October 2025

A focused poker player making a decision at the table, illustrating the challenge of LLMs in poker AI.

Poker is much more than just a game. It’s a proving ground for advanced machine reasoning, especially for evaluating the capabilities of LLMs in poker AI.

These features are problems for AI algorithms based on classical machine learning principles because all these features use the ability to think in time without time, bluff, and adapt. This is the reason we use poker as a yardstick to see how well an algorithm can perform in a game in which there is uncertainty and ambiguity.

Why Solvers Are Great Problem Solvers, But Not Instinct

The classical solvers are great at devising a balanced strategy where the aim is to discover a strategy that will prevent your opponents from being able to improve on the expected value (EV) in the long run at the tables. but they are not so good at making quick adjustments out of the emotions or instinct.

Human players on the other hand are great at making adjustments according to what is happening in the course of the game, and they tend to think instinctively or out of emotion in adjusting their strategies.

Why ToolPoker Was Invented for LLMs in Poker AI

As LLMs in poker AI continue to face the knowing-doing gap, new approaches are needed. The researchers invented ToolPoker to try to combine the best of both solvers (balance and non-exploitable) and the players (instinct) and adaptability.

The researchers did an experiment to compare the performance of six models at two games. The first game is a simplified version of poker called Leduc Hold’em. The second game is a version of Limit Texas Hold’em which presents more realistic poker games. Six different models (GPT-4, Qwen 2.5, LLaMA 3, o4-mini, and the baseline solvers: CFR+, NFSP) were given the same parameters and stakes and were judged on their performance on three criteria:

HR (Hand Rationality): How much does the model reason like a professional player?
FA (Final Action): Did the model make the GTO (game theoretic optimal) move?
AC (Average Chips): Did the model win?Why the Researchers Wanted to Test the LLMs Initially Before Making the Decisions

In order to understand how well the models would do the researchers had each model explain why it wanted to make a particular decision prior to making that decision. The results were astounding; many of the models gave perfect reasons for their decisions but then proceeded to make the opposite decision.

This phenomenon has been called the “knowing-doing gap” that is the difference between knowing something on an intellectual level and doing something on a practical level.

Three Major Ways LLMs Still Fail in Poker

While the researchers have improved LLMs with reinforcement learning (RL) training, they still possess great flaws as poker players. Heuristics-Based Logic

LLMs tend to rely on familiar linguistic heuristics. For example, a model may decide to bet aggressively in a situation because it has been exposed to examples of situations in which aggressive betting is successful even though the current situation may be very different. Poor Understanding of Mathematical Nuances

LLMs seem to understand mathematical concepts like pot odds and implied odds. But when it comes to employing those calculations in practice, they tend to be inconsistent.

Mot that an LLM will utilize a logical process to evaluate a decision to make, but that does not mean that it will execute that decision in the way that it ought to . This is due to the fact that the LLM makes a distinction between “the logical thought process it took to come to that decision” and “the actions it takes which flow logically from that decision.”

Continue reading →

Poker as a Psychological Arena

Scattered poker chips conveying the unpredictable bluffing in Leduc Hold'em by AI

Poker is not merely a game; it is a psychological arena dressed as a card table. Each raise and/or fold is a motion in a silent drama of partial truths. That is the reason why researchers studying artificial intelligence are interested in poker, as it provides an environment that includes logic, un-predictability and strategy.

Using the Leduc Hold’em format of poker, the 2025 study, titled “Analysis of Bluffing by DQN and CFR in Leduc Hold’em Poker,” stripped poker to its bare essentials in order to determine if systems based upon math and feedback can develop deceptions without first learning to deceive.

Spoiler alert: they did.

Bluffing in Leduc Hold’em by AI: DQN vs CFR

CFR is the control freak. It analyzes each game decision in reverse, in order to find out what it should have done and adjust accordingly until the regret is eliminated. Eventually, CFR develops a level of balance in its play such that no player can take advantage of it. That is the theoretical concept behind CFR.

DQN is the experiential learner. It makes a guess, attempts the guess, fails, attempts again. It is not attempting to achieve perfection – it is attempting to achieve reward. It finds what works and focuses on that.

Neither CFR nor DQN had pre-programmed bluffs or pre-defined tactics. Both simply began with cards, rewards and logic.

Although neither CFR nor DQN were programmed to bluff, both developed bluffing strategies. These results are not merely impressive – they also reveal characteristics.

Experimental Design of Bluffing in Leduc Hold’em by AI

Two artificial intelligent agents. Equal stacks. Equal blinds. Each agent has one private card and one public card. Two betting rounds. No noise. No multi-player interaction. Simply, clear, distilled decision-making.

Each decision was recorded and analyzed for hand strength, bet size, and the response of the opponent. The objective of the experiment was to record the frequency of bluffing (a weak hand, a large bet), the effectiveness of bluffing, and the development of the bluffing behavior of each agent.