acceptodds
Under review as a conference paper at ICLR 2027

Adaptive Local Perturbation Counterfactual Regret Minimization

Abstract

Imperfect-information games can admit Nash equilibria whose behavior at low- probability or unreachable information sets is strategically implausible. We in- troduce ALP-CFR, a counterfactual-regret-minimization framework that uses information-set-level local perturbations and adapts their magnitude from cumu- lative reach information. Under-reached descendants generate protection requests that propagate toward ancestors, allowing the method to allocate larger local per- turbations where downstream behavior receives too little effective reach. This pa- per develops the framework, its full-tree regret analysis, and the conditions needed for a sampling-based extension. We establish the scope of the claims carefully: the results concern perturbed-game regret, vanishing local perturbations, and sam- pling estimators under explicit assumptions; they do not, in general, establish convergence to an extensive-form perfect equilibrium (EFPE). Experiments on imperfect-information benchmark games will evaluate the effect of the protection mechanism on Nash and information-set regret metrics.

Primary area:
reinforcement learning
Keywords:
CFR, imperfect-information, two-player zero-sum games, regret minimization, information-set regret, NE refinement, sequential equilibrium, external sampling

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_kyxyOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper proposes a variant of Perturb CFR.

Critical strengths

Honest data reporting. Partial and unavailable runs are marked. The not good results are reported.

Critical weaknesses
  1. The paper's central contrast against prior work is factually incorrect, and its distinguishing idea is largely intrinsic to the EFPE framework. The following works are directly relevant and are not discussed or compared [1-4].

[1] Ren et al., Last-Iterate Convergence in Adaptive Regret Minimization for Approximate Extensive-Form Perfect Equilibrium, EC 2025. This is the closest work that exists to ALP-CFR's stated contribution, with a convergence guarantee ALP-CFR lacks, and it appeared more than a year before this submission.

[2] Čermák et al., Practical Performance of Refinements of Nash Equilibria in Extensive-Form Zero-Sum Games, ECAI 2014. An experimental comparison of refinements in exactly this setting; the paper's experimental mission largely repeats it a decade later without acknowledgment.

[3] Masaka et al., The Power of Perturbation under Sampling in Solving Extensive-Form Games, JSAI 2025. Perturbation under sampling, the exact setting of MC-ALP-CFR in the paper.

[4] Li et al., Efficient Online Pruning and Abstraction for Imperfect Information Extensive-Form Games, ICLR 2025. The paper's protection mechanism decides where to increase training effort; Pruning decide where to remove it. This is the same resource-allocation question and the connection is never discussed.

  1. No theory supporting the paper's objective. Every theoretical result is conditional in a way that voids the objective. The paper actually has no progress in theory.

  2. No empirical advantage; core claim never measured. Motivating narrative (low-reach information sets in large games) is untested: all games are tiny, and the largest has no result. The experiment shows ALP-CFR do not work well in small games.

  3. The paper is written in a defensive register that prevents the reader from extracting any positive claim.

Questions for authors
  1. How does ALP-CFR differ from [1], which already introduces an adaptive, information-set-level perturbation scheme for ε-EFPE with last-iterate convergence? Please provide a head-to-head comparison or a precise mechanism-level distinction.

  2. What is the formal definition and provenance of "RTCFR+"? Why was SPCFR excluded from the whitelist while RTCFR+ was included?

  3. Why is there no completed ALP-CFR run on Leduc-5, and no completed MC-ALP-CFR run outside Kuhn?

Reviewer_y19qOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes ALP-CFR, which incorporates adaptive local perturbations at each information set into CFR. It adjusts lower bounds on action probabilities according to cumulative reach and also adjusts perturbations at ancestors of information sets with low cumulative reach. It experimentally evaluates exploitability and information-set regret for full-tree and external-sampling versions.

Critical strengths

I intuitively agree with the concern that behavior at low-probability information sets is not sufficiently evaluated by overall exploitability alone.

Critical weaknesses
  1. The research objective is unclear.

It is unclear whether the goal is to approach a stronger equilibrium concept or to reduce regret at each information set. The connection between the motivating solution concepts and the empirical success criteria therefore needs clarification.

  1. The problems caused by insufficient reach are not sufficiently explained.

I agree with the intuition that insufficient reach can be undesirable, but the paper does not provide a concrete failure case connecting it to the problem the method aims to solve. This is particularly relevant for full-tree CFR, which updates counterfactual regrets even at information sets unreachable under the updating player’s own strategy. It remains unclear when low cumulative reach indicates a learning deficiency that requires the proposed mechanism.

  1. Experiments isolating the protection mechanism and evaluating its applicability are insufficient.

Experiments isolating the effect of the protection mechanism are insufficient. Evaluating the proposed mechanism within CFR+ or PCFR would also help establish its applicability across update methods.

  1. Comparisons with simple alternatives are missing.

Comparisons with simple methods, such as ε-greedy, are needed to assess whether the proposed mechanism offers benefits beyond basic exploration.

Questions for authors
  1. What properties and success criteria does this paper aim to achieve? How do they correspond to the reported evaluation metrics?

  2. Cloud you provide a concrete failure case in full-tree CFR that demonstrates the need for the proposed mechanism?

Reviewer_yyYFOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper seems to have tried to introduce a new variant of the CFR algorithm with the help of large language models.

Critical strengths

None, the whole paper is disorganized and appears LLM generated.

Critical weaknesses

The whole paper appears LLM generated, below are just some examples of the sloppiness.

Section 4 presents what seems to be pseudocode, which points to two functions (FullTreeUpdate, ExternalSampleUpdate) that are nowhere in the paper. It also points to Appendix B for “complete lemmas and proofs”, yet Appendix B does not contain any lemmas or theorems and their proofs.

The whole Section 5 reads like an LLM generated summary of experimental runs with multiple references to Git commits, etc., rather than interpretation and discussion by the authors. The last paragraph in Section 5.1 even points to a CSV file. Section 5.3 tries to read the results, but provides claims that directly contradict the shown results. For example, lines 442 - 445 claim that ALP-CFR’s performance remains within the same order of magnitude as CFR+ and DCFR, but the main figure shows ALP-CFR at roughly 10^-5, while CFR+ and DCFR sit below 10^-6. Additionally, the last row only contains curves for CFR, CFR+ and DCFR. Even the text on lines 460 - 464 mentions that other curves are absent, but doesn’t specify the reason.

Questions for authors

Is this all just LLM generated?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.