Research map / Multi-agent systems and games / Human-AI and goal-conditioned
CooT: Learning to Coordinate In-Context with Coordination Transformers
Huai-Chih Wang, Hsiang-Chun Chuang, Hsi-Chun Cheng, Dai-Jie Wu, Shao-Hua Sun
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 CooT in-context best-response transformer instead of HSP | HSP · Overcooked overall (5 layouts) | average return: 56.16 -> 68.79 | Improved |
| use CooT instead of HSP on coordination-heavy layout | HSP · Overcooked Coord. Ring Multi-recipe | average return: 29.35 -> 45.96 | Improved |
| use CooT instead of HSP on layout where agents act independently | HSP · Overcooked Asymm. Adv. | average return: 134.01 -> 129.48 | No clear effect |
| add gradient-based fine-tuning at test time to HSP (HSP-ft) | HSP · Overcooked overall | average return: 56.16 -> 56.33 | No clear effect |
| add PEARL-style trajectory encoder to HSP (HSP-meta) | HSP · Overcooked overall | average return: 56.16 -> 40.19 | Hurt |
| add peer identifier on top of latent context encoder (PACE vs HSP-meta) | HSP-meta · Overcooked overall | average return: 40.19 -> 35.95 | Hurt |
| tune HSP-meta learning rate to 1e-4 (ablation) | HSP-meta · Overcooked overall | average return: 40.19 -> 37.60 | Hurt |
| increase context length from 1 up to 5 episodes (ablation) | CooT · Overcooked Coord. Ring | mean return: best 41.1 at 5 episodes | Improved |
| increase context length from 5 to 7 episodes (ablation) | CooT · Overcooked Coord. Ring | mean return | Hurt |
| step-wise temporal shuffling of context as augmentation (ablation) | CooT · Overcooked | return | Hurt |
| chunk-wise shuffling (20-step blocks) of context as augmentation (ablation) | CooT · Overcooked | return | No clear effect |
| reduce number of training partners from 36 to 12 (ablation) | CooT vs BC · Overcooked Coord. Ring | performance drop: 24.0% drop (CooT) vs 41.5% (BC) | Hurt |
Closest papers by meaning
- Adaptively Coordinating with Novel Partners via Learned Latent Strategies (NeurIPS 2025 · similarity 0.91)
- Learning “Partner-Aware” Collaborators in Multi-Party Collaboration (NeurIPS 2025 · similarity 0.88)
- NestRL: A Nested Training Regime for Mutual Adaptation in Human–AI Teaming (NeurIPS 2026 · similarity 0.88)
- Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination (ICML 2025 · similarity 0.88)
- Partner Modelling Emerges in Recurrent Agents (But Only When It Matters) (NeurIPS 2025 · similarity 0.88)
- Learning to Cooperate with Humans using Generative Agents (NeurIPS 2024 · 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.
