A game AI that learns entirely from self-play reinforcement learning and next-token prediction of raw game history, with truly zero domain knowledge — no policy priors, no hand-crafted features, what you see is what you get, surpassing hand-crafted SOTA in GuanDan (掼蛋) and DouDiZhu (斗地主), two complex multi-player trick-taking card games hugely popular across China.
Disclaimer: This project was developed through ~100% vibe coding (powered by Claude Opus 4.6). While extensively tested, the code and documentation may contain critical bugs, hallucinations, or inaccuracies. We are actively working on fixing these issues. Use at your own risk and verify critical results independently. If you encounter any problems, feel free to open an issue.
(2026-07-11) 🪐 Call for collaboration: We have noticed a profound connection between our work and the recently popular concept of world models. Microsoft's ECHO and PaW adopted very similar approaches, co-training NTP with RL objectives to learn an implicit on-policy world model for LLM agents. Our team is also working toward this direction to bridge the gap between world models and agentic LLMs. We already have some great ideas and are looking for collaborators to realize them. If you are interested, feel free to drop me an email.
BTW, we noticed an interesting work FableDan that implemented a variant of DanLM, provided fully open-source training code, and is also SOTA on the Botzone leaderboard. Refer to this work if you are interested in training a DanLM-like model yourself.
(2026-05-01) 🔥 We proposed DouLM, the DouDiZhu version of DanLM, which has reached #1 on the Botzone FightTheLandlord leaderboard, surpassing all the other 491 bots.
(2026-04-03) 🔥 DanLM has reached #1 on the Botzone GuanDan leaderboard, surpassing all the other 30 bots.
graph LR
feat["567-dim Hand-Crafted Features<br/>(manual game history filtering, pre-calculated statistics, secondary information like remaining cards, etc.)"] --> mlp["MLP"] --> q1["Q(s,a)"]
style feat fill:#ffebee,stroke:#E53935
style mlp fill:#ffebee,stroke:#E53935
style q1 fill:#ffebee,stroke:#E53935
DanLM (This Work) — zero domain knowledge, learn everything from raw observations with next-token prediction
graph LR
history["Game History<br/>(tokenized play record, raw public information only)"] --> encoder["TinyLM Encoder<br/>(game world model)"]
hand["Hand + Action<br/>(simple count/onehot vectors)"] --> handmlp["Hand MLP"]
encoder -->|"predictive embedding"| qhead["Q-Value Head"]
encoder -->ntp["Next-Token Prediction for game dynamics learning<br/>(auxiliary task)"]
handmlp -->|"hand state"| qhead
qhead --> q2["Q(s,a)"]
style ntp fill:#f3e5f5,stroke:#9C27B0
style history fill:#e8f5e9,stroke:#43A047
style hand fill:#e8f5e9,stroke:#43A047
style encoder fill:#fff3e0,stroke:#FF9800
style handmlp fill:#e8f5e9,stroke:#4CAF50
style qhead fill:#e8f4f8,stroke:#2196F3
style q2 fill:#e8f5e9,stroke:#43A047
Existing card game AI systems (DouZero, DanZero, PerfectDou, Suphx, etc.) usually rely on carefully designed hand-crafted features, including too much domain knowledge like pre-calculated statistics and secondary information of the game.
Inspired by modern LLMs, we model card play as a causal sequence-modeling problem and leverage the powerful next-token prediction objective to learn game dynamics from raw game records.
DanLM shows that raw game history speaks for itself just by predicting the next token. The input is simply the raw play-by-play game transcript — who played what cards, in order — tokenized like natural language. The model learns what matters from scratch, through self-play RL and causal sequence modeling.
| Aspect | Previous SOTA (DanZero) | DanLM |
|---|---|---|
| State features | 567-dim hand-crafted | Raw token sequence |
| Architecture | MLP | TinyLM + MLP |
| Domain knowledge | Yes | No |
| Training | DMC self-play | DMC self-play + NTP |
| Model | Main Architecture | State Representation |
|---|---|---|
| DanZero | MLP | 567-dim hand-crafted |
| DanZero V1T | MLP | 964-dim hand-crafted |
| DanLM V1 | causal Transformer | raw info tokenization (~90 vocab) |
DanZero V1T is our enhanced reproduction of DanZero with bug fixes, stronger state representation, and fully model-driven action selection (original DanZero uses heuristics for tribute). It exceeds the original DanZero's performance under the same training settings, serving as a stronger hand-crafted baseline.
- Single-Round Win Rate: One round with a random level, random tribute/back, and random deal. The team whose player finishes first wins. This is a stricter metric with less variance amplification.
- Whole-Game Win Rate: A complete game consisting of multiple rounds with level progression from 2 to A. This is the metric used in the original DanZero paper (Table I). Single-round advantages are amplified over multiple rounds (e.g., 54% single-round → ~66% whole-game).
We include 16 competition bots from the 1st National GuanDan AI Algorithm Competition (首届中国人工智能掼蛋算法大赛) as standardized evaluation baselines, consistent with the DanZero paper's evaluation protocol.
Key finding: competition rankings do NOT reflect actual bot strength. Many bots have critical bugs in their source code that caused crashes during the competition. For example, njupt-guandan-ai only won a consolation prize due to a None-check bug causing ~49% of games to crash — but it is actually the strongest bot after the fix. We fixed all bots one by one and evaluated against their bug-free versions.
See baselines/ for the full bots collection.
Evaluation results of DanLM against DanZero, DanZero V1T, and the 5 strongest baseline bots: njupt-guandan-ai, chick-squad, guanglan-iot, egg-expert, and egg-pancake.
Random single-round win rate (1000 rounds, seed=42):
| DanZero | DanZero V1T | njupt-guandan-ai | chick-squad | guanglan-iot | egg-expert | egg-pancake | |
|---|---|---|---|---|---|---|---|
| DanZero | - | - | 71.4% | 74.6% | 77.3% | 78.8% | 79.1% |
| DanZero V1T | 62.1% | - | 77.8% | 80.8% | 81.8% | 83.2% | 85.9% |
| DanLM | 65.1% | 59.6% | 81.9% | 82.1% | 80.9% | 82.3% | 85.9% |
Whole-game win rate (1000 games, seed=42):
| DanZero | DanZero V1T | njupt-guandan-ai | chick-squad | guanglan-iot | egg-expert | egg-pancake | |
|---|---|---|---|---|---|---|---|
| DanZero | - | - | 95.5% | 98.9% | 98.6% | 99.1% | 99.0% |
| DanZero V1T | 87.5% | - | 99.1% | 99.9% | 100.0% | 100.0% | 99.8% |
| DanLM | 97.5% | 74.9% | 100.0% | 100.0% | 99.8% | 99.9% | 99.9% |
We ablated training w/ NTP (red) and w/o NTP (orange). NTP loss helps regularize predictive representation learning and boosts the performance by a large margin.
- Python 3.12
- macOS ARM64 (Apple Silicon)
pip install torch numpy onnxruntime
pip install fastapi uvicorn # for UI# DanLM vs random
PYTHONPATH=. python scripts/evaluate.py \
--model ckpts/DanLM_v1/dansformer_v1_best_eval.pt \
--games 100
# DanLM vs DanZero V1T (hand-crafted SOTA)
PYTHONPATH=. python scripts/evaluate.py \
--model ckpts/DanLM_v1/dansformer_v1_best_eval.pt \
--model-b ckpts/DanZero_v3_rep_v1t/v3_rep_v1t_best_eval_001_int8.onnx \
--games 500
# DanLM vs baseline bot
PYTHONPATH=. python scripts/evaluate.py \
--model ckpts/DanLM_v1/dansformer_v1_best_eval.pt \
--model-b bot:fin-njupt-guandan-ai \
--games 500
# Whole-game evaluation
PYTHONPATH=. python scripts/evaluate_game.py \
--model ckpts/DanLM_v1/dansformer_v1_best_eval.pt \
--model-b ckpts/DanZero_v3_rep_v1t/v3_rep_v1t_best_eval_001_int8.onnx \
--games 100
Play GuanDan against the AI in your browser with built-in AI Hint support showing Q-value estimates for each legal play.
PYTHONPATH=. python ui/server.py
# Open http://localhost:8000Choose from 3 AI agents:
- DanZero V0 — MLP baseline
- DanZero V1T — Hand-crafted feature SOTA
- DanLM V1 — Feature-free TinyLM agent (ours)
Apache License 2.0 with additional non-commercial restriction. See LICENSE for details.
Free for academic research and personal use. Commercial use requires written permission from the author.
- DanZero: Lu et al., "DanZero: Mastering GuanDan Game with Reinforcement Learning", AAAI 2023. [paper]
- DouZero: Zha et al., "DouZero: Mastering DouDizhu with Self-Play Deep Reinforcement Learning", ICML 2021. [paper] [code]
- PerfectDou: Yang et al., "PerfectDou: Dominating DouDizhu with Perfect Information Distillation", NeurIPS 2022. [paper]
- Suphx: Li et al., "Suphx: Mastering Mahjong with Deep Reinforcement Learning", 2020. [paper]
If you use this work, please cite this repository.