PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers

Researchers introduce PokerSkill, a framework that enables large language models to play expert-level poker without training or computational solvers by combining rule-based poker skills with LLM reasoning. The approach achieves competitive performance against state-of-the-art GTO benchmarks, reducing losses by 49-61% compared to standard LLM prompting and outperforming established poker bots.

PokerSkill represents a significant breakthrough in applying large language models to complex strategic domains traditionally dominated by specialized AI systems. The framework cleverly sidesteps the computational burden of equilibrium solvers—which require millions of core-hours—by using human-expert-designed poker skills as structured constraints that guide LLM decision-making. This hybrid approach demonstrates that LLMs possess latent strategic knowledge that becomes actionable when grounded in domain-specific rule systems, rather than relying on raw parametric learning.

Historically, poker AI evolved through two distinct paths: rule-based systems offering interpretability but limited performance, and solver-based agents achieving near-equilibrium play at enormous computational cost. PokerSkill bridges this gap by treating the LLM as a strategic reasoner rather than a game-theoretic optimizer. The deterministic context engine acts as a filter, presenting only relevant poker skills to the model for each game state, preventing hallucination and keeping decisions within reasonable bounds. This modular architecture reflects broader trends in AI toward hybrid systems that combine symbolic reasoning with neural capabilities.

n

The competitive results against GTOWizard and Slumbot carry implications beyond poker. The framework's success suggests LLMs can acquire and apply domain expertise more efficiently when paired with human-curated knowledge structures. For AI researchers, this validates the value of grounding large models in structured skill libraries rather than pursuing pure end-to-end learning. The solver-free nature also democratizes access to expert-level performance, removing the infrastructure barrier that previously required massive computational resources.

Future developments may explore whether PokerSkill's hybrid approach scales to other imperfect-information games or complex strategic domains, potentially reshaping how specialized AI systems are designed across finance, negotiation, and military applications.

• →PokerSkill combines rule-based poker skills with LLM reasoning to achieve expert-level play without training or computational solvers.
• →The framework reduces losses against GTO benchmarks by 49-61% compared to standard LLM approaches and outperforms established poker bots.
• →Structured skill grounding prevents LLM hallucination and constrains decisions to strategically sound actions in imperfect-information games.
• →The hybrid architecture demonstrates LLMs can leverage latent domain knowledge when guided by human-expert-designed constraints rather than pure parametric learning.
• →This approach removes computational barriers to expert AI performance and may generalize beyond poker to other strategic domains.