Research map / Multi-agent systems and games / Cooperative multi-agent RL
Learning Multi-Agent Coordination via Sheaf-ADMM
Jeffrey Seely, Bartłomiej Cupiał, Llion Jones
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 |
|---|---|---|---|
| add ADMM coordination iterations (K=0 -> K=20) (ablation) | Sheaf-ADMM · Sudoku | solved rate (%): 0.0% -> 92.5% | Improved |
| use learned shared restriction maps instead of identity maps (ablation) | Sheaf-ADMM · Sudoku | solved rate (%): 6.1% -> 92.5% | Improved |
| add LoRA modulation of restriction maps to shared maps (ablation) | Sheaf-ADMM · Maze | solved rate (%): 8.9% -> 99.8% | Improved |
| add LoRA modulation of restriction maps to shared maps (ablation) | Sheaf-ADMM · Sudoku | solved rate (%): 92.5% -> 87.5% | Hurt |
| add l1 regularization to the diagonal quadratic x-update objective (ablation) | Sheaf-ADMM · Sudoku | solved rate (%): 47.0% -> 92.5% | Improved |
| make the quadratic term learned (diagonal quad vs fixed Q) (ablation) | Sheaf-ADMM · Maze | solved rate (%): 38.9% -> 99.2% | Improved |
| x-update variants on MNIST (fixed Q vs quad vs quad + l1) (ablation) | Sheaf-ADMM · MNIST | accuracy: 98.4% -> 98.2% -> 98.5% | No clear effect |
| use conjugate gradient instead of GD + Nesterov for the z-update (ablation) | Sheaf-ADMM · Sudoku | solved rate (%): 0.0% -> 92.5% | Improved |
| use conjugate gradient instead of GD + Nesterov for the z-update (ablation) | Sheaf-ADMM · MNIST | accuracy: 52.2% -> 98.5% | Improved |
| use conjugate gradient instead of GD + Nesterov for the z-update (ablation) | Sheaf-ADMM · Maze | solved rate (%): 98.2% -> 99.8% | Mixed |
| increase ADMM iterations from K=10 to K=30 (ablation) | Sheaf-ADMM · MNIST | accuracy: 98.5% -> 90.3% | Hurt |
| increase ADMM iterations from K=5 to K=30 (ablation) | Sheaf-ADMM · Maze | solved rate (%): 1.2% -> 99.8% | Improved |
Closest papers by meaning
- Distilling Task-Level Coordination Policies for Generalizable Multi-Agent Cooperation (ICML 2026 · similarity 0.87)
- Towards Efficient Collaboration via Graph Modeling in Reinforcement Learning (AAAI 2025 · similarity 0.86)
- MAT-Agent: Adaptive Multi-Agent Training Optimization (NeurIPS 2025 · similarity 0.85)
- Graph Diffusion for Robust Multi-Agent Coordination (ICML 2025 · similarity 0.85)
- Dr. MAS: Stable Reinforcement Learning for Multi-Agent LLM Systems (NeurIPS 2026 · similarity 0.85)
- Learning to Share: Selective Memory for Efficient Parallel Agentic Systems (ICML 2026 · similarity 0.85)
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.
