How Far Are LLMs from Professional Poker Players? Revisiting Game-Theoretic Reasoning with Agentic Tool Use

As Large Language Models (LLMs) are increasingly applied in high-stakes domains, their ability to reason strategically under uncertainty becomes critical. Poker provides a rigorous testbed, requiring not only strong actions but also principled, game-theoretic reasoning. In this paper, we conduct a systematic study of LLMs in multiple realistic poker tasks, evaluating both gameplay outcomes and reasoning traces. Our analysis reveals LLMs fail to compete against traditional algorithms and identifies three recurring flaws: reliance on heuristics, factual misunderstandings, and a “knowing-doing” gap where actions diverge from reasoning. An initial attempt with behavior cloning and step-level reinforcement learning improves reasoning style but remains insufficient for accurate game-theoretic play. Motivated by these limitations, we propose ToolPoker, a tool-integrated reasoning framework that combines external solvers for GTO-consistent actions with more precise professional-style explanations. Experiments demonstrate that ToolPoker achieves state-of-the-art gameplay while producing reasoning traces that closely reflect game-theoretic principles.

💡 Research Summary

This paper presents a comprehensive investigation into the capabilities of large language models (LLMs) when applied to professional‑level poker, a canonical imperfect‑information game that demands rigorous game‑theoretic reasoning. The authors first conduct a systematic empirical study across two widely used poker variants—Leduc Hold’em and Limit Texas Hold’em—comparing a broad spectrum of open‑source and closed‑source LLMs (ranging from 3 B to 72 B parameters, as well as GPT‑4.1‑mini, GPT‑4o, and o4‑mini) against established poker AI baselines such as NFSP, DQN, DMC, CFR+, and DeepCFR. The evaluation protocol involves 50 games per seed with position swapping, measuring net chip gain as the primary performance metric, and also collecting the LLMs’ natural‑language reasoning traces for each decision.

The results reveal a stark performance gap: virtually all vanilla LLMs underperform the equilibrium‑focused CFR+ and DeepCFR algorithms, with only the largest closed‑source models occasionally beating non‑equilibrium baselines. A deeper qualitative and quantitative analysis of the reasoning traces uncovers three recurring flaws. First, heuristic reasoning dominates; models rely on surface‑level patterns (“bet strong hands, fold weak hands”) rather than explicit equity or range calculations. Second, factual misunderstandings occur, where the model misestimates hand strength or pot odds, leading to systematically flawed decisions. Third, a knowing‑doing gap emerges: even when the model articulates correct strategic reasoning, its final action often diverges (e.g., reasoning to fold but raising instead). To quantify these issues, the authors introduce an LLM‑as‑a‑Judge framework with three metrics—heuristic reasoning (HR), factual alignment (FA), and action‑reasoning consistency (AC)—scored on a 0‑2 scale. Professional‑level reasoning scores a perfect 2 on all metrics, while LLMs range from 0.5 to 1.8, confirming the severity of the identified flaws.

Attempting to remediate these deficiencies, the authors experiment with a two‑stage internal training pipeline: behavior cloning on expert traces followed by step‑level reinforcement learning (RL) with reward shaping. This approach improves the fluency and apparent reasoning style of the models but fails to achieve GTO‑consistent actions; the underlying game‑theoretic deficiencies persist.

Motivated by these findings, the paper introduces ToolPoker, a novel tool‑integrated reasoning (TIR) framework that equips LLMs with the ability to invoke external poker solvers during inference. The key innovations are: (i) a unified API that aggregates multiple solver functionalities (action selection, equity calculation, range analysis) into a single call, reducing error propagation and simplifying training; (ii) a compact expert‑style reasoning dataset, constructed by emulating professional thought processes and automatically augmenting each entry with standardized tool‑invocation templates and the corresponding solver outputs. This dataset serves as high‑quality supervision for both the behavior‑cloning stage and the subsequent RL fine‑tuning.

Empirical evaluation shows that ToolPoker dramatically narrows the performance gap. Across both Leduc Hold’em and Limit Texas Hold’em, ToolPoker achieves average chip gains of +28 and +85 respectively, outperforming all vanilla LLM baselines by 2–3× and matching or surpassing CFR+ in several scenarios. Moreover, the reasoning metrics improve to near‑expert levels (HR, FA, AC scores around 1.8–1.9), indicating that the model’s explanations now align closely with game‑theoretic principles. The unified tool interface also reduces data‑collection costs by roughly 60 % compared with naïve multi‑tool pipelines.

In conclusion, the study demonstrates that while LLMs excel at natural‑language generation, they lack the intrinsic capacity to perform the precise, mathematically grounded calculations required for optimal poker play. By integrating external solvers through a carefully designed tool‑use framework, ToolPoker leverages the strengths of both LLMs (flexible reasoning and explanation) and specialized algorithms (guaranteed GTO actions). This work opens a promising direction for deploying LLMs in high‑stakes decision‑making domains where rigorous game‑theoretic reasoning is essential, and suggests future research avenues such as scaling to multi‑player settings, optimizing tool‑call latency, and extending the approach to other imperfect‑information games.