Research map / Label-efficient and robust learning / Knowledge distillation

Multi-Agent Coordination via Support-Preserving Distillation

NeurIPS 2026 · Label-efficient and robust learning › Knowledge distillation · approach: Teacher properties and robustness

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MoSDOT, which assigns noise to capacity-matched replay modes via conditional semi-discrete optimal transport before flow-teacher training and distills into decentralized one-step actors for offline MARL, plus a shared-randomness variant to close the strict-product gap.

Problem

Generative teachers in offline MARL route nearby noise samples to conflicting joint-action modes, producing off-support mass that decentralized one-step actors inherit through conditional-mean distillation; independent local actors also cannot represent correlated joint support.

Approach

Summarize multimodal replay into a finite mode support with replay-frequency capacities and use semi-discrete OT (Laguerre cells) to assign source noise to single joint modes before training a flow teacher, then distill into strict-product actors; a shared public signal variant lets agents coordinate mode choice without communication.

Key results

• On the landmark teacher diagnostic, MoSDOT achieves route consistency 0.972 vs. 0.856 for flow matching, 0.857 for IMLE, and 0.653 for Drifting.
• On SMACv1, MoSDOT reaches average reward 16.7 vs. 15.6 for MAC-Flow and DoF, the best average among listed methods.
• On SMACv2, MoSDOT averages 11.9, below DoF (14.0) and MAC-Flow (13.1), so it does not win on SMACv2 overall.

Surprising findings

• On SMACv2 terran_10_vs_10 and the zerg/terran averages, MoSDOT underperforms DoF and MAC-Flow, which the authors attribute to identity aggregation approaching sample-level coupling.
• Drifting teacher has high endpoint success but the lowest route consistency, showing endpoint accuracy does not imply distillable assignments.

Limitations

• Strict-product execution cannot represent arbitrary correlated joint support; shared-randomness variant needs a public coordination channel.
• Empirical scope limited to at most 10 agents and low-dimensional or discrete action spaces; no online fine-tuning.
• Identity aggregation rule approaches sample-level coupling in high-dimensional joint action spaces (e.g., terran_10_vs_10), weakening gains.

Experimental setup

Datasets: SMACv1 offline datasets (Good/Medium/Poor) SMACv2 replay datasets MPE Simple Spread (Expert/Medium/Medium-Replay/Random)

Benchmarks: SMACv1 SMACv2 MPE Simple Spread landmark diagnostic XOR strict-product task

Metrics: fan-free route consistency success balance average reward Jacobian off-diagonal ratio

Baselines: Flow matching IMLE policy Drifting model MABCQ MACQL BC Diffusion BC MADiff DoF Flow BC MAC-Flow MATD3BC

Builds on: AlignFlow MAC-Flow flow matching IMLE Drifting model MADiff DoF

Closest papers by meaning

• Beyond Training Time, Test-Time Coordination is Essential for Cooperative MARL (NeurIPS 2026 · similarity 0.87)
• Multi-agent Coordination via Flow Matching (ICLR 2026 · similarity 0.87)
• Dispelling the Mirage of Progress in Offline MARL through Standardised Baselines and Evaluation (NeurIPS 2024 · similarity 0.86)
• Skill-SD: Skill-Conditioned Self-Distillation for Multi-turn LLM Agents (NeurIPS 2026 · similarity 0.86)
• Learning Multimodal One-step Flow Policy via Value-weighted Optimal Transport (NeurIPS 2026 · similarity 0.86)
• Kaleidoscope: Learnable Masks for Heterogeneous Multi-agent Reinforcement Learning (NeurIPS 2024 · similarity 0.85)