LLMs Achieve Expert-Level Poker Without Training or Solvers, New Study Shows

The Claim: Expert-Level Poker via Pure Prompting

A new paper posted on arXiv on Friday, May 29, 2026, makes a striking claim: large language models can play heads-up no-limit Texas hold'em at expert level without any specialized training, fine-tuning, or access to game-theoretic solvers. The paper, titled "PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers" (arXiv:2605.30094), describes a prompting strategy that enables LLMs to reason about incomplete information and deception — abilities long considered outside the reach of simple language model inference.

According to the paper, authored by Boning Li, Baoxiang Wang, and Longbo Huang, the approach relies on structured prompting that guides the LLM through reasoning steps such as opponent modeling, hand range estimation, and pot odds calculation. The authors report that their method achieves performance statistically indistinguishable from top human professionals and significantly outperforms prior AI systems like DeepStack and Pluribus in certain settings, despite those systems requiring months of self-play training or explicit counterfactual regret minimization solvers.

The paper comprises 45 pages and includes three figures illustrating the prompting pipeline and experimental results.

How PokerSkill Works: Prompting In-Context Reasoning

The core insight of PokerSkill is that LLMs, when prompted appropriately, can carry out the complex multi-step reasoning required for poker without any gradient updates or external game solvers. The authors design a set of chain-of-thought prompts that decompose each decision into sub-problems: estimating the opponent's likely hand range based on betting patterns, evaluating the strength of one's own hand relative to that range, and calculating expected value of possible actions.

Specifically, the system provides the LLM with the current game state, including position, pot size, betting history, and community cards, and then guides it through a series of reasoning steps. The prompts are engineered to avoid common pitfalls such as anchor bias or failure to consider bluffing strategies. The LLM outputs both a recommended action (fold, call, raise) and a justification, which allows for debugging and validation of its reasoning.

The authors tested their method across thousands of hands against multiple baselines, including the open-source solver "OpenSpiel" and state-of-the-art RL-based agents. They found that PokerSkill achieved a win rate exceeding 55% against these baselines over large samples, and in human evaluation sessions with professional poker players, the LLM's play was indistinguishable from expert-level human decision-making.

Significance for AI Research: Reasoning Under Uncertainty

The results have broader implications beyond games. Poker is a canonical example of an imperfect-information game where optimal play requires reasoning about hidden information, deceptive behavior, and probabilistic outcomes. The fact that an LLM can perform this reasoning purely through in-context learning suggests that current models possess latent capabilities for strategic reasoning that have not been fully exploited.

This contradicts a common narrative that LLMs are merely "stochastic parrots" that cannot engage in genuine reasoning. The paper provides evidence that, under the right prompting conditions, LLMs can simulate reasoning processes analogous to those used by humans in strategic decision-making. This opens up possibilities for applying LLM-based reasoning to other domains with incomplete information, such as negotiation, bidding, portfolio management, and cybersecurity threat assessment.

Moreover, because the method requires no training or solver components, it is immediately deployable with any existing LLM API. This democratizes access to high-quality decision-making in complex environments — a domain previously dominated by specialized AI systems that were expensive to build and maintain.

Limitations and Open Questions

Despite the impressive results, the paper also acknowledges several limitations. First, the experiments were conducted primarily in heads-up (two-player) no-limit Texas hold'em. Multi-player games, which involve more complex social dynamics and coalition formation, remain largely untested. Second, the LLM's performance is heavily dependent on prompt engineering; small changes in prompt phrasing can lead to significant performance drops. The authors note that their prompts are specifically tailored to the particular LLM architecture used (in their experiments, GPT-4 and Claude 3.5 models), and transferability to other models is not guaranteed.

Third, the method incurs significant computational overhead due to the lengthy chain-of-thought reasoning. Each decision may require generating hundreds or thousands of tokens, leading to latency that could be problematic in real-time games. The authors report an average reasoning time of 2–3 seconds per action on modern hardware, which is borderline acceptable for online poker but would be too slow for live play.

Finally, there is the question of generalization: does the LLM truly understand game theory, or is it pattern-matching from the poker content in its training data? The paper argues that because the LLM generalizes to novel situations not explicitly in its training set, genuine reasoning is occurring. However, the lack of interpretability tools makes it difficult to fully rule out memorization-based strategies.

Implications for Real-World Applications

The PokerSkill paper represents a notable advance in demonstrating that LLMs can serve as general-purpose reasoning engines for strategic decision-making under uncertainty. If these results hold up to scrutiny, they could reshape how we approach autonomous agents in domains such as finance, logistics, and security — areas where actions must be taken with incomplete information and adversarial participants.

For the AI community, the key takeaway is that current LLMs are far more capable than typical benchmarks suggest, but unlocking these capabilities requires careful prompt design. The paper also adds evidence to the growing body of work showing that scaled-up inference-time compute (i.e., longer chain-of-thought reasoning) can compensate for lack of training in many tasks.

As with any breakthrough, independent replication is critical. The authors have made their code and evaluation data available on GitHub, allowing other researchers to verify their findings. Whether the poker world will soon see LLM-based opponents at professional tables remains an open question, but the technical feasibility now seems plausible. For now, PokerSkill serves as a provocative demonstration that the next generation of AI reasoning may not require bespoke training — just a well-crafted conversation.