CoFlow addresses three linked obstacles in offline MARL generation:
Challenge 1: Step Bottleneck
Diffusion and flow baselines often rely on repeated sampling steps, and each step repeats cross-agent attention.
CoFlow: a joint averaged velocity field supports single-pass shortcut sampling.
Challenge 2: Coordination Loss
Per-agent distillation and independent one-step velocity fields remove the communication path needed for cooperation.
CoFlow: CVA inserts cross-agent attention directly into the velocity field.
Challenge 3: Memory Cost
JVP-based consistency regularization is expensive when a model must represent multiple agents jointly.
CoFlow: two stop-gradient forward passes form a finite-difference surrogate.
A weight-shared temporal U-Net embeds Coordinated Velocity Attention at every skip connection. Adaptive gates start at zero, so training begins from an independent-agent baseline and activates coordination only when gradients support it.
Contributions
- Joint-coupled velocity field: CVA with adaptive coordination gates keeps inter-agent communication inside each few-step forward pass.
- Memory-efficient consistency: a finite-difference surrogate replaces second-order JVP backpropagation with ordinary forward passes.
- Large-scale validation: 60 configurations across MPE, MA-MuJoCo, and SMAC evaluate returns, coordination probes, and denoising-step budgets.