Research map / Large language models / Instruction and fine-tuning
Can Large Language Models Master Complex Card Games?
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.
Fine-tunes Qwen2.5, Llama3.1 and GLM4 on high-quality gameplay data from strong game AIs across eight card games to test whether LLMs can master complex card games, learn several at once, and keep general capabilities.
Problem
Whether LLMs can learn to master complex card games (learning ability rather than prompt-only play), how much data is needed, whether one model can handle several games at once, and whether general capabilities survive game fine-tuning.
Approach
Instead of learning from scratch through environment exploration, convert trajectories from strong game AIs (DouZero, DanZero, Mortal, RLCard DQN, rule models) into instruction-output SFT pairs with game-specific prompt templates, then fine-tune LLMs with LoRA on single games and on a mixture of all eight games.
Key results
- On DouDizhu, fine-tuned Qwen2.5-7B reaches an average win rate of 0.806 against the rule opponents, and Qwen2.5-14B shows a landlord win rate of 0.858 but farmer win rate of 0.570.
- The mixed-game GLM4-9B-Chat-mix achieves the best DouDizhu (0.882) and GuanDan (0.698) scores among fine-tuned models, while API models such as DeepSeek-R1 peak at 0.185 DouDizhu and 0.020 GuanDan.
- After fine-tuning on game data, the mixed Llama3.1-8B-Instruct model drops from 47.95 to 44.74 on MMLU-Pro and from 46.60 to 35.20 on Math-500 (GLM4-9B drops from 52.40 to 37.20 on Math-500), and general-data recovery only partly restores these.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · DouDizhu | win rate: 0.087 -> 0.852 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · GuanDan | round win rate: 0.000 -> 0.634 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · Riichi Mahjong | reward score: 0.04 -> 1.08 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · Uno | reward score: 0.032 -> 0.108 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · Gin Rummy | reward score: -0.530 -> 0.177 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · Leduc Hold’em | reward score: 0.63 -> 1.24 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · Limit Texas Hold’em | reward score: 1.05 -> 2.66 | Helped |
| Fine-tune on high-quality gameplay data from all eight games instead of using the base model | Qwen2.5-7B-Instruct · No-limit Texas Hold’em | reward score: 1.25 -> 4.86 | Helped |
| Add general knowledge, math, and coding data to the mixed game fine-tuning data (ablation) | Llama3.1-8B-Instruct-mix · MMLU-Pro | score: 44.74 -> 45.18 | Helped |
| Add general knowledge, math, and coding data to the mixed game fine-tuning data (ablation) | Llama3.1-8B-Instruct-mix · Math-500 | score: 35.20 -> 47.20 | Helped |
| Add general knowledge, math, and coding data to the mixed game fine-tuning data (ablation) | Llama3.1-8B-Instruct-mix · HumanEval | score: 60.98 -> 65.24 | Helped |
| Add general knowledge, math, and coding data to the mixed game fine-tuning data (ablation) | Llama3.1-8B-Instruct-mix · Uno | reward score: 0.164 -> 0.208 | Mixed |
Surprising findings
- The 14B Qwen2.5 model performs worse overall than the 7B model on DouDizhu because its farmer win rate falls to 0.570, despite a stronger landlord role.
- Fine-tuning on the mixed game set improves Doudizhu and GuanDan but declines performance on the other six games relative to single-game models, indicating conflicts between game groups.
Limitations
- Training data comes from strong game AIs and rule models rather than LLM self-play, and the comparison to specialized AIs is not fair in compute or data.
- Only rule-based or weak opponents (random, rule models) are used for several games, and Mahjong has no available teacher model for comparison.
- Farmer-role performance is limited by a filtering rule that keeps low-quality data from the non-key farmer.
Experimental setup
Datasets: DouZero-generated DouDizhu trajectories DanZero-generated GuanDan trajectories Tenhou expert Riichi Mahjong data Rule-model Uno and Gin Rummy trajectories RLCard DQN Leduc, Limit and No-limit Hold'em trajectories Tulu-3 post-training data
Benchmarks: MMLU-Pro Math-500 HumanEval GPQA-Diamond AIME2024 LiveCodeBench IFEval
Metrics: win rate round win rate reward score
Baselines: GPT-4o GPT-4o-mini DeepSeek-V3 DeepSeek-R1 GLM-4-plus GLM-4-air Qwen2.5-7B-Instruct Llama3.1-8B-Instruct GLM4-9B-Chat
Builds on: DouZero DanZero Mortal RLCard LLaMA-Factory LoRA
Closest papers by meaning
- Mastering Board Games by External and Internal Planning with Language Models (ICML 2025 · similarity 0.90)
- PokerBench: Training Large Language Models to Become Professional Poker Players (AAAI 2025 · similarity 0.90)
- GTBench: Uncovering the Strategic Reasoning Capabilities of LLMs via Game-Theoretic Evaluations (NeurIPS 2024 · similarity 0.88)
- Large Language Models Play StarCraft II:Benchmarks and A Chain of Summarization Approach (NeurIPS 2024 · similarity 0.88)
- Mixing Expert Knowledge: Bring Human Thoughts Back To the Game of Go (NeurIPS 2025 · similarity 0.87)
- LMRL Gym: Benchmarks for Multi-Turn Reinforcement Learning with Language Models (ICML 2025 · 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.