Learning strategic poker decision-making with Large Language Models

Abstract

Decision-making in the poker variant No-Limit Texas Hold’em involves integrating categorical, numerical, and sequential information in an imperfect-information environment. These elements can be represented as structured text, enabling the investigation of Large Language Models (LLMs) as predictors of strategic decisions. This dissertation examines that possibility in 6-max No-Limit Texas Hold’em, an imperfect-information setting in which each decision involves an action and, in aggressive cases, a bet or raise sizing. The task is formulated as supervised prediction over textualized states, separating its categorical and numerical components. A hybrid pipeline selects actions through log-probability scoring and obtains sizings through deterministic generation followed by numerical parsing, enabling separate analysis of action selection and numerical parameterization. PokerBench examples are evaluated in preflop and postflop settings. Six open LLMs are compared under few-shot prompting, and the best-performing model, Qwen3-14B, is specialized through supervised fine-tuning with LoRA and QLoRA for each game stage. In addition to Action Accuracy, a continuous formulation of Actionand-Sizing Accuracy (Ac-s) is developed to preserve action-first semantics while assigning proportional credit to sizing predictions when the aggressive action is correct. Few-shot prompting provides an informative baseline but remains below supervised adaptation. After fine-tuning, Qwen3-14B achieves 93.3% Action Accuracy preflop and 91.8% postflop, corresponding to gains of 17.2 and 40.3 percentage points, respectively. Analyses show consistent improvements across actions and contexts, while conditional diagnostics indicate high proportional sizing agreement in both stages after the aggressive action is correctly identified. The dissertation therefore contributes an operational formulation, a continuous metric, a reproducible evaluation pipeline, and an empirical analysis of open LLMs for structured strategic decision-making in an imperfect-information domain.