DPBench: Structural Determinants of Multi-Agent LLM Coordination (arxiv.org)

Researchers have introduced DPBench, a groundbreaking benchmark designed to assess coordination among multi-agent systems utilizing large language models (LLMs). Unlike existing benchmarks that focus solely on task-level success, DPBench emphasizes the structural conditions affecting coordination outcomes. It reimagines the classic Dining Philosophers problem into a versatile testing environment, where variables such as action protocol, communication structure, and group size can be independently varied. The study evaluated six LLMs, revealing significant differences in coordination success; for instance, GPT-5.2 experienced a deadlock rate of 25%, while the Gemini 2.5 Flash peaked at 90% under default conditions. This development is particularly significant for the AI/ML community as it provides a structured framework to understand and improve agent coordination, crucial for effectively deploying multi-agent systems in real-world scenarios. The findings indicate that coordination success is heavily influenced by protocol design rather than the capabilities of individual models. For example, implementing pre-commitment communication and specific prompting strategies can dramatically reduce deadlock rates. DPBench not only paves the way for future research on LLM interaction but also serves as a vital tool for optimizing multi-agent performance in diverse applications.