"PokerBench: Training Large Language Models to become Professional Poker Players"

POKERBENCH provides a new benchmark for teaching LLMs optimal poker strategies, with 11,000 carefully curated scenarios testing complex decision-making abilities in uncertain conditions.

https://arxiv.org/abs/2501.08328

Methods in this Paper 💡:

→ POKERBENCH introduces a comprehensive benchmark with 11,000 poker scenarios split between pre-flop and post-flop play.

→ The scenarios are carefully filtered from billions of possible game states using board texture analysis and optimal action probability thresholds.

→ The benchmark evaluates both action accuracy (fold/call/raise decisions) and exact match accuracy (precise bet sizing).

→ POKERBENCH is validated through extensive gameplay testing between models with different benchmark scores.

-----

Key Insights 🔑:

→ All current LLMs significantly underperform at poker compared to their capabilities in other domains

→ Fine-tuning on POKERBENCH dramatically improves poker performance

→ Higher POKERBENCH scores correlate strongly with better gameplay results

→ Simple supervised learning alone may be insufficient for optimal poker strategy

-----

Results 📊:

→ GPT-4: 53.55% overall accuracy (best among pre-trained models)

→ Fine-tuned Llama-3-8B: 78.26% accuracy, outperforming GPT-4

→ Win rate correlation: Models with higher POKERBENCH scores consistently beat lower-scoring models