Research map / Multi-agent systems and games / Cooperative multi-agent RL
Role-Level Inductive Bias for Cross-Task Generalization in Multi-Agent Reinforcement Learning
Chang Yao, Youfang Lin, Shoucheng Song, Hao Wu, Shengkun Yang, Yuqing Ma, Kai Lv
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 |
|---|---|---|---|
| remove UW-DPP diverse role selection (ablation) | GTR · SMAC zero-shot 4s5z_vs_4s7z | test win rate: 0.08 -> 0.47 (adding UW-DPP) | Improved |
| add intention prediction regularization (ablation) | GTR · SMAC zero-shot 4s5z_vs_4s7z | test win rate: 0.09 -> 0.47 | Improved |
| add influence prediction regularization (ablation) | GTR · SMAC zero-shot 4s5z_vs_4s7z | test win rate: 0.18 -> 0.47 | Improved |
| add role embedding in observation encoder (top-down perception) (ablation) | GTR · SMAC zero-shot 4s5z_vs_4s7z | test win rate: 0.14 -> 0.47 | Improved |
| add role embedding in the mixer (ablation) | GTR · SMAC zero-shot 4s5z_vs_4s7z | test win rate: 0.38 -> 0.47 | Improved |
| add role components on source tasks (ablation) | GTR · SMAC source tasks (5m_vs_6m, 3s_vs_5z, 3s5z_vs_3s6z) | test win rate: e.g. 5m_vs_6m 0.89 -> 0.92 (vs w/o intention) | Improved |
| use GTR role-level bias instead of entity-level baselines (HPN, RPG, ASN, UPDeT) | zero-shot transfer · SMAC unseen tasks | test win rate: 8m vs 9m: 0.56 (RPG) -> 0.91; 4s vs 8z: 0.25 -> 0.72 | Improved |
| use GTR instead of baselines on source tasks | zero-shot transfer · SMAC source tasks | test win rate: 5m vs 6m: 0.95 ± 0.08 (HPN) vs 0.92 ± 0.04 | No clear effect |
| initialize from pretrained GTR instead of training from scratch | GTR · SMAC few-shot 4s5z_vs_4s7z | test win rate | Improved |
Closest papers by meaning
- Attention-Guided Contrastive Role Representations for Multi-agent Reinforcement Learning (ICLR 2024 · similarity 0.87)
- R3DM: Enabling Role Discovery and Diversity Through Dynamics Models in Multi-agent Reinforcement Learning (ICML 2025 · similarity 0.86)
- Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation (ICLR 2025 · similarity 0.86)
- Training Generalizable Collaborative Agents via Strategic Risk Aversion (NeurIPS 2026 · similarity 0.85)
- Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts (ICLR 2024 · similarity 0.85)
- PLATO: Pointer Learner for Agent and Task Openness (NeurIPS 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.
