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

Recognize Your Orchestrator: An Entropy Dynamics Perspective for LLM Multi-Agent Systems

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Robust and efficient agent execution · 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.

Proposes a mean-field entropy dynamics model (damped oscillation plus log context growth) of LLM orchestrator scheduling entropy, plus an inverse workflow generation pipeline for synthesizing process-verifiable multi-agent benchmarks, and reports a "Reasoning Trap" where reasoning-heavy models underperform as orchestrators.

Problem

Centralized orchestrator agents in LLM multi-agent systems are fragile, yet there is no formal model or step-level benchmark explaining how orchestrator scheduling uncertainty evolves over long-horizon tasks.

Approach

Models the orchestrator's per-step scheduling entropy as the sum of a damped sinusoidal task-resolution term (momentum-like hunting) and a logarithmic context-accumulation term, fits these parameters to step-level logs from a synthesized benchmark built by inverse planning with deterministic checkpoints.

Key results

• Claude-4.5-Sonnet achieves the highest Step-SR (60.83%) and orchestrator-level average (77.18) among evaluated models, while GPT-5 scores only 22.49 system-level average.
• For GPT-5, switching from High to Low reasoning reduces token usage by about 69% and raises Step-SR from 0.251 to 0.302; Qwen3-Max NoThink reaches Step-SR 0.413.
• The entropy model fits empirical step-wise entropy trajectories with reported p < 0.005 across six orchestrators, though fitted parameters are reported without a quantitative fit-quality metric.

Surprising findings

• Reasoning-heavy models (e.g., DeepSeek-R1, GPT-5 High) perform worse as orchestrators than light-thinking variants, which the authors term the Reasoning Trap.
• The paper reports the Orchestrator accounts for about 67.7% of failures across four MAS systems, while executors account for only 32.3%.

Limitations

• Fitted parameters lack identifiability and rigorous justification; the mean-field derivation relies on heuristic assumptions (e.g., diffusion-to-log-entropy and momentum-to-damped-oscillator mappings).
• Executors are built on GPT-4/GPT-4o and the benchmark is synthesized with GPT-4o and Qwen-based validators, creating possible circularity and model-bias concerns.
• Fitting uses only the first six steps, and the paper's own chaotic-regime claims and long-horizon test are limited.

Experimental setup

Datasets: GAIA AssistantBench HotpotQA HumanEval BrowserComp OCR-VQA DeepResearchBench

Benchmarks: IWG benchmark (679 instances)

Metrics: LCS-F1 Task Success Step-SR EH-F1 Faithfulness Consistency Token Efficiency

Baselines: GPT-5 Gemini-2.5-Pro Claude-4.5-Sonnet DeepSeek-V3.1 Qwen3-Max GPT-4.1 Grok-4

Builds on: ReAct Magentic-One Nesterov accelerated gradient Su-Boyd-Candès ODE

Closest papers by meaning

• MAS-Orchestra: Understanding and Improving Multi-Agent Reasoning Through Holistic Orchestration and Controlled Benchmarks (ICML 2026 · similarity 0.88)
• Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.88)
• RAGEN-2: Reasoning Collapse in Agentic RL (ICML 2026 · similarity 0.88)
• Learning Latency-Aware Orchestration for Multi-Agent Systems (NeurIPS 2026 · similarity 0.88)
• Reward Modeling for Multi-Agent Orchestration (NeurIPS 2026 · similarity 0.87)
• Emergent Coordination in Multi-Agent Language Models (ICLR 2026 · similarity 0.87)