acceptodds
Withdrawn from ICLR 2027

Safe Probing of Censored Showdown Records: What an Agent Can Learn and How to Act on It

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

In poker-like imperfect-information games every hand leaves a public record of betting and payoffs, yet an opponent's private cards appear only at showdown. Weak hands fold, so showdown records systematically under-represent the hands an opponent model needs. Safe exploitation rests on the sequence-form exploitability certificate, which bounds what a response can lose against any opponent. It certifies safety but does not quantify observation: how much a probe may collect inside a safety budget, or what the censored records identify. Even the full terminal record can leave the opponent plan only partially identified. This paper couples the two questions and deploys against the residual ambiguity. A linear observation-rate objective composed with the certificate yields a *capacity linear program* (LP-C), whose optimum is the maximum target-showdown mass extractable inside a bounded exploitability budget under a nominal opponent model. The sharp identified set of opponent plans is a moment polytope: the opponent's *plan* may remain partially identified, while the probe's own *value* is point-identified from observable payoffs. Deploying robustly over the identified set is a single LP whose optimum certifies a one-sided finite-sample lower bound on deployed value at nominal coverage. Stacking probes refines the identified set and its guarantees. Under a fixed total sample budget, a dual-certificate analysis bounds the width of any deployment whose payoff vector lies in the certificate row space. It decomposes the effect of adding a probe into a weight term and a radius term. On the exhaustively enumerated Leduc Hold'em tree, the cell-polytope certificate over two probes attains a mean margin over the game value of against for a certificate built only from the two probes' mean payoffs. Over three probes, empirical–Bernstein intervals lift it to from under a Hoeffding box. The constructions apply to any finite perfect-recall two-player zero-sum game with hero-observable terminal cells.

Primary area:
reinforcement learning
Keywords:
safe opponent exploitation, imperfect-information games, linear programming, opponent modeling, poker

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_48eTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper considers 2 player zero-sum game with imperfect-information, and more precisely poker, as the notion of showdown of the hand is central in the paper as well as in Poker. The paper studies Leduc Hold'em version of poker, which is a minimalist version with very few cards (2 players, 1 card in hand each, 1 public, out of 6 cards) to scale down the combinatorial complexity.

This paper is very technical and very hard to follow. It is a succession of 10+lines technical lemmas (called surprisingly theorem in the paper), with very little explanation of the rationale, and no big picture presented early to help the reader understand why the lemmas are provided.

There seems to be 2 parts:

The first one (section 4 and 5), I finally understood (the key definition of capacity is only provided in the middle of a "theorem" statement Theorem 5.1 page 5), but an english presentation without technical jargon plagging the paper could have easily be given: "maximize the showdown of the opponent hand while only playing epsilon-optimal strategies".

The main focus here is to show that the operators are linear, and thus LP programs can efficiently optimize the policies. The results are possibly new, but quite classical. Also, the result may not be so important if what the LP programs can handle is Leduc Holdem and not a full game of Poker.

The second focus (section 6 and 7), looks more interesting, but I gave up trying to understand the details: My best guess is: use several policies to identify the policy of the opponent, and thus improve the gains. Techniques could be more interesting there, but I really can't say for sure. Importantly, there are 2 core problems in this methodology:

  1. the strategy space of the opponent seems overly restricted to these "censoring policies" defined by one scalar.
  2. worse, the definition of the value for the hero player is that we want a hero strategy which is epsilon optimal versus the worse counter strategy of the opponent. In section 6 and 7 at least, it seems to be the opposite: the opponent strategy seems to be fixed, and a set of hero strategies are provided to identify the fixed opponent strategy.

These 2 points should have been discussed, and most of the technical lemmas should have been put in appendix to keep only the high level interesting and understandable statements.

Also, how these two parts are linked is unclear to me.

Last, the experimental part is very short, less than one page on page 9, impossible to understand for a non expert.

Further, there is no conclusion.

Critical strengths
  • The 2 LP programs seem sound. However, they are obtain with what seems to me classical technique. It would have been interesting to show how complex poker games they scale to.

-The technique to identify the opponent policy may be more interesting, but I cannot judge with certainty. Also what is the relationship of this part with the LP program part is unclear to me.

Critical weaknesses
  • Presentation issues: This paper should be written in engish. Now, it is mainly a sequence of long technical lemmas. Jargon should be removed from abstract and introduction, all the more since the jargon used is absolutely unnecessary. Theorem should be understandable with limited statement to the point. Big picture should be provided early. Experimental part should be understandable. A conclusion should be provided...

  • possible soudness issue with the opponent policy, which seems fixed and taken in a very narrow set.

  • scaling to useful level seems unlikely, at least in the identification part.

Questions for authors
  • answer on the soundess part.

  • also try to explain what you are trying to solve in section 6 and 7 without using jargon.

  • how are section 4 and 5 linked to section 6 and 7?

Reviewer_U3mJOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies safe probing in two-player zero-sum extensive-form games with censored observations. It formulates probe optimization and robust deployment as linear programs, characterizes the partial identification of opponent strategies and values, and provides finite-sample performance guarantees.

Critical strengths

The motivation is clear, and the framework coherently connects safe data collection, partial identification, and robust deployment.

Critical weaknesses

Polynomial complexity in the explicit game-tree size does not imply computational feasibility in large sequential games, where the tree itself can be prohibitively large. Without scalable solution methods or evidence on runtime and memory requirements, the practical value of the LP formulations remains unclear.

The theoretical presentation is fragmented, making it difficult to identify the central contribution. Many results appear to follow directly from standard LP duality, linear algebra, and confidence-set arguments. The paper should more clearly distinguish its substantive new insights from these relatively straightforward consequences.

A central question is how to design probes that support effective subsequent learning and exploitation. Although LP-C optimizes target-showdown collection under a nominal opponent model, the connection between this objective and downstream deployment performance is insufficiently justified. The paper offers limited guidance on how to select probes, allocate samples, and balance exploration and exploitation in a concrete game.

The experiments are confined to a small Leduc game. While useful for checking the mathematical relationships and implementation, this evaluation provides limited evidence of the approach’s effectiveness or computational feasibility in larger sequential games.

Questions for authors

See weaknesses

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

This paper studies safe opponent exploitation in two-player zero-sum games like poker when certain events hide the opponent’s private information. As a result, the revealed information during multiple rounds may not be complete. In such games, authors investigate whether an agent can exploit an opponent to achieve a better value while remaining within an exploitability budget, what those observations reveal about the opponent’s strategy, and how to use them to choose a deployment strategy with a certified value guarantee. The central observation is that censored records may leave the opponent’s strategy ambiguous while still identifying the value of a deployment strategy.

Critical strengths

Safe exploitation from censored observations is a relevant problem in computational game theory. The observation that the opponent's plan can stay ambiguous while the value of a deployment is determined is interesting.

The theoretical development mostly appears sound and complete. See my response to weaknesses. I didn't find any major red flags.

Critical weaknesses

The problem is niche for the ICLR audience, and I'm not convinced that the proposed constructions improve meaningfully over simpler pipelines, or that the certificates remain meaningful beyond Leduc.

  1. The claim that LP-R improves deployed value is not well-established. As far as I understand, robustness over the identified set guarantees a lower bound on value. It doesn't necessarily yield a better exploit. Since the Nash and payoff-only certificate collapse to a Nash equilibrium, by construction, they may not be sufficient to separate the value and exploit. Thus, the contribution of LP-R would only be a certificate, not a better exploit. It is not clear why LP-R improves deployed value.

  2. The claim that any finite perfect-recall two-player zero-sum game with hero-observable terminal cells appears to hold theoretically, but is not backed by evidence.* The method requires simultaneous confidence bounds over all terminal cells, using Bonferroni corrections with an additive term that remains large for rarely observed cells. However, once the number of cells grows with game size (combinatorially), the certificate may become vacuous unless the number of samples also increases significantly. Even in Leduc, the certified margin falls to 0.0035 at N = 6000 with four probes (Table 3).

  3. The paper, in its current form, fails to motivate connections to learning representations or primary focus area for ICLR. It's based on LP-based game theory with partial identification and finite-sample statistics.

Questions for authors

Response to my concerns in weaknesses section.

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

Related papers

Discussion (0)

Sign in to comment.