acceptodds
Under review as a conference paper at ICLR 2027

A Generalized Theory of Card Counting

Abstract

Existing card counting systems have largely been developed through heuristic design, without a unified mathematical definition of what constitutes an optimal count. This paper introduces a framework for defining and deriving optimal card-counting systems by formulating card counting as a long-term growth-rate maximization problem. The central idea is to view a count as a low-dimensional representation of the high-dimensional information contained in the exposed-card composition: the representation should preserve the information most relevant to future returns while remaining sufficiently compact for practical use. Based on this formulation, an alternating optimization procedure jointly adapts the count representation and the playing policy, combining a growth-optimal count update with reinforcement learning for count-dependent strategy deviations. The framework establishes connections between card counting, Kelly criterion, information-theoretic value of side information, and representation learning, and can be generalized beyond classical blackjack to variants with different rules and payoff structures. Experiments improve both long-term growth and fixed-spread return, increasing growth by up to 30.92% over Hi-Lo with the Illustrious 18 under 3:2 blackjack and by 76.98% under 6:5; the framework also demonstrates its value on new variants, achieving positive growth and spread in the evaluated Spanish 21 game where baseline methods loses.

Primary area:
reinforcement learning
Keywords:
Reinforcement Learning, Q-Learning, Representation Learning, Sequential Decision Making, Growth-Optimal Betting, Kelly Criterion, Blackjack

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_SGZ8OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper addresses a general theory of card counting. A count is a statistic of the discarded cards. Eg, "Hi–Lo", made famous from TV (eg, 21), adds/subtracts 11 when low/high cards are discarded (namely, 2–6/10–A); a high count (normalised by the number of cards remaining) indicates favourable conditions.

The paper introduces a unified mathematical framework for defining, and deriving, optimal card-counting systems - something that the paper claims is absent in the field. This is connected with other topics, such as information theory (obviously) and representation learning.

The method is evaluated on a variety card games and comparisons with strategies, demonstrating the flexibility of the approach

Critical strengths

Overall, I liked the paper. It isn't my area - I come from probability, including card shuffling, but not card counting. Nevertheless, I found the introduction compelling, and a clear case for the paper was made:

  • I feel quite strongly that progress can be held back by lack of clear definitions of what people are trying to do. As much as anything else, it hinders comparisons of approaches, when different authors have different goals
  • Whenever I see "... without a unified mathematical definition of what constitutes an optim[um]. This paper introduces a framework for defining and deriving [such optima]..." I'm very happy
  • Setting up a clear framework that crosses games - eg, "Spanish 21 is a key test of the theory because it changes both the deck composition and the game rules, precisely where transferred counting systems may become mismatched (§4.3) - is very important

I definitely recommend this paper for acceptance. That said, I cannot give a really detailed review, as I just don't know the topic so well. I leave the more nuanced comments to the other reviewers. But, on the more abstract question of "Is this type of paper important?" my instinctive feeling is a strong "yes".

The fac that I could not check all the details precisely, including understanding what the standard baselines are, leads me to restrict my rating to "3/4: Weak acceptance" - instinctively, I'd like to give "4/4: Clear acceptance", but my understanding can't justify that.

Critical weaknesses

For someone not in the field, I felt that the paper was a bit inaccessible.

  1. The abstract was a bit jargon-y:
    • I couldn't understand the description of the main results from the abstract alone
    • Likely, these are understandable to someone in the field, but I feel that the abstract could be more accessible
    • Even explaining what a "count" is exactly, just as "a statistic of the discarded/used cards", would have helped a lot
  2. The structure of §2 felt a little like a braindump:
    • It just felt like reading a list of statements, but without understanding why
    • There was no foreshadowing or guidance, to understand what the purpose was
    • Again, this might be simply because I don't know the area

Unrelatedly, I do think that it perhaps oversells a little:

  • The impression given by the introduction (§1) is that this is for quite general card-counting systems
  • It actually only handles the case in which the count is a weighted sum of the number of each rank discarded so far (§2.2)
  • The "framework for defining and deriving optimal card-counting systems" (§Abstract) only applies to this one-dimensional projection

There were several other minor notes that didn't affect my understanding substantially, but could be fixed or clarified.

  1. The notation c=(c2,c3,...,cA)c = (c_2, c_3, ..., c_A):
    • Why not start at 11? Is it already exposed, hence not needed to be conditioned on? Where is the number AA introduced?
    • I realised after that "AA" → "Ace", and this is "(2, 3, ..., 10, J, Q, K, A)". This could be made clearer
  2. Betting function and the ff-notation:
    • The expression "betting function" wasn't known to me
    • It wasn't clear how 1+f(Yw)Xπ(w)>01 + f(Y_w) X^{\pi(w)} > 0 could be enforced
    • Sometimes it's f(Yw)f(Y_w) but sometimes just ff
    • This is likely due to not really understanding the definitions fully
  3. Typo in §2.5:
    • The second display is missing an ff on the right-hand side
  4. Notation in §2.6:
    • Why did the notation change from G(w,π(w))G(w, \pi(w)) to JπJ_\pi?
    • I guess that the point is to be able to use ∇wJπ(w)\nabla_w J_\pi(w)
    • Tbh, I didn't really like G(w,π(w))G(w, \pi(w))
      • Why not G(w,π)G(w, \pi)? Indeed, this is the case in §2.8
      • It's not like G(w,p)G(w, p) was defined for any (w,p)(w, p) and then using p=π(w)p = \pi(w)
  5. Zero-gradient in §2.6 (bottom of page 4):
    • Surely this is a condition on ww, not a statement?
    • That is, it's at the local maximum/minimum
  6. Citation format:
    • Should use \citep rather than \cite for many in-line citations, such as in §2.9
    • This is really a nit-pick!
  7. Notation change (§2.8 → §3.1):
    • The notation in §3.1 suddenly changes to PkP_k instead of πk\pi_k
    • §2.8, in which πk\pi_k is used, is even referenced to explain this sequence
Questions for authors

None

3 comments

I have received an email entitled "[ICLR 2027] Discussion reminder", saying,

We are about to enter the reviewer-AC discussion period. The authors of the following papers have posted responses to your review, but it appears that you have not yet replied:

<link to this page>

To facilitate the next discussion phase, we would appreciate it if you would reply to these discussion threads letting the other reviewers, AC, and authors know if your concerns have been addressed.

However, on my view of the page at least, the authors have not posted any response to my review - which, assuming that my view is accurate, is somewhat disappointing. If they have, and I just can't see it, please accept my apologies - I am not refusing to engage, I simply cannot see the comment.

AuthorsOpenReview ↗

Comment to Reviewer SGZ8

I am not sure why you can't see my response, but I will paste it here:

$\textit{ Thank you for the positive and thoughtful review! I really appreciate your comments and suggestions, and I'll use them to improve the clarity and accessibility of the paper.

Regarding the generality of the framework, I think you raise a fair point. The generality comes from defining the components of card counting mathematically, allowing the framework to be applied across different games and rules. The restriction to one-dimensional linear counts is intentional, as our goal is to develop counting systems that a human player can realistically maintain and use in real time. I'll make this distinction clearer in the paper. }$

I want to take this opportunity to thank you again for the positive and useful review.

Thanks for responding, even so close to the deadline. Idk why I can't see it. Perhaps an AC+ should look into this?

Here's an image of my view of the page

My identifying information has been removed and it's been uploaded anonymously, of course.

Regarding your response, that all sounds fine and reasonable. I don't think that it's enough for me to raise my score. Indeed, I didn't raise concrete enough concerns for that. My lack of familiarity with the area - other results, baselines and the like - limits my judgement. But, from a slightly outsider point of view, I certainly enjoyed the more qualitative aspects of the paper

Reviewer_4viLOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops a general framework for card counting by framing it as a long-term growth rate maximization problem. It treats a count as a low-dimensional statistic that describes potentially high-dimensional information relevant to making decisions that maximize long-term bankroll growth. The proposed method alternates between learning count-dependent deviations from a basic strategy using Q-learning and refitting count weights under the resulting policy. The method is validated empirically in Blackjack and Spanish 21, outperforming existing heuristic-based count and deviation methods

Critical strengths

Replaces heuristics with a principled objective. The paper gives a principled definition of count optimality, maximizing long-term logarithmic growth rate (the Kelly objective), for a problem that has historically relied on hand-tuned methods and metrics. This results in the benefit that both the algorithm and evaluation transfers naturally to different rulesets.

Reliable and reproducible experiments. The simulator is validated against published values and the learned policies are checked against published baselines (Appendix D).

Critical weaknesses

Narrow evaluation The method is evaluated on Blackjack and Spanish 21, but the core claim is the methods generality.

Weak connection to the ICLR community The representation is a simple linear statistic with tabular Q-learning as the learning algorithm. The paper mentions richer representations, but leaves experiments with them as future work. The paper cites theory as its main contribution, but the theoretical novelty of the contributions is questionable in my opinion. It seems more like a principled application that connects some relatively simple theoretical ideas (e.g. value of information, Kelly). As a result, I'm not sure that ICLR is the appropriate venue for the work.

Overstatement / conflation of results The abstract lists up to a 30% improvement over Hi-Lo + Illustrious 18, but the comparison that gives that number does not seem fair. The baseline uses a limited set of deviations, while the proposed method uses a full tabular deviation set. The authors reported numbers against Hi-Lo with the full deviation set in Table 6 of the Appendix show a more modest gain, and perhaps those are the fairer headline numbers that should be reported.

Questions for authors

Can the authors elaborate on the generality of the approach beyond blackjack-family games? How well does it apply to other settings?

Reviewer_rHaTOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper formulates card counting in the blackjack game as an optimization problem. The authors introduce a long-run logarithmic growth rate and propose L-GOC, which alternates between learning the playing policy using tabular Q-learning and calculating count weights via a closed-form constrained solve. This work includes a theoretical analysis which justifies the use of Q-learning for the policy. Experiments cover traditional blackjack and Spanish 21, showing consistent improvements over classical counting systems.

Critical strengths
  1. The core strength of this work is the optimization formulation of the card-counting problem, which replaces traditional hand-tuned tags and allows a unified learning paradigm across different game settings, particularly under modified rules such as 6:5 and Spanish 21.
  2. The experimental settings are comprehensive, and overall performance is promising compared to the most popular prior approach. The ablation study is also convincing.
Critical weaknesses
  1. The whole pipeline rests on the small-wager approximation. However, in adversarial games such as blackjack, most profit comes from the rare high-count rounds where the wager is largest — precisely where the approximation is weakest. Results in the appendix also confirm the resulting decision reversals begin at wager fractions well within the normal operating range, and capping high exposure discards profitable opportunities. More theoretical or empirical results are needed to convince the readers that this approximation is sound across all game scenarios.
  2. The comparison with baseline approaches may be unfair. For example, the Illustrious 18 indices are derived for multi-deck games, whereas the headline numbers are typically from the single-deck cases. The Spanish 21 comparison has similar issues: the omitted bonus payouts disadvantage Walker's tags, and Hi-Lo's wager calibration sits at a search boundary. Furthermore, Appendix K.4 already compares against stronger classical systems (e.g., Wong Halves with full deviations), and in 3:2 blackjack the confidence intervals of the two overlap at two and six decks.
Questions for authors
  1. Could you provide statistical results of optimizing directly for logarithmic growth? This would provide evidence that the small-wager approximation is sound. Also, could you elaborate on the convergence of the alternating optimization, given that the two steps optimize different surrogates?
  2. Analysis comparing L-GOC with more recent baselines (such as Appendix K.4) needs more discussion. The generalization of L-GOC across different rules is solid, but could you clarify the improvement over the strongest classical baseline in the main text?
  3. Could the authors add an upper-bound reference, such as the full-information policy's growth rate or an estimate of I(X;Y), so that readers can judge how much of the achievable gain is captured?
  4. The paper is hard to follow, especially for readers who do not have basic knowledge of blackjack and card counting. Two suggestions:
  • The introduction enumerates historical counting systems rather than establishing the conceptual foundations a non-specialist needs: what card counting is, why it works, and what the two decision problems (bet sizing and playing deviations) are. Some are mentioned in Section 2, but I think a dedicated background subsection is better.
  • The logarithmic growth objective is introduced by citation but never motivated for a general ML audience. An explanation of why expected log wealth is the appropriate objective is needed.
  1. In Section 2.4, why is the betting rule restricted to f(Y_w)=Y_w? I didn't find the justification, and to my understanding, the former is a fraction of bankroll while the latter is a dimensionless statistic of the discard tray.
  2. What is the transferable insight beyond blackjack and its variants? Could you demonstrate it on games outside the blackjack family, or alternatively moderate the generality suggested by the title? This would contribute more to the broader community.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.