Research map / Multi-agent systems and games / Cooperative multi-agent RL
COMBO: Compositional World Models for Embodied Multi-Agent Cooperation
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 COMBO, a compositional video world model conditioned on factorized multi-agent joint actions, combined with VLM-based action proposer, intent tracker, and outcome evaluator in a tree search for embodied multi-agent cooperation.
Problem
Embodied multi-agent cooperation from only partial egocentric views requires simulating world dynamics conditioned on an arbitrary number of agents' joint actions, which single-agent world models do not handle.
Approach
Factorize joint actions into per-agent text components and compose their score functions in a video diffusion model (composable diffusion), with agent-dependent loss scaling; plug this compositional world model into a VLM-driven tree search that proposes own actions, infers other agents' intents, and evaluates outcomes.
Key results
- On TDW-Game, COMBO reaches 1.00 success rate with 17.5 and 17.4 average steps for the two cooperators, versus 0.55-0.60 success and 28.4-29.2 steps for LLaVA.
- Compositional world model achieves 75% multi-agent video accuracy on TDW-Game (vs 20% for AVDC) and 100% on TDW-Cook (vs 25% for AVDC) in human evaluation.
- On 2D-FetchQ, COMBO achieves 81.3% success rate versus 53.3% for Co-GAIL.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the Intent Tracker to COMBO (ablation) | COMBO · TDW-Game | success rate: 0.65 -> 1.00; 0.60 -> 1.00 | Helped |
| add the Intent Tracker to COMBO (ablation) | COMBO · TDW-Cook | success rate: 0.80 -> 0.90; 0.80 -> 1.00 | Helped |
| add Agent-Dependent Loss Scaling (ablation) | COMBO world model · TDW-Game | multiple-action video accuracy: 55% -> 75% | Helped |
| add Agent-Dependent Loss Scaling (ablation) | COMBO world model · TDW-Cook | multiple-action video accuracy: 70% -> 100% | Helped |
| increase action proposals, rollout depth, and beam from 3/1/1 to 2/2/2 (ablation) | COMBO planning · TDW-Game | success rate: 50% -> 70% | Helped |
| increase action proposals, rollout depth, and beam from 2/2/2 to 3/3/3 (ablation) | COMBO planning · TDW-Game | success rate: 70% -> 100% | Helped |
| use COMBO instead of LLaVA | embodied multi-agent planning · 2D-FetchQ | success rate: 53.3 -> 81.3 | Helped |
Surprising findings
- A simpler supervised model with CLIP and BERT encoders is far worse than LLaVA on the outcome evaluator (62.7 vs 99.7 accuracy), showing VLM advantages.
- Model trained only on four-agent TDW-Game data generalizes to three- and two-agent settings without retraining.
Limitations
- Inference is slow due to repeated calls to large generative models in tree search.
- Image quality errors (about 80% of world-model failures) arise from CFG on multiple text conditions, causing contradictory actions.
- Baselines like MAPPO and Recurrent World Models fail entirely on the tasks, and the Intent Tracker generalizes imperfectly (64.6-75.8% intent accuracy on unseen cooperators).
Experimental setup
Datasets: TDW-Game random rollouts TDW-Cook random rollouts 2D-FetchQ human-human demonstrations
Benchmarks: TDW-Game TDW-Cook 2D-FetchQ
Metrics: success rate average steps of successful episodes human accuracy of generated videos (single and multiple agen
Baselines: Recurrent World Models MAPPO CoELA LLaVA-1.5-7B Shared Belief Cooperator AVDC BC-single BC-GAIL Co-GAIL
Builds on: AVDC composable diffusion models VLP RoboDreamer Recurrent World Models
Closest papers by meaning
- LLawCo: Learning Laws of Cooperation for Modeling Embodied Multi-Agent Behavior (ICML 2026 · similarity 0.89)
- VIKI‑R: Coordinating Embodied Multi-Agent Cooperation via Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.88)
- Embodied Multi-Modal Agent trained by an LLM from a Parallel TextWorld (CVPR 2024 · similarity 0.87)
- World Models as Group Actions (NeurIPS 2026 · similarity 0.87)
- Collaborative Tree Search for Enhancing Embodied Multi-Agent Collaboration (CVPR 2025 · similarity 0.87)
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.