acceptodds
Under review as a conference paper at ICLR 2027

Which Nash Equilibrium? Solver-Dependent Selection on Zero-Sum Nash Polytopes

Abstract

Regularized self-play—the algorithm family behind DeepNash's master-level Stratego play—drives a two-player zero-sum policy to a Nash equilibrium by repeatedly best-responding to an entropy-regularized, slowly moving reference policy ρ. When a game has a polytope of value-equivalent equilibria, this regularizer silently breaks the tie, and with a uniform reference it selects the maximum-entropy member. We ask whether the reference can be used to pick the equilibrium on purpose. We study this on five games plus a two-dimensional Nash polytope, where the full Nash set and exact best responses are known, and back every claim with seed-bootstrap confidence intervals, equivalence tests, or hypothesis tests. Anchoring the reference at a target member and running ordinary refinement steers self-play to that member with mean coordinate error 0.007 (95% CI [0.002, 0.015]) at median exploitability 5e-5 (TOST equivalence within ±0.05, p=3e-16). The anchoring persists through refinement, and the landing point follows the reference, not the initialization. The selected member approximately follows the reach-weighted information projection of ρ onto the polytope (slope 0.969 [0.950, 0.987], R²=0.993). We report the failure modes just as prominently. A fixed off-manifold reference steers only at an exploitability cost of 0.08–0.25. Too large a mirror step makes runs look converged while failing to express the selection. Targets near the boundary of the stiffest family undershoot. Table and MLP steering maps are statistically equivalent; attention adds seed variance. Finally, against a best response the selection–robustness trade-off is degenerate. The recipe—initialize the reference at the desired member and refine—recasts the KL-to-reference term of trust-region and RLHF-style RL as a selection knob, not only a stabilizer.

Primary area:
reinforcement learning
Keywords:
equilibrium selection, two-player zero-sum games, Nash polytope, counterfactual regret minimization, regularized Nash dynamics, magnetic mirror descent, maximum entropy, information projection, last-iterate convergence

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_j6JtOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper empirically studies what sort of Nash equilibria do popular families of zero-sum game solvers tend to converge towards. The paper considers popular last-iterate and regret averaging algorithms applied to normal form (matrix) and extensive form zero-sum games with known Nash solution sets, and compares the behavior of the converged methods, measuring which tend to yield higher entropy outcomes. The main conclusions of the paper through empirical demonstrations are that (i) for matrix games, regularized nash dynamics tend to yield Nash solutions with max entropy and regret averaging methods like CFR/CFR+ yield lower entropy Nash solutions, and (ii) on extensive form game of Kuhn poker this trend does not hold.

Critical strengths
  • The premise of the paper is interesting and if actually answered, could have informed practitioners on choosing solver algorithms.
Critical weaknesses
  • Paper results are unconvincing: The biggest weakness is that the conclusions of the paper are very chaotic and the only conclusion that can be certainly claimed is rather unsurprising/not new. I do not understand how the fact that regularized Nash dynamics yield higher entropy Nash solutions for matrix games is new knowledge for the community. Further, for regret averaging methods like CFR/CFR+, the paper could not draw concrete conclusions in the first place, especially for the extensive form setting, where CFR/CFR+ excel/are mostly used anyway.
  • Paper presentation resembles AI slop: It is very evident that large sections of the paper have been taken verbatim from an AI model and with due respect, resemble AI slop (several instances of the theme of "doing/claiming X not Y", sentences that dont add any value and are repetitions, etc.). Further, many concepts are arbitrarily stated without any explanation/definition (for example, reach-weighted formulations are mentioned as something which could help answer Conjecture 1, but never even defined).
Questions for authors
  • Do the authors have any solid intuition on why the behavior of CFR/CFR+ flips (from yielding low entropy solutions in matrix games to high ones in the extensive form Kuhn poker game)? I find the discussion in the paper inadequate.
  • Why did the authors only choose one extensive form game?
2 comments
AuthorsOpenReview ↗

Response to Reviewer j6Jt

thank you for the review. We answer your points in order.

"The only conclusion that can be certainly claimed is rather unsurprising / not new."

You are right that the principle is known, and the submitted version did not make clear what was new. The revision changes this in two ways.

  • It says what is known. For linear models, mirror descent that reaches a solution returns the one closest to its initialization, and exponentiated gradient from a uniform start returns the maximum-entropy solution (Gunasekar et al., ICML 2018). We now cite this as the mechanism (Section 7, Appendix G).
  • It replaces the conjecture by a theorem with an explicit condition. The general statement "R-NaD selects maximum entropy in matrix games" is in fact false; two reviewers gave a 2×2 counterexample. Theorem 1 (Section 5) proves the selection under condition (E), for any opponent, and for a class that includes R-NaD, a fixed magnet, MWU and optimistic MWU. Proposition 2 gives the exact limit when the condition fails. Neither statement was in the submitted version. We are not aware of either in the literature for games, and the existing convergence result for R-NaD assumes interior equilibria, which in a matrix game forces uniqueness (Appendix G).

"Conclusions are very chaotic."

The abstract, contributions, discussion and conclusion are rewritten around the theorem. Section 6 now lists what is open and the limitations.

"For CFR/CFR+ the paper could not draw concrete conclusions, especially for the extensive-form setting."

That is accurate, and we say so. What the paper establishes for regret averaging is limited to matrix games: CFR+ lands on lower-entropy members in 153 of the 162 ensemble games where both solvers converge (94%), the gap widens with more iterations, and the max(R,0) projection is not the cause. The theorem explains why averaging methods can differ: averages leave the set on which the argument works (Section 5). Why they land where they do is listed as open.

Q1. "Why does the behaviour of CFR/CFR+ flip between matrix games and Kuhn?"

Most of the flip was an implementation effect, which the submission reported in its appendix and which we should have put first. Our engine's CFR divides counterfactual values by the opponent's reach and averages without reach weights. In matrix games this is identical to textbook CFR; in Kuhn it is not. Textbook CFR and CFR+ land at bluff 0.203 and 0.242 in Kuhn, at or above the max-entropy member (0.201); the low values (0.005, 0.068) belong to the reach-normalized variant (Section 4.1, Table 3). We do not have a mechanism for where textbook CFR lands, and we do not claim one.

Q2. "Why only one extensive-form game?"

Kuhn was chosen because its equilibrium set is known in closed form. We added a second, minimal example (Appendix J): asym_safe written as a two-stage game with identical payoffs. R-NaD's selection moves from P(Play) = 0.654 in the matrix form to 0.494 in the two-stage form, so in extensive form the selection depends on how the game is written. Two small games do not support a general claim about sequential games, and the paper says so.

"Reach-weighted formulations are mentioned but never defined."

Corrected: Section 2 now defines it (the entropy at each information set is weighted by the probability that the player's own actions reach it).

Presentation and the ethics flag.

The submission contains the AI-use statement that ICLR requires, describing which tools were used and for what. It has been updated for the revision: it states that AI tools helped draft the new proofs and run the new experiments, and that the proofs were checked line by line by the authors. The results do not depend on trust in the text: every number is regenerated by the supplementary notebook, and two reviewers independently reproduced our update rule and its outputs to seven digits. We have rewritten the abstract, contributions, discussion and conclusion, and revised the writing of the main text: long sentences are split, the "X, not Y" constructions you pointed to are removed where they carried no content, a recap paragraph that repeated the abstract and the conclusion is deleted, and Section 5 introduces each formal statement in plain words before stating it.

Thank you for your response, I have no more questions and will finalize my score after discussion with other reviewers/AC.

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

The paper studies which equilibrium different algorithms select when a two-player zero-sum game has a non-singleton Nash set. Using five small matrix games, Kuhn poker, and 180 random games formed by augmenting a mixed 2 × 2 core with constant safe actions, it compares selection coordinates, policy entropy, and exploitability. On the tested matrix games, moving-reference R-NaD approaches the maximum-entropy equilibrium, whereas regret-averaging methods often approach lower-entropy members; a budget sweep shows that the CFR+ discrepancy need not disappear with additional iterations. The paper formulates maximum-entropy selection by R-NaD as a conjecture for matrix games. It also reports important qualifications: the naive extensive-form analogue fails in Kuhn, the selected Kuhn strategy depends on regularisation temperature and CFR implementation, and maximum entropy is not necessarily best against the specified imperfect opponents.

Critical strengths
  1. The distinction between convergence to the Nash set and selection within that set is useful. Analytically specified equilibrium families and exact best-response calculations make it possible to separate strategic selection from residual exploitability. The paired random-game comparison and the iteration-budget sweep provide concrete evidence of solver-dependent selection within the studied family.

  2. The paper is unusually explicit about qualifications to its initial picture. In particular, it distinguishes the non-standard reach-normalised Kuhn CFR implementation from textbook CFR, reports the failure of the chosen entropy characterisation in Kuhn, and explains why its matrix-game equilibria are payoff-equivalent against every opponent. These disclosures make the restricted findings interpretable and identify useful boundaries for subsequent analysis.

Critical weaknesses
  1. Conjecture 1 is false for the stated update, and the experimental construction does not test the condition responsible for this failure.

Consider the row-maximising, column-minimising game

A=(0100).A = \begin{pmatrix} 0 & 1 \\ 0 & 0 \end{pmatrix}.

Its Nash set is Δ2×{(1,0)}\Delta_2\times\{(1,0)\}, so the maximum-entropy equilibrium has row strategy (1/2,1/2)(1/2,1/2). A direct implementation of Eq. (4), with uniform policies and references, η=1\eta=1, λ=0.5\lambda=0.5, and reference resets every 60 iterations, instead gives row strategy approximately (0.6597965,0.3402035)(0.6597965,0.3402035). NashConv is approximately 3.94×10−123.94\times10^{-12} after 1,200 iterations and 2.75×10−1152.75\times10^{-115} after 12,000 iterations, while the row strategy remains unchanged at the displayed precision.

This is also an asymptotic counterexample. While the second column has positive probability, it favours the first row; each reference reset retains a strictly positive increase in the row log-odds. The second-column probability decays geometrically across reset blocks, and the row log-odds have a finite, strictly positive limit. Thus the dynamics converge to Nash without selecting maximum entropy. No assumption in Conjecture 1 excludes this game.

The tested safe-action families instead impose affine equaliser constraints. Moreover, polytope4 and the ensemble’s higher-dimensional faces add identical safe rows: uniform initialisation preserves equality of their probabilities. In polytope4, p2=p3=(1−3p0)/2p_2=p_3=(1-3p_0)/2 at equilibrium, so the two reported coordinates do not provide the independent selection constraints claimed in Section 4.2 and Appendix D. These issues leave a substantially narrower empirical result than the proposed general matrix-game characterisation. The concern is not merely the absence of a proof; the proposed generalisation fails in an elementary game.

  1. The fixed-magnet comparison conflates regularisation bias with discretisation instability.

Appendix E uses η=1\eta=1 and λ=0.1\lambda=0.1 for MMD, and Figure 5 varies λ\lambda while keeping η=1\eta=1. However, the cited MMD analysis couples the stable step size to the regularisation strength: Sokota et al. (ICLR 2023), Theorem 3.4, gives the sufficient condition η≤λ/L2\eta\leq\lambda/L^2 under its assumptions. A temperature sweep at a fixed large step size therefore does not establish the intrinsic bias–stability limitation asserted in Section 4.2.

This can be checked on the paper’s asym_safe game. Directly applying Eq. (4) reproduces NashConv approximately 3.466 at the published setting. Keeping λ=0.1\lambda=0.1 but reducing η\eta to 0.1 produces a stable regularised fixed point with NashConv approximately 0.04496 after the same 20,000 iterations. With λ=0.03\lambda=0.03, η=0.03/9\eta=0.03/9, and 200,000 iterations, NashConv is approximately 0.01377, below the paper’s operational threshold. The latter is a stability check, not a matched-cost comparison, and positive-temperature MMD remains biased away from exact Nash. Nevertheless, these checks separate that bias from the reported cycling. A stable fixed-reference control is needed to support the interpretation of what specifically requires a moving reference.

Overall, I lean towards rejection. The controlled observations are useful, but the general selection conjecture fails and the baseline interpretation requires revision. A precise restricted characterisation, together with the corresponding stability-controlled comparison, could materially change my assessment.

Questions for authors
  1. Can you confirm the 2 × 2 counterexample above and state the additional assumptions under which you believe a corrected version of Conjecture 1 holds? In particular, how should the claimed evidence from polytope4 be interpreted given the preserved equality of its duplicate safe-action probabilities?

  2. Can you repeat the existing fixed-magnet temperature comparison with a step size chosen in a stable regime for each temperature, reporting convergence to the regularised fixed point separately from Nash exploitability? Which conclusions about moving-reference selection remain after this control?

2 comments
AuthorsOpenReview ↗

Response to Reviewer PvuP

Thank you for the two concrete checks. Both are correct, and the revision is organized around them.

Q1a. "Can you confirm the 2×2 counterexample?"

Yes. Our supplementary implementation gives the row strategy (0.6597965, 0.3402035), with NashConv 3.94e-12 at 1,200 iterations and 2.75e-115 at 12,000, as you report (Appendix H). Conjecture 1 is false and has been removed.

Q1b. "Under which additional assumptions does a corrected version hold?"

Under condition (E): all equilibrium strategies of the player induce the same payoff vector, (x − x′)ᵀA = 0 on the Nash set. This is the "affine equaliser constraint" you identified in our safe-action families, stated as an explicit hypothesis. Theorem 1 (Section 5) proves that under (E), if the exploitability goes to zero, the iterates converge to the KL projection of the starting point onto the Nash set. Your game violates (E): the second column is never played in equilibrium, yet its payoff x₁ varies over the Nash set.

Your asymptotic argument ("each reference reset retains a strictly positive increase in the row log-odds") is made exact by Proposition 2. The limit minimizes KL(x‖x₀) − ū₂·x₁, where ū₂ is the accumulated weight the column player put on its second column. For a uniform start this gives p = 1/(1 + e^(−ū₂)). Prediction and simulation agree to 7 digits in six settings (Table 5); the limit moves with the column player's start, the step size, λ and the reset period.

Q1c. "How should the polytope4 evidence be interpreted?"

As you say: its two safe rows are identical, a uniform start keeps their probabilities equal, and the second coordinate is determined by the first. We withdrew the "two independent degrees of freedom" claim in Section 4.2 and in the appendix, and we state that the ensemble's higher-dimensional faces have the same symmetry (Section 4.3). The replacement test (Section 5, Table 4) uses random starts, which break the symmetry, and 60 rank-deficient random games with no duplicate rows, 31 of which have 2-D or 3-D Nash sets. All converged runs match the prediction within 5e-8.

Q2a. "Repeat the fixed-magnet comparison with a stable step size, reporting convergence to the regularised fixed point separately from Nash exploitability."

Done (Section 4.2, Figure 6, Table 6). For each λ we use η = λ/L² and run until the one-step change is below 1e-14 (convergence to the regularized fixed point), then report NashConv separately. We reproduce your numbers on asym_safe: NashConv 3.466 at η = 1, and 0.04496 with a stable step at λ = 0.1. With the stable step the fixed magnet converges at every λ we tried, on asym_safe and on Kuhn. NashConv is proportional to λ and the number of iterations to 1/λ². For example, at λ = 0.01 on asym_safe it reaches 0.0046 after 2.3 million iterations; R-NaD reaches 5e-16 in 18,000. Annealing λ follows the same cost law (Appendix I). The MMD column of Table 1 now uses the stable step and is labelled as a regularized fixed point. The statements that a weak fixed magnet "cycles" are removed.

Q2b. "Which conclusions about moving-reference selection remain after this control?"

  • Remains: the moving reference reaches an exact equilibrium at a fixed, small budget, and under (E) that equilibrium is the KL projection of the start.
  • Changed: the fixed magnet is not unstable. It is biased, and removing the bias is expensive.
  • New: a fixed magnet placed at the start is covered by Theorem 1 as well. On asym_safe its fixed point approaches the same max-entropy member as λ → 0 (0.2192 at λ = 0.01, against 0.2180). The selection therefore comes from the multiplicative form of the update, not from the moving reference; optimistic MWU, which has no regularization, selects the same member (Section 5).

What we do not claim. Convergence is a hypothesis of Theorem 1, and we have no converse for (E). Both are listed as open in Section 6.

Thank you for the response and for revising the manuscript. I appreciate the explicit withdrawal of Conjecture 1 and of the claim that polytope4 provides two independent selection degrees of freedom, together with the reported corrections to the fixed-magnet discussion. These changes directly engage with the main concerns raised in my review.

The replacement theorem and its supporting evidence now provide the appropriate basis for reassessment. In particular, the key questions are whether the result characterises the finite-step moving-reference update studied in the experiments under clearly stated assumptions, and whether the revised fixed-magnet comparison separates regularisation bias from discretisation instability.

My concern was with the generality of the proposed selection characterisation and the interpretation of the baseline, rather than with the value of studying solver-dependent equilibrium selection. A sound restricted characterisation could therefore materially improve my assessment. I welcome the additional evaluation and will take these revisions into account when finalising my recommendation.

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

The paper studies which Nash equilibrium different solvers return when the equilibrium set is not a singleton. It compares regret-based methods, fictitious play, and regularized dynamics on small games with known Nash sets, including a randomized safe-action family and Kuhn poker. Its main hypothesis is that the specified moving-reference R-NaD update selects the maximum-entropy equilibrium from a uniform start in matrix games. The paper also documents where this picture fails in Kuhn and how the selected strategy affects play against imperfect opponents.

Critical strengths

The selection question is worth asking: sharing a minimax value does not determine an agent's behavior, and the Kuhn example shows that the choice can affect payoffs against imperfect opponents. The analytically known Nash sets and explicit selection coordinates make the solver differences easier to interpret than exploitability alone.

Critical weaknesses

My main concern is that Conjecture 1 appears false for the update specified in Eq. (4), even when the iteration converges. Consider the row-player payoff matrix A=[[0,1],[0,0]], with the row player maximizing and the column player minimizing. Write their strategies as (p,1−p) and (1−q,q). The Nash set is all p in [0,1] with q=0, and its maximum-entropy member has p=1/2. Starting both policies and references uniformly, an independent implementation of Eq. (4) with the stated η=1, λ=0.5 and reset period 60 gives p≈0.6597965 and q≈1.36×10^−23 after 2,400 iterations; NashConv is exactly q in this game. I still see value in the comparisons on the tested family; my concern is with the general characterization that motivates Conjecture 1 and the conclusion.

Questions for authors

Could you check this 2×2 example with the supplementary implementation and explain any difference from the update described in Eq. (4) and Appendix E? If the example reproduces, what explicit restriction on the game or reference dynamics would support the intended maximum-entropy statement? A clearly justified restricted claim would materially affect my assessment.

1 comment
AuthorsOpenReview ↗

Response to Reviewer enmh

Thank you for the counterexample and for stating what would change your assessment.

"Could you check this 2×2 example with the supplementary implementation and explain any difference from Eq. (4) and Appendix E?"

There is no difference. The supplementary engine and a second, vectorized implementation both give p = 0.6597965 and q = 1.36e-23 after 2,400 iterations with η = 1, λ = 0.5 and reset period 60 (Appendix H). Conjecture 1 is false as stated and has been removed.

"What explicit restriction on the game or reference dynamics would support the intended maximum-entropy statement?"

A restriction on the game, which we call condition (E): all equilibrium strategies of the player induce the same payoff vector, (x − x′)ᵀA = 0 on the Nash set. It holds whenever the opponent has a full-support equilibrium, and it can be tested with two linear programs per column. Your game violates it: column 2 is never used in equilibrium, yet it pays p, which varies over the Nash set.

The restricted claim is now a theorem (Section 5; proof in Appendix G):

Under (E), if the player's exploitability goes to zero (which NashConv → 0 implies), the iterates of Eq. (4) converge to the equilibrium closest in KL divergence to the starting point. For a uniform start this is the maximum-entropy equilibrium.

No restriction on the reference dynamics is needed: the theorem holds for a moving or a fixed reference, for MWU and optimistic MWU, for any step sizes, and for any behaviour of the opponent. The proof is short. The log-ratio of the iterate to the start stays in range(A) + span(1) whatever the opponent plays; under (E) that subspace is orthogonal to the Nash set; the Pythagorean identity for KL then identifies the limit. Convergence itself is assumed, not proved.

Why the claim fails in your game, exactly. Proposition 2 gives the limit without (E): the KL projection tilted by the weight ū the opponent put, on the way, on columns it does not use in equilibrium. In your game, p = 1/(1 + e^(−ū)). With your settings the measured ū is 0.662387, which gives 0.6597965. The formula matches the simulation to 7 digits in six settings (Table 5), and the limit moves with the column player's start, so no statement in terms of the game and the row start alone can hold there.

Evidence that the restriction is the right one (Section 5, Table 4).

  • 95 games satisfying (E), 290 runs from uniform and random starts: 278 converged, and all match the predicted equilibrium within 5e-8. The remaining 12 were still converging and are not counted.
  • 31 games violating (E), 93 runs: 89 converged, and all miss the prediction, by 0.015 to 0.37.
  • A certificate that does not rely on a nonlinear optimizer (one linear program per run) separates the two groups cleanly (Figure 2).

The comparisons on the tested family that you found valuable are unchanged: all of those games satisfy (E), so the theorem now explains them. We also corrected the fixed-magnet baseline, which was run with an unstable step size (Section 4.2).

What we do not claim. We do not prove that R-NaD converges, we have no converse for (E), and the theorem does not extend to extensive-form games; Kuhn poker remains outside it (Section 6).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response to all reviewers

We thank the three reviewers. Two of them (PvuP, enmh) found a counterexample to Conjecture 1, and PvuP showed that our fixed-magnet baseline was run outside its stable regime. Both criticisms are correct. We reproduced every number they reported, and the revision is built around what they found. Changes are in blue in the revised PDF.

1. Conjecture 1 is withdrawn and replaced by a theorem (new Section 5, proofs in Appendix G).

  • Condition (E): all equilibrium strategies of the player induce the same payoff vector, (x − x′)ᵀA = 0 on the Nash set. It holds whenever the opponent has a full-support equilibrium. Our testbed and ensemble satisfy it; the reviewers' game does not.
  • Theorem 1: under (E), any multiplicative-weights-type last iterate whose exploitability goes to zero converges to the equilibrium closest in KL divergence to its starting point (the maximum-entropy equilibrium for a uniform start). This covers R-NaD, a fixed magnet, MWU and optimistic MWU, for any step sizes and any behaviour of the opponent. Limits on the boundary of the simplex need no extra assumption. Convergence is a hypothesis; we do not prove it.
  • Proposition 2: without (E), the limit is that projection tilted by how much the opponent played, on the way, the columns it does not use in equilibrium. In the reviewers' 2×2 game this gives p = 1/(1 + e^(−u)); the prediction matches the simulation to 7 digits in six settings, including the reviewers' (0.6597965). This explains why no formula in the game and the start alone can work there.

2. The theorem is tested on 383 runs over 126 games (Section 5, Appendix H).

  • (E) holds, 95 games, 290 runs: 278 converged (exploitability below 1e-9) and all match the predicted equilibrium within 5e-8. This includes 31 games with 2-D or 3-D Nash sets and no duplicate rows, run from random starts. The other 12 runs were still converging and are not counted.
  • (E) fails, 31 games, 93 runs: 89 converged and all miss the prediction, by 0.015 to 0.37.
  • Optimistic MWU, which has no regularization, ends within 1.5e-7 of R-NaD in all 250 runs where both converged under (E).

3. The fixed-magnet claims are corrected (Section 4.2, Table 1, Figure 6, Appendix I). With the step size η = λ/L² the fixed magnet converges at every temperature we tried. "Cycles / cannot converge" was an artifact of η = 1 and has been removed. What remains is a bias-against-cost trade-off: exploitability proportional to λ, iterations proportional to 1/λ².

4. The polytope4 "two degrees of freedom" claim is withdrawn (Section 4.2), for the symmetry reason PvuP gave.

5. Relation to prior work is stated (Section 7, Appendix G). Theorem 1 is the game version of the implicit bias of mirror descent (Gunasekar et al., ICML 2018, who note that exponentiated gradient from a uniform start returns the maximum-entropy solution). We say what is specific to games: the invariant holds for every opponent sequence, condition (E), the boundary case, and Proposition 2.

6. Other additions. A second small extensive-form example (Appendix J); R-NaD on Kuhn at smaller regularization weights, where its bluff frequency rises, passes the max-entropy value near λ = 0.06 and levels off about 0.001 above it (Table 8); a definition of "reach-weighted" (Section 2); the Limitations section moved to Appendix K to respect the page limit, with a short paragraph kept in Section 6.

What the paper does not claim. It does not prove convergence of R-NaD, it does not give a converse for condition (E), and the theorem does not extend to extensive-form games (Kuhn poker remains a counterexample to the extensive-form analogue).

Reproducibility. The anonymized supplementary material is now one self-contained notebook that regenerates every number and figure, including the reviewers' counterexample. We ran it on two software stacks with the same results.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.