Research map / Large language models / Instruction and fine-tuning
PokerBench: Training Large Language Models to Become Professional Poker Players
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 PokerBench, an 11k-spot benchmark (1k pre-flop, 10k post-flop) of GTO poker decisions with a 280k-spot training set, evaluates LLMs, and fine-tunes Llama-3-8B to beat GPT-4 on accuracy and in heads-up play.
Problem
LLMs are poorly evaluated on strategic incomplete-information games like poker, and there is no quick reliable benchmark for measuring their poker-playing ability against game-theory-optimal (GTO) play.
Approach
A balanced, GTO-derived set of pre-flop and post-flop poker decision spots, formatted as text prompts, serves both as a fast benchmark and as supervised fine-tuning data; scores are validated by heads-up gameplay between checkpoints.
Key results
- GPT-4 is the best pre-trained model with 53.55% exact-match and 65.54% action accuracy on PokerBench in few-shot evaluation.
- Fine-tuned Llama-3-8B reaches 78.26% EM and 80.64% AA, outperforming GPT-4 on the benchmark.
- In 1,000 heads-up hands, the fine-tuned Llama-3-8B lost to GPT-4 overall (GPT-4 won at 22.20 bb/100), attributed to GPT-4's donking and aggression that the GTO-based training data does not cover.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| fine-tune Llama-3-8B on PokerBench training data | Llama-3-8B · PokerBench overall | EM: 26.02 -> 78.26 | Helped |
| fine-tune Llama-3-8B on PokerBench training data | Llama-3-8B · PokerBench overall | AA: 40.03 -> 80.64 | Helped |
| fine-tune Llama-2-7B on PokerBench training data | Llama-2-7B · PokerBench overall | EM | Helped |
| fine-tune Gemma-2B on PokerBench training data | Gemma-2B · PokerBench overall | EM | Helped |
| fine-tune Llama-3-8B on GTO-labeled PokerBench data | Fine-tuned Llama-3-8B · Heads-up vs GPT-4 | win rate: 22.20 bb/100 | Hurt |
| increase PokerBench checkpoint score from 61.32 AA to 69.60 AA | Llama-3-8B checkpoints · Heads-up vs lower-scoring checkpoint | win rate: 34.99 bb/100 | Helped |
| increase PokerBench checkpoint score from 69.60 AA to 80.64 AA | Llama-3-8B checkpoints · Heads-up vs lower-scoring checkpoint | win rate: 24.79 bb/100 | Helped |
| use balanced action-distribution resampling for post-flop training data | PokerBench fine-tuning · PokerBench post-flop | fine-tuning performance | Helped |
| raise a wider range of hands pre-flop | Llama-3-8B 5000 vs GPT-4 · Heads-up pre-flop | win rate: 17.80 bb/100 | Helped |
| use supervised GTO fine-tuning without adapting to donk bets | Fine-tuned Llama-3-8B · Heads-up vs GPT-4 flop and turn | win rate: 34.97 bb/100 | Hurt |
Surprising findings
- Despite higher benchmark accuracy, the fine-tuned Llama-3-8B was outplayed by GPT-4 in heads-up play.
- GPT-4 won 100% of hands in which it donked (out-of-position bets), a move rare in GTO training data.
Limitations
- Fine-tuned model underperforms GPT-4 in actual heads-up play despite higher benchmark scores, indicating limits of simple SFT for strategy.
- Benchmark assumes GTO play by both players, so it under-represents exploitative opponent behavior such as donking.
- Heads-up gameplay sample (1,000 hands) is small for statistical significance.
Experimental setup
Datasets: PokerBench training set (60k pre-flop, 500k post-flop spots) GTOWizard strategies (pre-flop) WASM-Postflop solver outputs (post-flop)
Benchmarks: PokerBench
Metrics: action accuracy exact match accuracy win rate (bb/100)
Baselines: GPT-4 ChatGPT 3.5 Llama-3-70B Llama-2-70B Gemma-2B
Builds on: WASM-Postflop GTOWizard
Closest papers by meaning
- Can Large Language Models Master Complex Card Games? (NeurIPS 2025 · similarity 0.90)
- Mastering Board Games by External and Internal Planning with Language Models (ICML 2025 · similarity 0.89)
- GTBench: Uncovering the Strategic Reasoning Capabilities of LLMs via Game-Theoretic Evaluations (NeurIPS 2024 · similarity 0.87)
- Outbidding and Outbluffing Elite Humans: Mastering Liar’s Poker via Self-Play and Reinforcement Learning (NeurIPS 2026 · similarity 0.86)
- Large Language Models Play StarCraft II:Benchmarks and A Chain of Summarization Approach (NeurIPS 2024 · similarity 0.86)
- Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs (NeurIPS 2025 · similarity 0.85)
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.