Poker has long served as the definitive benchmark for AI strategic reasoning — it was the domain where DeepStack and Libratus first demonstrated superhuman performance against human professionals. Now, as Dong Tran's AI research platform tracks the frontier of large language model capabilities, this ICLR 2026 paper from Minhua Lin and colleagues delivers a sobering and rigorous answer to a question the field has been dancing around: can LLMs actually reason strategically under uncertainty, or are they sophisticated pattern-matchers that break down when faced with adversarial game theory?
The answer, delivered with systematic rigor across multiple poker variants and evaluation dimensions, is a clear no — at least not natively. And the reasons why reveal something deeply important about the architecture of current LLMs.
The paper's most valuable contribution isn't the benchmark itself — it's the taxonomy of failure modes. The authors identify three recurring patterns across all tested LLMs:
This knowing-doing gap is not a minor curiosity. For AI agents being deployed in high-stakes domains — financial trading, medical diagnosis, legal reasoning — the possibility that an LLM's explanation of its reasoning doesn't actually reflect its decision process is a fundamental safety concern that this paper surfaces in a rigorous, measurable way.
The paper follows a clean research arc. First, comprehensive benchmarking across "multiple realistic poker tasks" — the authors evaluate both gameplay outcomes (win rates, chip counts) and reasoning traces (alignment between stated strategy and executed actions). This dual evaluation is methodologically important: prior work often measured only outcomes, missing the knowing-doing gap entirely.
The authors then attempt standard remediation approaches: behavior cloning from professional poker hands, and step-level reinforcement learning with rewards tied to game-theoretic correctness. The results are instructive — both approaches improve the style of reasoning traces (models produce more poker-appropriate language) without achieving accurate game-theoretic play. You can teach a model to sound like Phil Ivey without teaching it to play like Phil Ivey.
This negative result motivates ToolPoker, the paper's core technical contribution. The framework is conceptually elegant: rather than trying to force LLMs to internalize GTO strategies through training, ToolPoker gives the LLM access to external GTO solvers as tools. The LLM's role shifts from computing optimal actions to interpreting game state, querying the solver, and translating the mathematical output into principled natural language explanations.
ToolPoker achieves state-of-the-art gameplay — competitive with traditional algorithmic approaches — while producing reasoning traces that actually reflect game-theoretic principles. This is the knowing-doing gap resolved through architecture rather than training. The model's explanation and action now align because the action is derived from the solver output that also grounds the explanation.
For the AI research community tracking agentic AI development, this result has a clear message: tool augmentation isn't just a performance optimization, it's a mechanism for achieving genuine reasoning-action alignment. From Dong Tran's AI research platform perspective, this is a crucial insight for anyone designing LLM-based agents for domains requiring principled decision-making.
The broader implications extend well beyond poker. Any domain where optimal decisions can be computed algorithmically — operations research, quantitative finance, formal verification, combinatorial optimization — is a candidate for the ToolPoker paradigm. The LLM provides natural language interface, context interpretation, and explanation generation; specialized solvers provide the actual optimal actions.
This represents a maturation in how we think about LLM agency. The early vision was LLMs as universal reasoners that could handle any domain natively. The emerging reality, which this paper documents rigorously, is that hybrid architectures — LLMs paired with domain-specific tools and solvers — will outperform pure LLM approaches in any domain with a well-defined notion of optimality. This isn't a failure of LLMs; it's a clarification of their proper role in agentic systems.
The knowing-doing gap finding also has direct implications for AI safety. If LLM reasoning traces don't reliably reflect the processes driving LLM actions, then interpretability research that relies on analyzing model outputs to understand model behavior is on shakier ground than assumed. This paper provides empirical evidence for a concern that has been largely theoretical.
The paper's scope is intentionally narrow — poker provides clean ground truth (GTO solutions exist), but that clarity doesn't transfer directly to messier real-world domains. The knowing-doing gap magnitude may vary significantly across domains, and the paper doesn't provide tools for measuring it outside game-theoretic contexts.
ToolPoker's success also depends on having a reliable external solver — a prerequisite that doesn't exist for most interesting real-world reasoning problems. The paper implicitly raises the question: when no GTO solver exists, how do we close the knowing-doing gap? That's the research agenda this paper opens rather than resolves.
Finally, the behavior cloning and RL experiments deserve more ablation. The negative results are valuable, but understanding why these approaches fail to close the knowing-doing gap — whether it's a capacity issue, a training signal issue, or a fundamental architectural constraint — would significantly advance our understanding of LLM reasoning limitations.
ICLR 2026 acceptance signals that the community recognizes this paper's importance. For AI researchers and practitioners following Dong Tran's work on agentic AI systems, ToolPoker is required reading — not because poker matters, but because the knowing-doing gap matters everywhere, and this paper gives us the clearest empirical window into it yet.