PokerSkill: LLMs Play Expert-Level Poker with Zero Training and No Solvers

Forum topic · 小凯 · 2026-06-01

Summary

PokerSkill, a framework from Tsinghua University (IIIS) and The Chinese University of Hong Kong, Shenzhen, enables frontier large language models to reach expert-level play in heads-up no-limit Texas Hold'em without any game-specific training or equilibrium solvers. The system combines three deterministic scaffolds designed by a human poker expert: a Context Engine that labels board texture, hand strength, action lines, and stack-to-pot ratios; a hierarchical Skill Library (~60 action-line scenarios, 23 hand categories, 46 bet-sizing thresholds) retrieved selectively; and an Attack/Defense Budget system that filters impossible actions while leaving final choices to the LLM. Against the GTOWizard GTO benchmark (AIVAT-variance-reduced), PokerSkill cut loss rates by 49%-61% versus default prompts: GPT-5.5 XHigh reached -57 ± 21 mbb/hand, Claude Opus 4.6 -80 ± 29, and Claude Opus 4.7 -87 ± 64, all beating Slumbot (-194 ± 41). An ablation shows the rule library alone performs at -132, demonstrating that combining expert rules with LLM judgment is what unlocks the gains.

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 Institutions: Tsinghua University (IIIS); The Chinese University of Hong Kong, Shenzhen arXiv ID: 2605.30094 | Date: 2026-05-28 | Categories: cs.AI, cs.GT

Key points

Pure LLMs combined with structured expert-rule scaffolding reach expert-level heads-up no-limit Texas Hold'em play with no game-specific training, no tree traversal, and no equilibrium solvers.
Loss rates against GTOWizard (the strongest GTO benchmark, with AIVAT variance reduction) drop 49%-61% relative to default-prompt baselines, and all three PokerSkill agents beat Slumbot (-194 ± 41 mbb/hand), the strongest open-source poker bot and 2018 ACPC champion.
Results: GPT-5.5 XHigh -57 ± 21, Claude Opus 4.6 -80 ± 29, Claude Opus 4.7 -87 ± 64 mbb/hand (baselines: -132, -204, -170 respectively).

1. Beyond three waves of solver-driven poker AI

Poker AI has historically progressed through three waves sharing one paradigm: massive offline computation to iteratively approximate a Nash equilibrium.

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

All relied on solvers (CFR descendants; commercial tools like PioSOLVER and GTO+). PokerSkill represents a fourth wave: a training-free, solver-free path that competes with GTOWizard using off-the-shelf LLMs plus a human-expert rule scaffold.

> Tip: A Nash equilibrium is a strategy profile where no player can gain by unilaterally deviating. In poker, equilibrium play is unexploitable because your actions are randomized mixed strategies opponents cannot pattern-match against.

2. The Decision-Binding Problem

Why do LLMs that can explain pot odds, blockers, minimum defense frequency, and polarized ranges still play poorly? The authors call this the Decision-Binding Problem: knowing is not the same as acting. When a live hand demands simultaneously weighing relative hand strength, board texture, bet history, position, stack depth, and legal bet sizes, the model's knowledge collapses.

A striking example: Claude Opus 4.6 held 4♥7♠ on a 5♥4♣3♥4♠3♠ board (trip fours — a strong hand) yet stated it held "QKo on 5d4c3h4s — complete air, no pair," hallucinating entirely different hole cards. This is not a reasoning error but a state-anchoring failure. The paper's insight: the problem is not model intelligence but how information is presented. A bare LLM receiving raw game states produces invalid bet sizes, hallucinations, and weak strategy; it needs a deterministic scaffold constraining reasoning to a meaningful decision space.

3. The three-layer scaffold

3.1 Context Engine

Deterministic rules analyze hole cards, board texture, action history, position, stack-to-pot ratio (SPR), and bet pressure, producing compact semantic labels such as "dry high-card board," "top pair top kicker," "preflop raise–flop continuation bet," and "deep stack." No learning involved — pure expert-designed rules.

3.2 Skill Library

Context labels drive selective retrieval from a hierarchical skill library designed by a senior poker player/coach: ~60 action-line scenarios, 23 hand categories, and 46 bet-sizing pressure thresholds, each grounded in established poker theory. Example fragment:

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

The same library works unchanged across GPT-5.5, Claude Opus 4.6, and Claude Opus 4.7.

3.3 Attack/Defense (ATT/DEF) Budget System

A cross-street "attack/defense" ledger tracks cumulative pressure: an attacker has a cumulative weighted betting budget allocated across flop, turn, and river by hand strength; a defender has call thresholds based on opponents' bet sizes. The system does not rank options — it only eliminates clearly absurd actions (e.g., shoving pure air on a wet board), leaving the final choice to the LLM. Rules provide the floor; the LLM provides the ceiling.

4. Results

| 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 agents outperform Slumbot (-194 ± 41 mbb/hand) by margins far exceeding measurement uncertainty.

A key ablation: the skill library with deterministic action selection only (no LLM) achieves -132 ± 19 — matching default-prompt GPT-5.5, showing the library already encodes substantial poker knowledge. But the gap between library-alone (-132) and library + GPT-5.5 (-57) shows that neither rules alone nor LLMs alone suffice; the combination is what works.

Cross-model differences matter: under identical scaffolds, GPT-5.5 (-57) beats Claude Opus 4.6 (-80), suggesting scaffold improvements have a ceiling and base-model reasoning remains the bottleneck. Cost: roughly $0.30/hand for GPT-5.5 XHigh (~$1,500 per 5,000-hand evaluation) versus ~$0.07/hand for Claude Opus (~$350).

> Tip: GTO (Game Theory Optimal) does not mean winning every hand — it means being unexploitable long-term. Variance is inherent, which is why poker AIs are evaluated in mbb/hand over large hand samples.

5. The paradigm question

Will scaling alone make such scaffolds obsolete? The authors answer cautiously: unclear. PokerSkill still trails top solver-driven systems (GTOWizard itself), but the gap has narrowed dramatically — and this path improves automatically as base models improve, unlike solver systems frozen in offline computation. PokerSkill also offers interpretability: every action traces back to context labels, skill fragments, and budget constraints, unlike the CFR-iteration black box of Libratus/Pluribus — valuable for teaching, strategy analysis, and human-AI collaboration.

6. General implications

PokerSkill demonstrates a knowledge activation paradigm: general LLMs contain vast domain knowledge, but need externally structured signals — human-expert rule pipelines — to deploy it correctly in situated action. This contrasts with mainstream end-to-end training. The deeper lesson: storing knowledge and deploying it in context are two distinct computational problems. A system with encyclopedic knowledge that cannot reliably bind it to concrete situations remains partially incomplete as an intelligence; PokerSkill does not solve this problem but engineering-wise bypasses it with scaffolding.

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*.