acceptodds
Under review as a conference paper at ICLR 2027

When Does Nash Averaging Ignore a Duplicate? The Exact Scope of Redundancy Invariance

Abstract

Nash averaging reads pairwise results as a zero-sum game and rates each agent by its expected result against an equilibrium. When there is more than one it takes the maximum-entropy equilibrium, the most spread-out one. The method was proposed with the property that an exact copy changes no rating. We show that the property can fail, and that the published proof covers the set of equilibria, not the maximum-entropy equilibrium. A copy tilts the entropy in proportion to the copied agent's probability. We prove that the maximum-entropy equilibrium stays in place exactly when no step inside the equilibrium set gains more from this tilt than it loses in entropy. When it moves, only agents that no equilibrium plays can be re-rated, while the rest tie at the top. A copy never moves a unique equilibrium, including those of the Atari score table and chatbot-arena leaderboard we check. Coarse scores leave more than one equilibrium in forty of the forty-four football seasons studied, often through exact draws. In seventeen of them, adding a copy of one team reorders two others. More than one equilibrium lets a rating jump at an exact draw. Any rule that never jumps is wrong somewhere nearby by at least half that rating's range; a vanishing relaxation of the equilibrium conditions costs at most the whole range and loses invariance to copies. A published Nash average should say which rule picked its equilibrium and whether others exist.

Primary area:
learning theory
Keywords:
Nash averaging, agent evaluation, maximum-entropy equilibrium, clone invariance, equilibrium selection, rating systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_HdRTOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies scenarios of agents that compete in pairwise competition (e.g. think of participants in an on-line chess site; you want to work out their strength at chess based on pairwise game results). We're in a setting of win-lose-draw competition. The paper studies Nash averaging, a rule for computing ratings of players that is intended to be resistant to the operation of duplicating a player (which ought not to affect the relative rating of the other players). Since Nash averaging can in fact affect these relative ratings, the paper studies the effects of such duplication, and conditions under which it doesn't change things.

Critical strengths

Highlights a problem with a technique that's been well-studied and used in practice. Studies the impact of the problem on real world data.

On para 2, it looks plausible that copying an agent looks like it shouldn't change any rating, since it's like duplicating a strategy in a game, which ought not to change the game's value. It's pointed-out that the property fails. Paragraph 3 gives a nice example of the property failing; the choice of "max entropy" when there are multiple solutions, is bad.

Critical weaknesses

The abstract is hard to read, I did not get much out of it. Para 1 of the paper is much better.

Some unclear/vague wording "Agent 4's own results are untouched". line 046 "even distribution", I'd prefer "uniform distribution". But I get the idea.

Questions for authors

Question: Given that the max-entropy rule has been previously used/studied, the paper considers the question of conditions under which it works. To me there is raised the bigger question of what, if any, alternatives to the max-entropy rule we ought to be using (which is touched-on in para in middle of p.3).

Suggestions: On p.2 it might be helpful to make the discussion: We have a symmetric 0-sum game whose value is zero, hence every agent (ie pure strategy) that gets played, achieved payoff zero, and any agent that doesn't get played (and is inessential) would get a negative payoff if they were played.

Line 080 "no step inside the equilibrium set gains more from this tilt", usage of "step" and "tilt" is unclear to me. They should be explained better, especially "tilt". Line 087 "As the copy becomes exact" -- I'm guessing I'm supposed to be thinking of a process whereby an inexact copy converges towards an exact one, is that right?

p.2, top, you should say what p and q denote, although it's not hard to figure out.

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

This paper revisits the redundancy invariance property of Nash averaging, which evaluates agents using a maximum-entropy Nash equilibrium of a zero-sum game constructed from pairwise performance scores. Although Nash averaging has been claimed to be invariant to the duplication of an agent, the paper demonstrates that this property does not hold in general when the game admits multiple equilibria.

The main theoretical contribution is a precise characterization of how duplication affects equilibrium selection. The authors show that duplicating an agent introduces a linear bias into the entropy-maximization objective, and derive a necessary and sufficient condition for the selected equilibrium to remain unchanged. They further characterize how duplication affects equilibrium probabilities and ratings, showing that only inessential agents can have their ratings changed.

The paper also studies sensitivity to small perturbations. It constructs examples where ratings are discontinuous even though the perturbed games have unique equilibria, and establishes a lower bound on the approximation error that any continuous rating rule must incur.

The theoretical findings are supported by small counterexamples and experiments using synthetic games and real-world evaluation data, including football results and AI-agent benchmarks. The experiments illustrate when multiple equilibria arise and how duplication may affect ratings and rankings.

Critical strengths

The paper identifies an important distinction between invariance of the equilibrium set and invariance of a particular equilibrium selected by maximum entropy. The counterexample is simple and convincing, and the entropy-tilting characterization provides a clear explanation of why redundancy invariance can fail. The necessary and sufficient condition for invariance goes beyond demonstrating a counterexample and gives a more complete theoretical understanding of the issue.

The results on discontinuity and the limitations of continuous rating rules are particularly interesting from a game-theoretic perspective. They highlight a fundamental difficulty in selecting a single equilibrium from a set of equilibria while maintaining desirable robustness properties. I find this theoretical aspect of the paper valuable, independently of the immediate application to agent evaluation.

Critical weaknesses

The failure of redundancy invariance depends on the existence of multiple equilibria and, for changes in ratings, on the presence of inessential agents. The paper provides empirical examples where this matters, but it is less clear how frequently the phenomenon would have a substantial impact in typical AI evaluation settings. In particular, some of the reported AI-agent evaluation matrices have unique equilibria, for which the identified problem does not arise. Thus, while the theoretical results are interesting, their practical implications may be relatively limited.

The distinction between set-valued solutions and single-valued equilibrium selection, as well as the sensitivity of equilibrium selection to perturbations, has a substantial history in game theory and social choice. The paper discusses related work on maximal lotteries and clone independence, but I am not sufficiently familiar with this specific literature to confidently assess how much of the theoretical contribution is fundamentally new.

In particular, I would appreciate a clearer explanation of which results are specific to Nash averaging and which represent new insights into equilibrium selection more generally. This uncertainty about novelty is my main reservation, rather than a concern about the correctness of the results.

Questions for authors

1. Novelty relative to existing equilibrium-selection results.

Could the authors clarify how the main results, particularly the entropy-tilting characterization and the continuity lower bound, relate to existing results on equilibrium selection, refinement, and clone independence in game theory and social choice? Which aspects should be regarded as the principal new theoretical contributions beyond the counterexample to the previously claimed invariance property?

A more explicit comparison would help me assess the significance and originality of the theoretical results.

2. Practical relevance to AI-agent evaluation.

Could the authors elaborate on the conditions under which the failure of redundancy invariance is likely to have a meaningful practical impact? In particular, how often should we expect multiple equilibria with materially different ratings in realistic AI-agent evaluation settings, as opposed to settings with exact ties or coarse outcome scores?

Clarifying this point would help me better understand the significance of the results for the intended applications.

Reviewer_WY2kOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper revisits the redundancy invariance of Nash averaging and demonstrates that duplicating an agent can change the maximum-entropy Nash equilibrium and the resulting ratings. The authors derive an exact characterization of when invariance holds, establish structural results on rating changes and discontinuities, and provide extensive experiments on synthetic games and real-world evaluation datasets.

Critical strengths
  1. A real gap in a widely restated claim, with a clean five-agent counterexample and an intuitive mechanism.
Critical weaknesses
  1. The main Atari and chatbot-arena datasets have unique equilibria, where the identified issue does not arise. Although non-uniqueness appears in subsampled leaderboards, most observed rating changes are smaller than the sampling uncertainty. Can the authors demonstrate practically significant ranking changes in a realistic ML evaluation setting?
  2. The paper characterizes the limitations of existing methods but does not provide a broadly applicable alternative.
Questions for authors

Same as weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.