Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Ad-Hoc Human-AI Coordination Challenge
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Introduces AH2AC2, a Hanabi human-AI coordination challenge with API-hosted human proxy agents (BC refined by KL-regularised IPPO), an open 3,079-game human dataset, and baselines including OBL, BR-BC, FCP, HDR-IPPO and DeepSeek-R1.
Problem
Human-AI ad-hoc coordination is hard to evaluate because human evaluation is costly and hard to reproduce, and closed human datasets and proxy agents block open research in Hanabi.
Approach
Train human proxy agents by behavioural cloning on ~147k hanab.live games, then refine them with IPPO under a KL penalty to the BC policy (HDR-IPPO), host them behind a pre-registered API as a standardised evaluation partner, and open-source only a small 3,079-game subset to encourage data-efficient methods.
Key results
- OBL (L4) achieves the highest 2P mean score among non-human-data methods at 21.04, while BR-BC reaches 19.41 and HDR-IPPO 12.76.
- In 3P, DeepSeek-R1 with H-Group conventions prompting reaches a mean of 14.62, the top baseline, slightly ahead of HDR-IPPO at 14.03.
- Human proxies reach self-play means of 20.88 to 22.97 and cut 3P zero-score games from 75.82% (BC) to 0.34%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use regularised IPPO after BC training | Human proxy agents · 2P Hanabi self-play | mean score: +3.0; +4.0 over BC | Helped |
| Use regularised IPPO after BC training | Human proxy agents · 3P Hanabi self-play | mean score: +15.7; +13.9 over BC | Helped |
| Prompt with H-conventions | DeepSeek-R1 · 2P Hanabi human-proxy coordination | mean score: 5.43 vs 9.91 | Helped |
| Use OBL instead of the strongest human-data baseline | 2P Hanabi coordination · 2P Hanabi human-proxy coordination | mean score: 19.41 -> 21.04 | Helped |
| Use HDR-IPPO instead of BC | 3P Hanabi coordination · 3P Hanabi human-proxy coordination | mean score: 3.31 -> 14.03 | Helped |
| Use BR-BC instead of HDR-IPPO | 2P Hanabi coordination · 2P Hanabi human-proxy coordination | mean score: 12.76 -> 19.41 | Helped |
| Use a human proxy instead of BC trained on limited data | 3P Hanabi coordination · 3P Hanabi human-proxy coordination | mean score: 3.31 -> 20.86 | Helped |
| Remove KL regularisation (λ = 0) (ablation) | HDR-IPPO human proxy · Hanabi self-play | self-play score | Hurt |
| Remove KL regularisation (λ = 0) (ablation) | HDR-IPPO human proxy · 3P Hanabi self-play | mean score: 7.19 -> 3.99 | Hurt |
| Replace the non-regularised policy with the BC policy (ablation) | 3P Hanabi cross-play with two human proxies · 3P Hanabi cross-play | mean score: 6.93 -> 18.92 | Helped |
| Add human-data KL regularisation during IPPO (ablation) | HDR-IPPO human proxy · Hanabi coordination with human proxies and BC policies | coordination performance | Helped |
| Fine-tune BC with regularised IPPO | Human proxy policies · Human-action prediction test set | accuracy: 0.63; -0.03 to BC | Hurt |
Surprising findings
- Zero-shot coordination method OBL outperforms all methods that use human data in the two-player setting.
- DeepSeek-R1 prompted without Hanabi-specific fine-tuning outperforms all other baselines in the three-player setting.
Limitations
- Human proxies have not been validated with direct human-AI play experiments.
- DeepSeek-R1 was evaluated on only 100 games versus 1,000 for other methods.
- Only standard Hanabi with two and three players is covered.
Experimental setup
Datasets: hanab.live human gameplay (147,621 games)
Benchmarks: AH2AC2 leaderboard (1,000 games with human proxies per 2P/3P) AH2AC2 action prediction test set
Metrics: mean score median score cross-entropy loss top-10% and top-20% accuracy information per play communicativeness
Baselines: OBL BR-BC FCP OP IPPO BC DeepSeek-R1 HDR-IPPO
Builds on: IPPO Off-Belief Learning Fictitious Co-Play BR-BC Other-Play
Closest papers by meaning
- Sparks of Cooperative Reasoning: LLMs as Strategic Hanabi Agents (ICML 2026 · similarity 0.87)
- A Generalist Hanabi Agent (ICLR 2025 · similarity 0.85)
- OvercookedV2: Rethinking Overcooked for Zero-Shot Coordination (ICLR 2025 · similarity 0.85)
- Holistic Agent Leaderboard: The Missing Infrastructure for AI Agent Evaluation (ICLR 2026 · similarity 0.84)
- Enhancing Human Experience in Human-Agent Collaboration: A Human-Centered Modeling Approach Based on Positive Human Gain (ICLR 2024 · similarity 0.84)
- Constrained Human-AI Cooperation: An Inclusive Embodied Social Intelligence Challenge (NeurIPS 2024 · similarity 0.84)
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.