acceptodds
Under review as a conference paper at ICLR 2027

PokerSkill: A Strong LLM Agent in Poker without Training or Solvers

Abstract

Poker is a landmark challenge for artificial intelligence. The dominant approach relies on equilibrium solvers built on counterfactual regret minimization, requiring millions of core-hours of training. Large Language Models (LLMs) possess extensive poker knowledge but perform far below solver-based agents when asked to play directly. Traditional rule-based poker agents are interpretable and training-free, but their strategic ceiling remains far below equilibrium play. We introduce **PokerSkill**, a training-free and solver-free framework that bridges this gap by using detailed rule-based poker skills as a structured action-grounding interface for LLMs. A deterministic context engine analyzes the current state and retrieves only the relevant fragments from a layered skill library, which is entirely designed by human poker experts, constraining the LLM's choice to reasonable actions. Against GTOWizard, a state-of-the-art GTO benchmark, GPT-5.5 XHigh with PokerSkill achieves mbb/hand, Claude Opus 4.6 achieves mbb/hand and Claude Opus 4.7 achieves mbb/hand, reducing losses by 49–61% compared to default-prompt baselines and outperforming the strong bot Slumbot. Our key finding is that rule-based skills alone do not constitute a strong strategy, and LLMs alone cannot play well, but their combination yields an agent that requires neither training nor solver access yet competes with systems built on millions of core-hours of computation. To our knowledge, this is the first demonstration of an LLM achieving competitive performance in a complex imperfect-information game without game-specific training or solver queries.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Poker AI, LLM Agent, Structured Prompting, Skill, Imperfect-Information Games

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_MPs2OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The authors introduce Pokerskill, a harness built from expert knowledge that extracts decision features for retrieval from a expert written skill library and filters the possible actions for each turn. They evaluate their system on five frontier LLMs against GTOWizard, claiming state of the art results for open-source poker AI systems, and examine how expert knowledge was used to create their system, claiming a generalizable pattern.

Critical strengths

None

Critical weaknesses

To be blunt, I really think this isn't a main conference level submission. While the authors establish their system is able to meaningfully increase LLM performance in HUNL there isn't nearly enough investigation into why (Table 2 is good, but the bare score alone isn't enough to show differences in behavior). The baselines are almost weak by construction. A bare prompt and a naive greedy algorithm are not sufficient to prove the strength of this approach. There are two critical baselines I consider to be missing: 1.) a version without the retrieval where the entirety of the library and everything is provided upfront to the model 2.) A GEPA style loop where the model is able to augment its own prompt and create its own harness (ADAS). As it stands, there is minimal evidence for their claim of a 'transferable design pattern', which would otherwise be the strongest claimed contribution.

Questions for authors

1.) You claim a 'transferable design pattern' but the skill library layer seems extremely bespoke for HUNL. Could you provide some examples of how this design might transfer to other domains?

2.) Could the authors provide more details about the rule-based baselines?

6 comments

We thank the reviewer for the frank assessment and for naming the two missing baselines. Both have now been run, along with the attribution study requested by Reviewer vaQd, and we address each point below. The full table of new results is in the general response.

Weaknesses

To be blunt, I really think this isn't a main conference level submission.

Poker has been a milestone problem for imperfect-information game AI, and every strong HUNL agent so far has paid for its strength with training: solver-based systems (DeepStack, Libratus, Slumbot, GTOWizard) require millions of core-hours of CFR, while LLM-based systems such as ToolPoker rely on a solver queried at inference time, which itself must be trained. PokerSkill is the first strong HUNL agent that trains nothing and queries no solver, pairing an unmodified general-purpose LLM with an expert-written skill library. Under these constraints, it loses 44–87 mbb/hand to GTOWizard across five frontier models (Table 1). For comparison, the top human HUNL professionals lost 147 mbb/hand to Libratus (Brown & Sandholm, 2018), so PokerSkill's loss against a near-equilibrium solver is lower than that of the strongest human specialists. A training-free, solver-free agent at this level is a new result.

Two further points highlight the value of the result. First, to our knowledge, PokerSkill is both the strongest reported LLM poker agent and the strongest open, deployable HUNL agent: Slumbot, GTOWizard, Libratus, and Pluribus have not been released as runnable systems. Second, an agent at this level has uses for which near-equilibrium solvers are poorly suited. It is strong enough to be a sparring partner for human players and is a useful evaluation opponent for other poker AIs: the strong public opponents available today are solver-based and so close to equilibrium that they measure a single quantity, whereas an agent that plays well but not at equilibrium exposes a different, complementary set of weaknesses.

While the authors establish their system is able to meaningfully increase LLM performance in HUNL there isn't nearly enough investigation into why (Table 2 is good, but the bare score alone isn't enough to show differences in behavior).

We agree that the score alone does not show how play changes, and we have added two types of evidence.

Mechanism. The context engine converts the raw state into the labels by which the library is indexed (hand class, board texture, action line, pot type, SPR, and position); retrieval provides the model with only the entries written for that situation; and the budget mechanism leaves it with a short list of viable actions. The LLM's role is to reason over that retrieved expert knowledge, read the opponent's line, and choose among the viable actions, including the bet size. The new attribution table isolates these components: structured labels alone do nothing (−152 vs. −153 for the default prompt), retrieved strategy text is the largest single factor (−101), the budget on its own is insufficient (−139), and only their combination reaches −73.

Behaviour. We compared the postflop action frequencies and bet sizes of the default-prompted model, PokerSkill, and the rule agents with GTOWizard's own play. Frequencies are in percent (fold/check/call/bet-or-raise); size is the median bet as a fraction of the pot. This is Table 7 of the revision.

AgentFlop f/k/c/b (%)sizeTurn f/k/c/b (%)sizeRiver f/k/c/b (%)size
GTOWizard8/48/18/260.339/54/14/220.7512/43/11/340.67
Default prompt10/52/15/230.5014/52/9/250.6720/51/11/180.67
PokerSkill10/41/16/330.2511/56/9/240.6619/42/11/280.75
Rule agent, expert-ranked11/36/14/380.2510/52/7/320.3720/46/10/230.66
Rule agent, random9/41/15/350.2511/58/8/230.3520/43/10/270.35

The default-prompted model bets half pot on every street and bets the river only 18% of the time (GTOWizard: 34%). PokerSkill moves flop betting to a third of the pot and river betting to three quarters, toward the solver's size ladder, and bets the river 28% of the time. The rule agents share the same options but over-bet the flop (38%) and use a third of the pot on the turn and river, where the solver uses three quarters; the differences between the rule agents and PokerSkill, which reflect the LLM's contribution, lie in when to bet and how much.

The baselines are almost weak by construction. A bare prompt and a naive greedy algorithm are not sufficient to prove the strength of this approach.

We have added five non-LLM and matched-LLM controls, listed here in full (all 10,000 AIVAT hands against GTOWizard with DeepSeek-V41-Flash, the successor of the retired V4-Flash; mean ± 1 SE, mbb/hand; columns mark which parts of the prompt the configuration receives: preflop play from the P2 table, context-engine labels, retrieved P3–P5 text, the whole library in the system prompt, attack/defense budget, viable-option line).

ConfigurationTableLabelsRetrievalWhole lib.BudgetOptionsmbb/hand
PokerSkill✅✅✅❌✅✅−73 ± 21
PokerSkill, hard enforcement of the viable set✅✅✅❌✅✅ (hard)−71 ± 23
PokerSkill, no viable-options line✅✅✅❌✅❌−84 ± 23
Default prompt❌❌❌❌❌❌−153 ± 16
Default prompt + preflop table✅❌❌❌❌❌−139 ± 26
(i) Labels only✅✅❌❌❌❌−152 ± 20
(ii) Labels + retrieved text✅✅✅❌❌❌−101 ± 27
(iii) Labels + budget + options✅✅❌❌✅✅−139 ± 33
Whole library, raw state❌❌❌✅❌❌−165 ± 32
Whole library + engine✅✅❌✅✅✅−118 ± 27

Rule agents run the same pipeline up to the viable-option list and then use a fixed selector instead of the model:

Selector over the viable optionsmbb/hand
First option (paper, Table 1)−132 ± 19
Expert-ranked option−120 ± 18
Uniformly random option−161 ± 21

What the two tables show. The structured state by itself changes nothing: the model plays as poorly with the engine's labels as with the raw cards (−152 versus −153), so the gain does not come from a better description of the state. Retrieved strategy text is the largest single ingredient (−101), but text alone leaves the model free to overcommit; the budget and viable options alone (−139) remove that freedom but leave it nothing to reason with; together they reach −73, more than the two effects add up to. Retrieval also beats exposure: the whole library in the context, with the engine kept, costs 45 mbb/hand (−118), and without the engine it is the worst configuration measured (−165), so the model cannot pick the applicable entries out of the library by itself. No fixed selector over the same options comes close to the model (−120 expert-ranked, −132 first, −161 random), and hard enforcement of the options changes fewer than 1% of the model's actions and 2 mbb/hand of the result: the model stays inside the option set of its own accord, and its choice within that set is where the strength lies. The expert-ranked and random rule agents are added to Table 1 of the revision, the LLM configurations form Table 3, and the appendix defines all three selectors.

There are two critical baselines I consider to be missing: 1.) a version without the retrieval where the entirety of the library and everything is provided upfront to the model

Done, in two versions. Putting the entire library (P1–P5, about 14k tokens) into the system prompt with only the raw game state scores −165 ± 32, the worst configuration we measured. Keeping the context engine, budget, and viable options while replacing the retrieved fragments with the whole library scores −118 ± 27, compared to −73 ± 21 with retrieval. Selecting the applicable entries is worth about 45 mbb/hand, and the whole library without the context engine is worse than no library.

2.) A GEPA style loop where the model is able to augment its own prompt and create its own harness (ADAS).

We ran this before submission. Using the same base model and evaluator, we tested (a) a reflector loop in which the LLM reads the hand logs and per-situation results and rewrites one part of the library at a time (RIVER-style one-factor mutation with an LLM judge); (b) self-consistency voting at multi-option decisions; (c) an empirical action-frequency prior mined from the agent's own showdown logs and bound as a majority action; (d) an empirical preflop table mined in the same way; (e) big-pot discipline rules mined from the largest losses; and (f) opponent-model exploitation rules mined from 400,000 hands against Slumbot. None of these improved upon the hand-written library. The two reflector rewrites were indistinguishable from the original (+11 ± 63 and +36 ± 74 mbb/hand); the four log-mined upgrades yielded +12 ± 38, +16 ± 39, +2 ± 35, and −33 ± 28; and the one rule that appeared positive against Slumbot in an arm-level comparison (+19.5 ± 5.4 at the decision level) fell to +1.8 ± 10.8 under an in-spot randomized control.

The reason is structural. Poker is an imperfect-information game with a large luck component. In our logs the raw outcome of a hand has a standard deviation of about 15 bb, so a candidate's win rate is known to ±150 mbb/hand after 10,000 hands and ±47 after 100,000; even a deliberately damaged library (with the top-pair advice inverted) was only 1.1 standard errors from the original in a 1,000-hand comparison. Separating two candidates that differ by 10 mbb/hand, the size of the smaller effects in the attribution table, takes about 2×1072 \times 10^{7} hands per candidate, and a search over fifty candidates about 10910^{9} hands, with each hand costing several model calls; every "improvement" a loop accepts within an affordable budget is indistinguishable from a lucky run. The loop also inherits the limits of the model running it: rewrites are only as good as the rewriter, so weaker models, which have the most to gain, are the least able to improve themselves. This is why the library is written once by an expert and then fixed: once written down, knowledge does not have to be selected amid the noise, and the same unchanged library serves every model.

As it stands, there is minimal evidence for their claim of a 'transferable design pattern', which would otherwise be the strongest claimed contribution.

The reviewer is right that we present no second domain, and the revised paper frames this pattern as an outlook. What the paper does establish is that a general LLM, given a deterministic context engine, layered retrieval of expert skills, and a bounded action set, succeeds in a large imperfect-information game with a heavy luck component, which earlier LLM-agent work had not shown; the transfer question is addressed in Q1 below.

Reviewer_vaQdOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces PokerSkill, a training-free and inference-time-solver-free agent for heads-up no-limit Texas hold'em. A deterministic context engine converts cards, board texture, action history, position, stack-to-pot ratio, and betting pressure into structured labels. These labels retrieve relevant portions of a five-layer, expert-authored skill library, while an attack/defense budget proposes a small set of strategically viable actions. Expert range tables determine most preflop play; postflop, an otherwise unmodified LLM chooses among the prompted options, after which a validator guarantees a legal action.

Critical strengths
  • The empirical improvement is large, consistent, and measured across meaningful settings. The same library improves all five tested models against GTOWizard, with displayed independent-run contrasts of approximately 2.3–3.2 standard errors for four models; the arithmetic of the reported 49–74% loss reduction checks out. The separate 400,000-hand DeepSeek–Slumbot run also shows a large improvement over the same model’s default prompt (+172 ± 34 mbb/hand, as reported in Section 4). Although these results do not prove parity with either opponent, they establish a practically important improvement over generic prompting.
Critical weaknesses
  • The evaluation does not isolate the claimed source of the improvement, and the non-LLM control is too weak to support the rules-versus-LLM narrative. Default prompting versus PokerSkill simultaneously changes deterministic state extraction, five prompt layers, the action-budget mechanism, the preflop controller, and the list of suggested actions. In eleven preflop families, PokerSkill samples from a hand-authored mixed strategy while the default model makes a language-model decision, so this is not a matched postflop-prompt comparison (Section 3.2, p. 5; Appendix C.2, pp. 15–16). The “rule-based” baseline simply chooses the first viable action (Section 4, p. 8; Appendix E.2, p. 26), with no comparison to random, frequency-weighted, or expert-priority selection and no report of how often there is only one viable action.
  • The expert interface is not yet sufficiently reliable or reproducible for the strength of the strategic claims. The hand trace’s geometric sizing is mathematically inconsistent with matched called bets: for initial pot P, remaining effective stack S, and n equal pot-fraction bets, the usual relation is �0��0�, hence �1��1�. For the trace’s flop values (P=18, S=191, n=3), this gives about 90.6% pot, not the displayed 126%; for the turn (P=28, S=186, n=2), it gives about 139%, not 176% (Appendix G, pp. 32 and 36).
Questions for authors
  1. Can the authors provide the exact implementation used for the reported runs (or a complete source-equivalent specification and version hash), including all preflop tables, all 47 pressure thresholds, scenario/override precedence, sizing code, validator substitutions, and the exact prompts?
  2. Please reconcile the uncertainty and count discrepancies: Table 1 versus Table 3 for Claude Fable 5, and Figure 4’s 10,000-hand-per-bar caption versus Appendix E.2’s 5,000-hand default control.
  3. A stronger attribution study should hold the preflop controller fixed and compare (i) a bare postflop LLM with the same structured state, (ii) context retrieval without budgets, (iii) budgets without retrieved strategic text, and (iv) competent deterministic selectors (including random/frequency-weighted and expert-ranked choices) with hard versus soft viable-action enforcement.
3 comments

We thank the reviewer for a precise review. We address each point in order.

Weaknesses

The evaluation does not isolate the claimed source of the improvement, and the non-LLM control is too weak to support the rules-versus-LLM narrative. Default prompting versus PokerSkill simultaneously changes deterministic state extraction, five prompt layers, the action-budget mechanism, the preflop controller, and the list of suggested actions.

We agree that the default-versus-PokerSkill contrast bundles these changes, and we have run the study the reviewer outlined in Q3, holding the preflop controller fixed and adding one ingredient at a time. All configurations use 10,000 AIVAT hands against GTOWizard, the same library, and the same model (DeepSeek-V41-Flash; since the paper's V4-Flash has been retired, the full agent was re-run on V41-Flash as the reference row). Results are reported as mean ± 1 SE in mbb/hand.

ConfigurationTableLabelsRetrievalBudgetOptionsmbb/hand
Default prompt❌❌❌❌❌−153 ± 16
Default prompt + preflop table✅❌❌❌❌−139 ± 26
(i) bare postflop LLM with the structured state✅✅❌❌❌−152 ± 20
(ii) retrieval without budgets✅✅✅❌❌−101 ± 27
(iii) budgets without retrieved text✅✅❌✅✅−139 ± 33
PokerSkill (soft enforcement, reference)✅✅✅✅✅−73 ± 21
PokerSkill, hard enforcement of the viable set✅✅✅✅✅ (hard)−71 ± 23
PokerSkill without the viable-options line✅✅✅✅❌−84 ± 23

Reading the table from top to bottom: the preflop controller on its own is worth about 14 mbb/hand (−153 → −139); adding the structured state to the postflop prompt adds nothing by itself (−152); retrieved strategy text is the largest single ingredient (−101); budgets without text are not enough (−139); and the combination reaches −73. The improvement comes from the model reasoning over the retrieved expert text within the bounded action set.

In eleven preflop families, PokerSkill samples from a hand-authored mixed strategy while the default model makes a language-model decision, so this is not a matched postflop-prompt comparison (Section 3.2, p. 5; Appendix C.2, pp. 15–16).

Correct, and the "default prompt + preflop table" configuration now provides the matched comparison: with identical table-driven preflop play, the default postflop prompt scores −139 ± 26, while PokerSkill scores −73 ± 21. We also logged the routing: when the table is present, 100% of preflop decisions are taken from it in every configuration, meaning all rows in the table above except the first share the exact same preflop policy.

The "rule-based" baseline simply chooses the first viable action (Section 4, p. 8; Appendix E.2, p. 26), with no comparison to random, frequency-weighted, or expert-priority selection and no report of how often there is only one viable action.

We added the selectors the reviewer requested. All rule agents run the same pipeline up to the viable-option list and differ only in the selector:

Selector (no LLM)AIVAT (mbb/hand)
First viable action (paper)−132 ± 19
Expert-ranked: the option the rules themselves prefer ("lean" marker, else the option whose unlocking threshold is closest below the remaining budget)−120 ± 18
Uniformly random viable action−161 ± 21
LLM choosing from the same list (PokerSkill)−73 ± 21

Frequency-weighted selection is not defined postflop because the library carries action frequencies only in the preflop tables (which all rule agents already sample from), and the postflop option list has no frequencies attached. The best deterministic selector is 47 mbb/hand behind the LLM, and a local best response (reported to Reviewer iSkm) exploits the expert-ranked and random agents by the same amount (+7.4 ± 1.3 and +8.0 ± 1.2 BB/hand); thus, the strength of the full agent comes from the LLM's choice. The two added rule agents appear in Table 1 of the revision next to the paper's first-option rule agent; the appendix defines the three selectors. On the frequency of a single viable action: 71% of all decisions in the logged runs have exactly one viable option, and the mean size of the viable set is 1.4.

The expert interface is not yet sufficiently reliable or reproducible for the strength of the strategic claims. The hand trace's geometric sizing is mathematically inconsistent with matched called bets: for initial pot P, remaining effective stack S, and n equal pot-fraction bets, the usual relation is S=P[(1+2f)^n −1]/2, hence f=((1+2S/P)^{1/n}−1)/2. For the trace's flop values (P=18, S=191, n=3), this gives about 90.6% pot, not the displayed 126%; for the turn (P=28, S=186, n=2), it gives about 139%, not 176% (Appendix G, pp. 32 and 36).

The reviewer is right, and we thank them for the derivation. The sizing code computed the fraction from (1+f)n=1+S/P(1+f)^n = 1 + S/P, i.e., it grew the pot by the hero's bet only and omitted the opponent's call. We have corrected it to f=((1+2S/P)1/n−1)/2f = \big((1 + 2S/P)^{1/n} - 1\big)/2, which is exactly the reviewer's formula; on the two quoted cases, the corrected code gives 90.6% and 139%, and these two cases are now unit tests. The consequences are as follows: (1) every new run in this rebuttal uses the corrected code; (2) the example trace in Appendix G has been regenerated (it was produced from the corrected code on a new hand and checked street by street, and its flop and turn geometric sizes are 86% and 133%).


Questions

Q1: Can the authors provide the exact implementation used for the reported runs (or a complete source-equivalent specification and version hash), including all preflop tables, all 47 pressure thresholds, scenario/override precedence, sizing code, validator substitutions, and the exact prompts?

The anonymized supplementary material already contains the runnable agent used for the reported runs. We cannot attach a meaningful version hash to "the reported runs", for two reasons. First, in response to W2, the sizing code changed between the submitted runs and the new ones, so no single hash covers both. Second, during review we can only upload an anonymized package that differs from our working repository, so a hash of the package would not identify the code that produced the runs either; we therefore do not think a hash is the right form of verification here. As stated in the paper, the full source code, including the corrected sizing code, will be released upon acceptance.

Q2: Please reconcile the uncertainty and count discrepancies: Table 1 versus Table 3 for Claude Fable 5, and Figure 4's 10,000-hand-per-bar caption versus Appendix E.2's 5,000-hand default control.

Both are reporting inconsistencies, and both are fixed. Table 3 showed a 95% confidence half-width (±48) for Claude Fable 5 whereas Table 1 showed 1 SE (±25); the revision reports 1 SE everywhere. The 5,000-hand default-prompt run is the control of Table 1 (every default-prompt configuration there plays at least 5,000 hands); the ablation runs use 10,000 hands each, and the figure caption no longer states a per-bar count.

Q3: A stronger attribution study should hold the preflop controller fixed and compare (i) a bare postflop LLM with the same structured state, (ii) context retrieval without budgets, (iii) budgets without retrieved strategic text, and (iv) competent deterministic selectors (including random/frequency-weighted and expert-ranked choices) with hard versus soft viable-action enforcement.

Done: (i)–(iii) are in the first table above, and (iv) is in the second. On hard versus soft enforcement: the paper's validator enforces legality only, while membership in the viable set is soft. We added a hard variant in which an action outside the viable set is mapped to the nearest viable one (bet/raise → call → check → fold, and the reverse for passive actions). It scores −71 ± 23 compared to −73 ± 21 for the soft version, and the hard validator changed the model's action in fewer than 1% of decisions (180 of 25,030); the model almost always follows the viable set of its own accord, so soft and hard enforcement yield empirically the same agent. The attribution table is Table 3 of the revision and the three rule agents are in Table 1.


We are grateful for a review that told us exactly which experiment to run; the paper's central claim is now backed by the attribution study the reviewer requested, and the sizing error has been fixed. We hope the attribution study and the corrections address the concerns raised; if they do, we would be grateful if the reviewer would consider revising the score. We would also be glad to address any further concern during the discussion period.

Dear Reviewer,

Thanks for your time and effort in reviewing our paper.

We hope our response has adequately addressed your concerns. If you feel that our rebuttal has clarified the issues raised, we kindly ask you to consider adjusting your score accordingly. Should you have any further questions or need additional clarification, we would be more than happy to discuss them with you.

Thanks once again for your valuable feedback.

Best regards,

Authors

Reviewer_iSkmOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The main idea of the paper is to guide a generic LLM agent how to play Head’s Up No Limit (HUNL) Texas Hold-em poker using a expert-written skill library at inference time. The context labels from a deterministic engine retrieve relevant parts of a 5 layer skill library. The attack / defence options narrows down the action space to a few viable options at each turn for the LLM to choose from. Against the GTO Wizard benchmark, a single fixed library reduces the losses of five frontier LLMs by 49–74%, to between −44 and −87 mbb/hand. Over 400,000 hands against Slumbot, the DeepSeek-V4-Flash agent scores −31 ± 23 mbb/hand, against −203 ± 25 without the library.

Critical strengths

S1: The problem is very well motivated that even frontier models fail on the GTOWizard benchmark (Provost et al., 2026) for various reasons and how even though Poken knowledge is present in these models, they are not reliably converted into actions and how at without training, at inference time we can improve the performance of these models by a skill library and the rules to extract relevant information from it. S2. A well designed skill-library with all possible actions thought of and made into skill labels. Given that this is developed by a deterministic context engine makes it easy to evaluate as well. The retrieval strategy of the relevant skills into layered skill libraries makes it a bit narrower as well. With the attack/defence the model again has a very few options to choose from as the capacity becom

Critical weaknesses

W1: The opponents in the game never adapt / change their strategy. This doesn’t resemble a real world player. So qualitatively we can’t say anything about the exploitable nature of PokerSkill’s rules since no LBR is reported. W2. The authors claim that PokerSkill competes with systems built on millions of core-hours,” and Section 5 calls it “the first to show” a strong HUNL agent. Yet against Slumbot it scores −31 ± 23 mbb/hand (95% CI - [−76, +14]), while GTO Wizard’s AI (+194 ± 41, an earlier Slumbot release) and ReBeL (+45 ± 5) have published wins against Slumbot. W3. The instructions regarding “bluff” in P5 which uses the overall bluff to portion of the range to calculate the bluff-to-value ratio contradicts the one in P3. The example shown in Appendix G contains state misclassifications as well.

Questions for authors
  1. Can you report LBR, or a learned or approximate best response, at least for the deterministic rule-only agent?
  2. Was the library edited after observing any GTO Wizard or Slumbot result?
  3. What exactly does “w/o P2” remove, given that preflop table sampling bypasses the LLM?
3 comments

We thank the reviewer for the careful summary and precise questions; both the LBR request and the P3/P5 inconsistency led to concrete additions and corrections.


W1: The opponents in the game never adapt / change their strategy. This doesn't resemble a real world player. So qualitatively we can't say anything about the exploitable nature of PokerSkill's rules since no LBR is reported.

We now report a local best response (LBR; Lisý & Bowling, 2017) against the deterministic rule component, which is the fixed part of the system that can be exploited exactly. The setting is standard: a fixed preflop line (BTN opens 2.5 BB, BB calls, with the BB's range given by the P2 table); the exploiter acting in position with the LBR action set {fold, check/call, bet ½, 1, 2 × pot, all-in} and check-calling afterwards; the rule agent's hand averaged over the range consistent with the history; and turn and river cards averaged within each maximisation. Values represent the exploiter's win in BB/hand (positive = the rule agent leaks) over 5 seeds × 8 exploiter hands each:

Rule agent (no LLM)LBR (BB/hand)
Expert-ranked viable action+7.4 ± 1.3
Random viable action+8.0 ± 1.2

For scale, Lisý & Bowling (2017) report that LBR with the same action set wins up to 4.0 ± 0.1 BB/hand against Slumbot 2016 and 3.3 ± 0.1 against Act1, the two ACPC finalists, so a deterministic rule agent being more exploitable than those bots is the expected outcome. Both rule agents are exploitable by a similar amount, and the choice of selector does not matter. The rule component only produces the viable-option set; the LLM chooses among those options and sets the bet size after reading the hand, the board, and the opponent's line, and that choice is what carries the agent from the rule-agent level (−120 to −161 mbb/hand against GTOWizard) to −73. The new attribution table in the general response makes the same point from the other side: every configuration that removes the LLM's choice or the retrieved knowledge it reasons over falls back toward the rule-agent level.

An LBR against the full agent is not feasible: the LLM's policy is stochastic and available only through API calls, and an LBR requires the opponent's action at every candidate node for every hand in the range (about 1,000 combinations × 6 exploiter actions per decision), amounting to millions of model calls per evaluated hand.


W2. The authors claim that PokerSkill competes with systems built on millions of core-hours," and Section 5 calls it "the first to show" a strong HUNL agent. Yet against Slumbot it scores −31 ± 23 mbb/hand (95% CI - [−76, +14]), while GTO Wizard's AI (+194 ± 41, an earlier Slumbot release) and ReBeL (+45 ± 5) have published wins against Slumbot.

We have removed both phrases from the abstract and Section 5, and the revised text makes no claim of parity with solver-based systems. Two facts put the Slumbot number in context. First, the 400,000-hand Slumbot run uses DeepSeek-V4-Flash, the weakest of the five models we tested and the only one cheap enough for that volume, so it does not reflect the strength of the method. Second, against GTOWizard, the same model scores −82 ± 17, while Claude Fable 5 with the same library scores −44 ± 25 (Table 1), a gap of about 40 mbb/hand. We therefore expect a frontier model with PokerSkill to perform considerably better against Slumbot than the −31 reported, and because the library is fixed and the model is unmodified, the agent improves automatically as frontier models advance.


W3. The instructions regarding "bluff" in P5 which uses the overall bluff to portion of the range to calculate the bluff-to-value ratio contradicts the one in P3.

Thank you for catching this; you are right. P3 states that the value-to-bluff ratio for a pot-sized river bet is 2:1, which is the indifference ratio (the bluff share of a polarized betting range is bet/(pot+2 bet)=1/3\mathrm{bet}/(\mathrm{pot}+2\,\mathrm{bet}) = 1/3 for a pot-sized bet). The P5 line "bet_size/(pot+bet_size) of range should be bluffs" gives 1/21/2 for the same bet, which is the opponent's fold-to-break-even frequency rather than the bluff share, contradicting P3. In the revision, the P5 line reads betsize/(pot+2 betsize)\mathrm{bet_size}/(\mathrm{pot}+2\,\mathrm{bet_size}), consistent with P3. P5 is a low-priority reference block attached only on the river, whereas P3's 2:1 rule is the one the prompt instructs the model to apply first; the correction changes one line of text.

The example shown in Appendix G contains state misclassifications as well.

The example trace in Appendix G has been replaced. It has been regenerated from the corrected code (the geometric-sizing correction described in our reply to Reviewer vaQd), and we checked every label street by street against the cards: on the flop (A♣6♣ on J♠2♠A♠), the engine reports top pair with a 6 kicker on a three-flush board; on the turn (A♦), it reports trips on a paired, three-flush board; and the geometric sizes (86% of the pot with three streets left at SPR 9.5; 133% with two streets at SPR 6.2) follow the corrected formula, with the model's turn bet using the 133% size. If the reviewer has a specific label from the old trace in mind, we would be glad to address it directly.


Q1: Can you report LBR, or a learned or approximate best response, at least for the deterministic rule-only agent?

Yes; see W1 above (+7.4 ± 1.3 BB/hand against the expert-ranked rule agent and +8.0 ± 1.2 against the random one).


Q2: Was the library edited after observing any GTO Wizard or Slumbot result?

No. The library was written by an author familiar with GTO strategy, debugged by playing against the code directly (checking that the prompt conveyed what the author intended in each situation), and frozen before the evaluation runs. The same frozen library is used for all five models in Table 1, the Slumbot run, the ablation, and every new configuration in this rebuttal; no entry was changed in response to a GTOWizard or Slumbot result.


Q3: What exactly does "w/o P2" remove, given that preflop table sampling bypasses the LLM?

The "w/o P2" setting removes the preflop range-table layer entirely: no preflop decision is sampled from the tables, and every preflop decision is made by the LLM using the P1 prompt and the raw state (postflop layers P3–P5 and the budget remain unchanged). With P2 present, 100% of preflop decisions go through the table in every logged configuration, so the ablation measures the difference between table-driven and LLM-driven preflop play with identical postflop play. We have added this definition to the caption of Table 3. The new "default prompt + preflop table" configuration (−139 ± 26 vs. −153 ± 16 for the default prompt) measures the same component from the other direction.


We are grateful for the review; the LBR, the P5 correction, and the clarified ablation all improve the paper. We hope these additions address the concerns raised; if they do, we would be grateful if the reviewer would consider revising the score. We would also be glad to address any further concern during the discussion period.

Dear Reviewer,

Thanks for your time and effort in reviewing our paper.

We hope our response has adequately addressed your concerns. If you feel that our rebuttal has clarified the issues raised, we kindly ask you to consider adjusting your score accordingly. Should you have any further questions or need additional clarification, we would be more than happy to discuss them with you.

Thanks once again for your valuable feedback.

Best regards,

Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to All Reviewers and the Area Chair

We thank all three reviewers for their careful reading and concrete suggestions. The reviews converge on two requests: a finer attribution of where PokerSkill's improvement comes from, using stronger non-LLM controls (vaQd, Mps2), and evidence showing how the agent's play changes (Mps2) as well as how exploitable the rule component is (iSkm). We have run the requested experiments and revised the paper accordingly. This post collects the new results and lists the changes; the per-reviewer replies refer back to it.

New experiments

All new configurations play 10,000 AIVAT hands against GTOWizard using the same library, the same prompt builder (with the geometric-sizing correction described in our reply to Reviewer vaQd), and the same base model. DeepSeek-V4-Flash, the model used in the paper, has been retired by the provider; the new runs therefore use its successor, DeepSeek-V41-Flash, on which we also re-ran the full PokerSkill agent as the reference row. Numbers are mean ± 1 SE in mbb/hand (higher is better). Columns: preflop play from the P2 table; context-engine labels; retrieved P3–P5 text; the whole library in the system prompt; attack/defense budget; viable-option line.

ConfigurationTableLabelsRetrievalWhole lib.BudgetOptionsmbb/hand
PokerSkill✅✅✅❌✅✅−73 ± 21
PokerSkill, hard enforcement of the viable set✅✅✅❌✅✅ (hard)−71 ± 23
PokerSkill, no viable-options line✅✅✅❌✅❌−84 ± 23
Default prompt❌❌❌❌❌❌−153 ± 16
Default prompt + preflop table✅❌❌❌❌❌−139 ± 26
(i) Labels only✅✅❌❌❌❌−152 ± 20
(ii) Labels + retrieved text✅✅✅❌❌❌−101 ± 27
(iii) Labels + budget + options✅✅❌❌✅✅−139 ± 33
Whole library, raw state❌❌❌✅❌❌−165 ± 32
Whole library + engine✅✅❌✅✅✅−118 ± 27

Rule agents (no LLM): the same pipeline up to the viable-option list, then a fixed selector instead of the model.

Selector over the viable optionsmbb/hand
First option (paper, Table 1)−132 ± 19
Expert-ranked option−120 ± 18
Uniformly random option−161 ± 21

We also report (in the reply to Reviewer iSkm) a local best response against the deterministic rule agents, and (in the reply to Reviewer Mps2) a behaviour profile of default prompting, PokerSkill and the rule agents against GTOWizard's revealed play, together with the outcome of our own attempts at self-improving the library.

Summary of changes in the revised PDF (changed words are marked in red)

  1. Section 4: A new attribution table (the LLM configurations above) is included with a paragraph of analysis; the expert-ranked and random rule agents are added to Table 1 next to the paper's first-option rule agent, and the appendix defines the three selectors.
  2. Appendix: A behaviour profile of default prompting, PokerSkill, and the rule agents against GTOWizard's revealed play (action frequencies and bet sizes by street) is added, together with the configuration definitions, the selector definitions, and the local best response; one sentence in Section 4 points to it.
  3. Geometric sizing: The formula in the context engine omitted the opponent's call; it is corrected in the code, and the example trace in Appendix G has been regenerated from the corrected code and checked street by street.
  4. P5 (river bluff reference): The bluff-frequency line is corrected to bet/(pot+2 bet)\mathrm{bet}/(\mathrm{pot}+2\,\mathrm{bet}), consistent with the 2:12{:}1 value-to-bluff ratio stated in P3.
  5. All uncertainties are now reported as 1 SE (Table 3 previously showed a 95% interval for Claude Fable 5); the ablation figure caption no longer states a per-bar hand count (the default-prompt control is the 5,000-hand run of Table 1).
  6. Table 3 now states what "w/o P2" removes.
  7. Abstract and Section 5: The strength claims are reworded; we no longer describe the agent as competing with solver-based systems, and "transferable design pattern" is now presented as an outlook alongside the counterparts of the three components in other games. A new appendix section explains why the library is written once and fixed; Section 4 gains a summary paragraph.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.