Research map / Multi-agent systems and games / Human-AI and goal-conditioned
An Efficient End-to-End Training Approach for Zero-Shot Human-AI Coordination
Xue Yan, Jiaxian Guo, Xingzhou Lou, Jun Wang, Haifeng Zhang, Yali Du
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 |
|---|---|---|---|
| Use E3T (mixture partner + partner modeling) instead of self-play | SP · Overcooked human proxy | average reward: 65.8 ± 5.5 -> 114.1 ± 6.4 | Improved |
| Replace MEP population training with E3T end-to-end | MEP · Overcooked human proxy | average reward: 89.7 ± 7.1 -> 114.1 ± 6.4 | Improved |
| Replace MEP with E3T for training cost | MEP · Overcooked Cramped Rm. | training time (h): 17.9 -> 1.9 | Improved |
| Add partner modeling module to self-play, E3T(+P) (ablation) | SP · Overcooked AI-AI coordination | average reward | Improved |
| Add mixture partner policy to self-play, E3T(+R) (ablation) | SP · Overcooked AI-AI coordination | average reward | Mixed |
| Combine mixture partner and partner modeling, E3T(R+P) (ablation) | E3T(+R), E3T(+P) · Overcooked AI-AI coordination | average reward over layouts | Improved |
| Vary balance parameter epsilon from 0.1 to 0.9 (ablation) | E3T · Overcooked Coord. Ring | average reward: optimal 0.3 and 0.5 | Mixed |
| Set epsilon to 0.0 instead of 0.5 (ablation) | E3T · Overcooked Forced Coord. | average reward | Improved |
| Use E3T with epsilon 0.0 (partner modeling only) instead of SP (ablation) | SP · Overcooked Forced Coord. | average reward | Improved |
| Raise epsilon from 0.0 to 0.5 (ablation) | E3T · Overcooked Coord. Ring | partner action prediction accuracy: 40% -> 55% | Improved |
| Add mixture partner policy to MAPPO self-play | Self-play · Google Football 3 vs 1 with Keeper | zero-shot winning rate (mean): 0.07 ± 0.08 -> 0.49 ± 0.21 | Improved |
| Add mixture partner policy to MAPPO self-play | Self-play · Google Football 3 vs 1 with Keeper | training winning rate: 0.79 ± 0.08 -> 0.76 ± 0.09 | No clear effect |
Closest papers by meaning
- Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination (ICML 2025 · similarity 0.88)
- OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination (ICLR 2025 · similarity 0.87)
- Learning to Cooperate with Humans using Generative Agents (NeurIPS 2024 · similarity 0.87)
- ZSC-Eval: An Evaluation Toolkit and Benchmark for Multi-agent Zero-shot Coordination (NeurIPS 2024 · similarity 0.87)
- Designs for Enabling Collaboration in Human-Machine Teaming via Interactive and Explainable Systems (NeurIPS 2024 · similarity 0.85)
- Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · 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.
