Title: Identifying the Orchestrator: An Entropy Dynamics Approach to LLM Multi-Agent Systems

Abstract: While the shift from single-turn models to Multi-Agent Systems (MAS) holds the potential for superior problem-solving, the current centralized orchestration architecture remains a significant vulnerability. To investigate this issue, we present a Mean-Field Entropy Dynamics framework, which characterizes the orchestration process through the tension between task resolution and the burden of accumulating context. For validation purposes, we introduce Inverse Workflow Generation (IWG), a multi-agent pipeline designed to create high-complexity benchmarks featuring dense intermediate checkpoints that allow for process verification. Our results show that the entropy dynamics model aligns with empirical trajectories, yielding physically interpretable parameters that measure both system stability and the onset of performance collapse. Notably, our analysis reveals a "Reasoning Trap": although reasoning-intensive models perform well in isolated scenarios, they often struggle as orchestrators because of context squeezing. By clarifying the physical mechanisms behind the Orchestrator’s role and quantifying systemic uncertainty, this work provides valuable guidance for the architectural design of MAS.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC