Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Scaling Multi-Agent Environment Co-Design with Diffusion Models
Hao Xiang Li, Michael Amir, Amanda Prorok
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| co-design the environment with DiCoDe versus a fixed environment | MAPPO agents · D-RWARE Corner | episode reward: 9.6 -> 12.1 (26%) | Improved |
| use DiCoDe instead of RL co-design baseline | co-design · D-RWARE Corner | boxes delivered: 8.7 -> 12.1; 39% more with 66% fewer samples | Improved |
| replace Projected Universal Guidance with gradient descent (ablation) | DiCoDe · D-RWARE Corner | episode reward: 12.1 -> 9.3 | Hurt |
| replace PUG with top-k sampling (ablation) | DiCoDe · D-RWARE Corner | episode reward: 12.1 -> 8.2 | Hurt |
| replace PUG with the ADD diffusion guidance (ablation) | DiCoDe · D-RWARE Corner | episode reward: 12.1 -> 9.1 | Hurt |
| add critic distillation instead of training environment critic on trajectory returns (ablation) | DiCoDe · D-RWARE Corner | episode reward: 10.9 -> 12.1 | Improved |
| use shelf coordinate representation instead of binary mask (ablation) | DiCoDe · D-RWARE Corner | episode reward: 12.1 -> 11.7 | No clear effect |
| add constraint projection to universal guidance (PUG vs UG) (ablation) | environment search with fixed critic · D-RWARE Corner | critic-estimated value | No clear effect |
| use learned guided diffusion rather than best of 1024 uniformly sampled environments (ablation) | environment design · D-RWARE Corner | critic-estimated value: best-of-1024 sampling is 12% worse than mean of PUG | Improved |
| use DiCoDe instead of RL co-design as turbines increase | co-design · WFCRL 2 to 16 turbines | episode reward: WFCRL16 254 -> 282 | Improved |
| use DiCoDe on WFCRL and ONav versus baselines | co-design · WFCRL2/4/8/16 and ONav | episode reward: 9.5% above Fixed, 10.3% above RL, 17.1% above DR | Improved |
| use DiCoDe versus gradient descent ablation on ONav (ablation) | DiCoDe · VMAS ONav | episode reward: 2.29 vs 2.29 | No clear effect |
Closest papers by meaning
- Adaptive Collaboration with Humans: Metacognitive Policy Optimization for Multi-Agent LLMs with Continual Learning (ICLR 2026 · similarity 0.85)
- Revisiting Multi-Agent World Modeling from a Diffusion-Inspired Perspective (NeurIPS 2025 · similarity 0.85)
- TodoEvolve: Learning to Architect Agent Planning Systems (ICML 2026 · similarity 0.85)
- HyperMARL: Adaptive Hypernetworks for Multi-Agent RL (NeurIPS 2025 · similarity 0.85)
- CoMAS: Co-Evolving Multi-Agent Systems via Interaction Rewards (ICLR 2026 · similarity 0.84)
- LNS2+RL: Combining Multi-agent Reinforcement Learning with Large Neighborhood Search in Multi-agent Path Finding (AAAI 2025 · similarity 0.84)
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.
