Research map / Multi-agent systems and games / Human-AI and goal-conditioned
K-level Reasoning for Zero-Shot Coordination in Hanabi
Brandon Cui, Hengyuan Hu, Luis Pineda, Jakob Nicolaus Foerster
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 |
|---|---|---|---|
| train KLR hierarchy levels synchronously rather than sequentially (ablation) | KLR · Hanabi cross-play | score: 17.17\pm 0.98 -> 22.14\pm 0.12 | Improved |
| train KLR hierarchy levels synchronously rather than sequentially (ablation) | KLR · Hanabi with Color Bot | score: 14.80\pm 1.77 -> 17.21 \pm 0.86 | Improved |
| train KLR hierarchy levels synchronously rather than sequentially (ablation) | KLR · Hanabi with Rank Bot | score: 12.36\pm1.44 -> 15.90\pm0.30 | Improved |
| train KLR hierarchy levels synchronously rather than sequentially (ablation) | KLR · Hanabi with Clone Bot | score: 13.03 \pm 1.91 -> 15.86 \pm 0.23 | Improved |
| co-train a best response to the KLR hierarchy (ablation) | Synchronous KLR · Hanabi cross-play | score: 22.14\pm 0.12 -> 23.29 \pm 0.05 | Improved |
| co-train a best response to the KLR hierarchy (ablation) | Synchronous KLR · Hanabi with Rank Bot | score: 15.90\pm0.30 -> 17.01 \pm 0.45 | Improved |
| co-train a best response to the KLR hierarchy (ablation) | Synchronous KLR · Hanabi with Clone Bot | score: 15.86 \pm 0.23 -> 16.59\pm 0.16 | Improved |
| replace KLR's preceding-level partner with a Poisson mixture of lower levels (ablation) | Synchronous KLR · Hanabi cross-play | score: 22.14\pm 0.12 -> 21.42\pm\textbf{0.10} | Hurt |
| replace KLR's preceding-level partner with a Poisson mixture of lower levels (ablation) | Synchronous KLR · Hanabi with Clone Bot | score: 15.86 \pm 0.23 -> 14.59 \pm 0.25 | Hurt |
| evaluate against independently trained level-4 partners rather than the training partner | Sequential KLR (level 5) · Hanabi level-4 partner cross-play | percentage of games bombing out: 0.4 \pm 0.2 -> 19.8 \pm 18.0 | Hurt |
| evaluate against independently trained level-4 partners rather than the training partner | Synchronous KLR (level 5) · Hanabi level-4 partner cross-play | percentage of games bombing out: 0.4 \pm 0.2 -> 0.6 \pm 0.3 | No clear effect |
Closest papers by meaning
- Cross-environment Cooperation Enables Zero-shot Multi-agent Coordination (ICML 2025 · similarity 0.88)
- Sparks of Cooperative Reasoning: LLMs as Strategic Hanabi Agents (ICML 2026 · similarity 0.88)
- Ad-Hoc Human-AI Coordination Challenge (ICML 2025 · similarity 0.87)
- A Generalist Hanabi Agent (ICLR 2025 · similarity 0.87)
- OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination (ICLR 2025 · similarity 0.87)
- ZSC-Eval: An Evaluation Toolkit and Benchmark for Multi-agent Zero-shot Coordination (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.
