Research map / Multi-agent systems and games / LLM agents
Distilling Task-Level Coordination Policies for Generalizable Multi-Agent Cooperation
Zimo Zhai, Manjie Xu, Wei Liang
Read the paperCheck your idea against 137,217 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 SynCoord, a self-supervised pipeline that distills LLM-derived task-level coordination policies into smaller models for generalizable multi-agent cooperation.
Problem
Directly deploying LLMs as high-level coordinators for multi-agent systems is hard because effective policies often fail to emerge reliably at the low-level control stage, and inference costs limit scalability.
Approach
Self-supervised synthetic coordination distillation (SynCoord) that transfers task-level coordination policies into a model for multi-agent cooperation, avoiding costly per-step LLM inference. Details beyond the abstract are not available.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the single-agent bootstrap phase (ablation) | SynCoord · Overcooked-AI | success rate: 0.65 -> 0.95 | Improved |
| add the single-agent bootstrap phase (ablation) | SynCoord · Overcooked-AI | teacher data accuracy: 0.90 -> 0.97 | Improved |
| use 235B teacher instead of 30B teacher (ablation) | SynCoord · Overcooked-AI | success rate: 0.92 -> 0.95 | Improved |
| use 235B teacher instead of 8B teacher (ablation) | SynCoord · Overcooked-AI | success rate / data accuracy: SR 0.80 -> 0.95; Data Acc 0.69 -> 0.97 | Improved |
| use 7B intuitive agent instead of 1.5B (ablation) | SynCoord · Overcooked-AI | success rate: 0.00 -> 0.95 | Improved |
| distill task-level coordination into a 7B model instead of prompting the 235B thinking agent | SynCoord · Overcooked-AI | decision latency: 15609 ms -> 350 ms | Improved |
| distilled SynCoord versus prompted LLMs on 5x13 layouts | SynCoord · Overcooked-AI 5x13 | avg success timesteps: 39.97 vs 44.24 (Gemini-3-flash), 46.88 (Qwen3-235B), 48.25 (GPT-5.2) | Improved |
| expand training layouts from 1x1 to 1x13 for RL baselines | MAPPO · Overcooked-AI | success rate: 1.00 -> 0.21 | Hurt |
| unseen versus seen recipes | SynCoord · Overcooked-AI 17 unseen recipes | success rate: 0.98 -> 0.97 | No clear effect |
| zero-shot transfer from 1-2 agents to larger LBF layout with 100 steps | SynCoord · LBF 15x15-3p-5f | success rate: 0.00 -> 0.80 | Improved |
| increase team size to three agents | SynCoord · Overcooked-AI random layouts | success rate: near-perfect -> 82.8% | Hurt |
Closest papers by meaning
- COOP$^2$: Defining, Observing, and Repairing Cooperation in LLM Multi-Agent Systems (NeurIPS 2026 · similarity 0.90)
- LLM-Assisted Semantically Diverse Teammate Generation for Efficient Multi-agent Coordination (ICML 2025 · similarity 0.90)
- Multi-Agent Collaboration via Evolving Orchestration (NeurIPS 2025 · similarity 0.89)
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.89)
- Learning Latency-Aware Orchestration for Multi-Agent Systems (NeurIPS 2026 · similarity 0.89)
- Learning to Deliberate: Meta-policy Collaboration for Agentic LLMs with Multi-agent Reinforcement Learning (AAAI 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
