DPBench: Structural Determinants of Multi-Agent LLM Coordination Under Simultaneous Resource Contention

DPBench introduces a benchmark for testing multi-agent LLM coordination using the Dining Philosophers problem, revealing that deadlock rates vary dramatically (25%-90%) across models under identical conditions. The research demonstrates that coordination success is primarily determined by protocol design—including communication structure and concurrency primitives—rather than model capability alone.

DPBench addresses a critical gap in AI evaluation: while existing benchmarks measure task success under fixed conditions, they ignore the structural factors determining whether coordination succeeds or fails entirely. By adapting the Dining Philosophers problem into a controlled testbed, researchers created a framework where action protocols, communication topology, and group size vary independently, enabling systematic analysis of multi-agent LLM behavior under resource contention.

The findings challenge conventional assumptions about model capability. GPT-5.2 achieves 25% deadlock while Gemini 2.5 Flash reaches 90% under identical conditions—suggesting raw model performance correlates weakly with coordination success. More provocatively, three protocol interventions drive Gemini's deadlock from 90% toward zero: adding three rounds of pre-commitment communication, encoding classical concurrency primitives in prompts, or scaling group size from five to ten agents. This indicates protocol design dominates model selection as a coordination lever.

These insights matter for deploying multi-agent LLM systems in production environments where resource contention mirrors real-world constraints. Applications spanning autonomous trading, distributed scheduling, and collaborative reasoning depend on reliable coordination. The research suggests that careful protocol engineering—rather than waiting for superior model architectures—can systematically improve reliability. Single-round messaging and memory mechanisms showed negligible impact at tested scales, hinting at phase transitions in coordination complexity.

Looking ahead, this work opens questions about scaling laws in multi-agent systems and whether coordination principles transfer across domains beyond philosophers and resources. Future research should explore how these findings apply to heterogeneous agent populations and whether emerging reasoning models fundamentally alter protocol requirements.

• →Deadlock rates in multi-agent LLM coordination vary 25%-90% across models under identical conditions, challenging the assumption that capability alone determines success.
• →Protocol design—including communication structure and concurrency primitives—emerges as the dominant factor controlling coordination outcomes, often surpassing model selection.
• →Three specific interventions (pre-commitment rounds, symmetry-breaking prompts, scaling group size) can reduce deadlock from 90% to near-zero in tested scenarios.
• →Current benchmarks fail to characterize the structural conditions enabling or preventing coordination, limiting insights into multi-agent LLM reliability.
• →Single-round messaging and memory mechanisms showed minimal impact on deadlock rates, suggesting protocol complexity requirements may increase non-linearly with agent count.