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

Researchers introduced PokerSkill, a framework that helps LLMs play poker using a deterministic context engine and a layered library of human-written skills.

Against the GTOWizard benchmark, frontier models with PokerSkill reduced losses versus default prompting and even beat Slumbot in tests.

The approach needs no game-specific training and no solver queries, instead grounding actions in expert-designed poker knowledge.

The result suggests a cheaper path to strong imperfect-information game agents without retraining or solver access.