PokerSkill lets LLMs play expert-level poker without training

A new framework called PokerSkill, from researchers at the Chinese University of Hong Kong, shows that large language models (LLMs) can play expert-level poker without any game-specific training or equilibrium solvers. Traditional AI poker agents rely on counterfactual regret minimization (CFR) solvers that require millions of core-hours of training. While LLMs already possess extensive poker knowledge, their raw performance lags far behind solver-based bots. PokerSkill bridges this gap by using a structured rule-based skill library—designed entirely by human experts—to ground the LLM's actions. A deterministic context engine analyzes the current game state and retrieves only the relevant skill fragments, constraining the LLM to reasonable, expert-level decisions.

Against GTOWizard, a state-of-the-art GTO benchmark, GPT-5.5 XHigh with PokerSkill achieved -57 ± 21 mbb/hand, while Claude Opus 4.6 scored -80 ± 29 mbb/hand and Claude Opus 4.7 scored -87 ± 64 mbb/hand. These results represent a 49–61% reduction in losses compared to default-prompt baselines, and they outperformed Slumbot, a strong poker bot. The key insight is that neither rule-based skills nor LLMs alone produce a strong strategy, but their combination yields an agent that requires neither training nor solver queries. This marks the first time an LLM has achieved competitive performance in a complex imperfect-information game without game-specific training, suggesting that similar hybrid approaches could crack other strategic domains.