CoFlow: Few-Step Multi-Agent Coordination via Joint Velocity Field
A novel generative framework for offline multi-agent reinforcement learning enables coordinated few-step inference while maintaining inter-agent collaboration. The CoFlow architecture integrates Coordinated Velocity Attention (CVA) and Adaptive Coordination Gating, employing a finite-difference consistency surrogate to bypass the memory-intensive Jacobian-vector product backpropagation. Evaluated across 60 setups in MPE, MA-MuJoCo, and SMAC environments, CoFlow shows that single-pass multi-agent generation can retain coordination even when the velocity field is inherently joint-coupled, contradicting earlier beliefs that independent agent policies are necessary for few-step inference.
Key facts
- CoFlow is a generative model for offline multi-agent reinforcement learning.
- It enables coordinated few-step inference without sacrificing inter-agent coordination.
- The architecture uses Coordinated Velocity Attention (CVA) and Adaptive Coordination Gating.
- A finite-difference consistency surrogate replaces Jacobian-vector product backpropagation.
- Tested across 60 configurations in MPE, MA-MuJoCo, and SMAC environments.
- Prior approaches required many iterative sampling steps or independent agent policies.
- CoFlow shows single-pass multi-agent generation can preserve coordination with a joint-coupled velocity field.
- The paper is published on arXiv with ID 2605.01457.
Entities
Institutions
- arXiv