations

• Individual agents rely on gpt-4o-mini API calls, so cost and results depend on a closed API model.
• Maximum of two interaction turns limits iterative refinement.
• Difficulty node caps (4, 7, 10) were set from statistics of SFT-generated samples, a heuristic choice.

Experimental setup

Datasets: APPS CodeContests LiveCodeBench HumanEval MBPP

Benchmarks: APPS LiveCodeBench (v4) CodeContests HumanEval MBPP GAIA HLE PopQA

Metrics: pass@1 prompt tokens completion tokens average topology density (S_complex) valid topology rate

Baselines: GPT-4o-mini AutoGen MetaGPT MapCoder AFlow FlowReasoner Chain-of-Agents GPTSwarm AgentPrune G-Designer MacNet

Builds on: GRPO Search-R1 AgentPrune G-Designer MacNet FlowReasoner Chain-of-Agents

Closest papers by meaning

• Assemble Your Crew: Automatic Multi-agent Communication Topology Design via Autoregressive Graph Generation (AAAI 2026 · similarity 0.90)
• CARD: Towards Conditional Design of Multi-agent Topological Structures (ICLR 2026 · similarity 0.90)
• When Parallelism Pays Off: Cohesion-Aware Task Partitioning for Multi-Agent Coding (NeurIPS 2026 · similarity 0.89)
• HieraMAS: Optimizing Intra-Node LLM Mixtures and Inter-Node Topology for Multi-Agent Systems (ICML 2026 · similarity 0.89)
• MAS$^2$: Self-Generative, Self-Configuring, Self-Rectifying Multi-Agent Systems (ICLR 2026 · similarity 0.88)
• G-Designer: Architecting Multi-agent Communication Topologies via Graph Neural Networks (ICML 2025 · similarity 0.88)