DanLM - 掼蛋斗地主AI模型（自博弈强化学习）

项目说明

DanLM: Next Token Is All You Need to Master Complex Card Games

A 游戏 AI that learns entirely from self-play reinforcement learning and next-token prediction of raw 游戏 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 de through ~100% vibe coding (powered by Claude Opus 4.6). While extensively tested, the 代码 and documentation may contain critical bugs, hallucinations, or inaccuracies. We are actively working on fixing these issues. Use at your own risk and critical results independently. If you encounter any problems, feel free to open an issue.

Updates

(2026-07-11) 🪐 Call for collaboration: We ha 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 of DanLM, provided fully open-源码 training 代码, 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 斗地主 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.

!Botzone GuanDan Ranking

---

Architecture

Previous SOTA (DanZero) — requires domain knowledge

DanLM (This Work) — zero domain knowledge, learn e from raw obser with next-token prediction

Key Idea

Existing card 游戏 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 游戏.

Intuition

Inspired by modern LLMs, we model card play as a causal sequence-modeling problem and le the powerful next-token prediction objective to learn 游戏 dynamics from raw 游戏 records.

DanLM shows that raw 游戏 history speaks for itself just by predicting the next token. The input is simply the raw play-by-play 游戏 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 |

E

Models

| 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-dri 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.

Metrics

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 amplification.

Whole-Game Win Rate: A complete 游戏 consisting of multiple rounds with le from 2 to A. This is the metric used in the original DanZero paper (Table I). Single-round ad are amplified o rounds (e.g., 54% single-round → ~66% whole-game).

Baseline Bots

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 e baselines, consistent with the DanZero paper's e 协议.

Key finding: competition rankings do NOT reflect actual bot strength. Many bots ha bugs in their 源码代码 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 e against their bug-free .

See baselines/ for the full bots collection.

Results

E 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-游戏 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% |

#### NTP ablation

We ablated training w/ NTP (red) and w/o NTP (orange). NTP loss helps regularize predicti learning and boosts the performance by a large margin.

Quick Start to Reproduce the Results

#### Requirements

Python 3.12

macOS ARM64 (Apple Silicon)

#### Install

#### E

Web UI - Enjoy playing with the AI yourself!

!Game UI

Play GuanDan against the AI in your browser with built-in AI Hint support showing Q-value estimates for each legal play.

Choose from 3 AI agents:

DanZero V0 — MLP baseline

DanZero V1T — Hand-crafted feature SOTA

DanLM V1 — Feature-free TinyLM agent (ours)

License

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.

References

DanZero: Lu et al., "DanZero: Mastering GuanDan 游戏 with Reinforcement Learning", AAAI 2023. [[paper]]()

DouZero: Zha et al., "DouZero: Mastering 斗地主 with Self-Play Deep Reinforcement Learning", ICML 2021. [[paper]]() [[代码]]()

PerfectDou: Yang et al., "PerfectDou: Dominating 斗地主 with Perfect Information Distillation", NeurIPS 2022. [[paper]]()

Suphx: Li et al., "Suphx: Mastering 麻将 with Deep Reinforcement Learning", 2020. [[paper]]()

Citation

If you use this work, please cite this repository.

技术栈

开发语言：未知

项目类型：掼蛋斗地主AI模型（自博弈强化学习）

Star 数：0

功能特点

这是一套掼蛋斗地主AI模型（自博弈强化学习），代码结构清晰，适合开发者学习棋牌游戏的服务端与客户端开发。项目涵盖了棋牌游戏的核心玩法逻辑、网络通信、用户管理、房间匹配等模块，对理解棋牌类游戏的完整开发流程具有较高的参考价值。

从技术实现角度看，棋牌类项目的难点在于：游戏逻辑的状态机设计、并发房间管理、牌型算法、以及服务端与客户端的协议同步。本项目在这些方面提供了可运行的实现，开发者可以在此基础上进行二次开发或学习研究。

部署说明

1. 下载源码压缩包，解压到本地开发环境。

2. 根据项目类型准备对应的运行环境（Node.js / Go / C++ / Java 等）。

3. 按照 README 中的步骤配置依赖、数据库和配置文件。

4. 先启动服务端，再启动客户端，即可进入游戏进行测试。

注意：本项目仅供学习研究使用，请勿用于商业或非法用途。

学习价值

对于想进入棋牌游戏开发领域的开发者来说，这套源码是一个很好的入门参考。通过阅读代码，可以学习到棋牌游戏的核心玩法实现、网络协议设计、房间与匹配机制、以及游戏状态管理等关键技术点。

同时，项目中的代码组织方式和模块划分，也展示了棋牌类项目在工程化方面的实践经验，值得开发者深入研究和借鉴。理解一个完整棋牌项目的架构，对后续开发其他实时对战类游戏也很有帮助。

免责声明

本资源来源于开源社区，仅供学习研究和技术交流使用，请勿用于任何商业或违法违规用途。使用本项目产生的一切后果由使用者自行承担。若涉及第三方权益，请联系删除。

登录后才能参与讨论

下载链接

会员才能下载

登录后开通会员

同分类推荐

暂无同类资源

资源标题	DanLM - 掼蛋斗地主AI模型（自博弈强化学习）
所属分类	游戏源码
发布作者	admin
发布时间	2026-10-05 17:47:40
下载权限	免费下载
附件数量	2 个下载链接
浏览次数	318
讨论数量	0
主题状态	正常讨论中