PokerSkill: LLMs Reach Expert-Level Poker Without Training or Solvers

Forum topic · 小凯 · 2026-06-01

Summary

Researchers from Tsinghua University (IIIS) and The Chinese University of Hong Kong, Shenzhen introduce PokerSkill, a training-free, solver-free framework that lets frontier LLMs play expert-level heads-up no-limit Texas Hold'em. Rather than iterating toward Nash equilibrium like CFR-based systems (DeepStack, Libratus, Pluribus), PokerSkill combines a deterministic rule-based Context Engine, a hierarchical expert-designed Skill Library (~60 betting lines, 23 hand categories, 46 sizing thresholds), and an attack/defense budget system that filters illegal or absurd actions. The LLM then makes the final choice within this constrained space. Against GTOWizard, the strongest GTO benchmark, losses per hand dropped 49%-61% versus default prompting: GPT-5.5 XHigh achieved -57 ± 21 mbb/hand (from -132), Claude Opus 4.6 -80 ± 29 (from -204), and Claude Opus 4.7 -87 ± 64 (from -170). All three outperformed Slumbot (-194 ± 41), the former ACPC champion. The authors identify a 'Decision-Binding Problem'—LLMs possess poker knowledge but fail to bind it to situated action—and argue structured human-prior scaffolding offers an interpretable, training-free alternative that improves automatically as base models scale.

PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers

Paper: PokerSkill: LLMs Can Play Expert-Level Poker without Training or Solvers Authors: Boning Li, Baoxiang Wang, Longbo Huang Affiliations: Tsinghua University (IIIS), The Chinese University of Hong Kong, Shenzhen arXiv ID: 2605.30094 | Date: 2026-05-28 | Categories: cs.AI, cs.GT

Key finding: A pure LLM combined with structured expert-rule scaffolding reaches expert level in heads-up no-limit Texas Hold'em without any game-specific training or equilibrium solvers, cutting loss rates by 49%–61% versus raw prompting and beating the historically strong bot Slumbot.

Beyond Three Waves of Poker AI

Poker has long been a touchstone for AI, testing deception, belief modeling, and incomplete-information reasoning. Since Zinkevich et al. introduced Counterfactual Regret Minimization (CFR) in 2007, three waves have defined the field:

1. DeepStack (University of Alberta, 2017): deep neural networks estimating subgame values with real-time solving at each decision point. 2. Libratus (CMU, 2018): blueprint strategy plus subgame re-solving; defeated four top human professionals over 120,000 hands. 3. Pluribus (2019): extended to six-player poker via Monte Carlo CFR with limited search depth.

All three share one paradigm: iterating toward Nash equilibrium with millions of core-hours of offline computation. Solvers (PioSOLVER, GTO+, GTOWizard) remain the professional standard.

PokerSkill offers a fourth path: no training, no solver—just an off-the-shelf LLM plus a human-expert-designed rule scaffold—competing with GTOWizard, the current strongest GTO benchmark.

The Decision-Binding Problem

Why do LLMs fail at the table despite having read virtually all poker strategy literature? The authors name this the Decision-Binding Problem: knowing pot odds, blockers, and polarized ranges is different from deploying that knowledge under real-game pressure.

A striking example: Claude Opus 4.6 held 4♥7♠ on a 5♥4♣3♥4♠3♠ board (trip fours, a strong hand) yet hallucinated "I hold QKo on 5d4c3h4s—complete air, no pair." This is not a reasoning error but a failure of state anchoring—the model misread its own hand. The insight: the problem is not model intelligence but how information is presented. Bare LLMs receiving raw game states produce invalid sizes, hallucinations, and weak strategies; a deterministic scaffold is needed to constrain reasoning into a meaningful decision space.

The Three-Layer Scaffold

1. Context Engine

Deterministic rules analyze the current state—hole cards, board texture, action history, position, SPR, cumulative betting pressure—and output compact semantic labels (e.g., "dry high-card board," "top set," "flush-draw board," "deep stack"). No learning involved.

2. Skill Library

Context labels trigger selective retrieval from a hierarchical skill library designed by a senior poker player versed in GTO theory: roughly 60 betting-line scenarios, 23 hand categories, and 46 sizing-pressure thresholds, each grounded in established theory. A sample fragment:

> "With top pair top kicker on a dry flop as the preflop raiser, make a small continuation bet (about 25–33% pot), value-focused with protection secondary. Facing a raise, usually call once unless the board is extremely wet."

The same library works across three different LLMs without modification.

3. Attack/Defense (ATT/DEF) Budget System

A cross-street ledger tracks cumulative pressure: attackers get a weighted betting allowance allocated across flop, turn, and river by hand strength; defenders get call thresholds based on opponent sizing. The system does not rank options—it only eliminates absurd ones (e.g., shoving pure air on a wet board). The LLM makes the final choice. Rules provide the floor; the LLM provides the ceiling.

Results: From −204 to −80

All evaluations were run against GTOWizard with AIVAT variance reduction:

| Model | Default-prompt baseline | PokerSkill | Improvement | |---|---|---|---| | GPT-5.5 XHigh | −132 ± 25 | −57 ± 21 | 57% | | Claude Opus 4.6 | −204 ± 44 | −80 ± 29 | 61% | | Claude Opus 4.7 | −170 ± 28 | −87 ± 64 | 49% |

All three outperformed Slumbot (−194 ± 41), the 2018 ACPC champion and strongest pre-2025 open poker bot, by margins far exceeding measurement uncertainty. This is the first time a pure LLM has reached this level in an imperfect-information game without game-tree traversal, iterative training, or solver queries.

A revealing ablation: the Skill Library plus deterministic action selection (no LLM) achieves −132 ± 19 mbb/hand—on par with default-prompted GPT-5.5. Rules alone are not a strong strategy; LLMs alone cannot play well; together they work remarkably well.

Cost: roughly $0.30/hand for GPT-5.5 XHigh (~$1,500 for 5,000 hands) and ~$0.07/hand for Claude models (~$350).

The Paradigm Question

The authors are candid: whether such scaffolds become obsolete as base models scale remains unclear. PokerSkill still trails top solver systems, but the gap has narrowed dramatically—and its performance improves automatically with base-model capability, unlike solver systems fixed to their offline computation.

Another dimension is interpretability: Libratus and Pluribus are near black boxes born of millions of CFR iterations, while every PokerSkill action traces back to context labels, skill fragments, and budget constraints—an auditable pipeline valuable for teaching, strategy analysis, and human-AI collaboration.

Broader Implications

PokerSkill demonstrates a knowledge activation paradigm: general LLMs embed vast domain knowledge, but correct invocation requires external structured signals encoded by human experts—not training. In poker: read the situation → recall principles → decide. Elsewhere: read the image → recall pathology → diagnose.

The deeper question: if LLMs already implicitly approximate GTO strategy, why is activation alone so fragile? The Decision-Binding Problem suggests that knowledge storage and situated deployment are two distinct computational problems—separated by a world. PokerSkill does not solve this; it engineers around it. That workaround is itself a form of engineering wisdom.

References

1. Brown & Sandholm (2018). "Superhuman AI for heads-up no-limit poker: Libratus beats top professionals." *Science*, 359(6374):418-424. 2. Brown & Sandholm (2019). "Superhuman AI for multiplayer poker." *Science*, 365(6456):885-890. 3. Moravčík et al. (2017). "DeepStack: Expert-level artificial intelligence in heads-up no-limit poker." *Science*, 356(6337):508-513. 4. Bowling et al. (2015). "Heads-up limit hold'em poker is solved." *Science*, 347(6218):145-149. 5. Zinkevich et al. (2007). "Regret Minimization in Games with Incomplete Information." *NIPS*.