acceptodds
Withdrawn from ICLR 2027

Auditing Learned Mean-Field Equilibrium Claims: Claim-Conditioned Response and Flow Evidence

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Mean-field game solvers return a policy together with a population flow, but standard endpoint evidence does not determine whether this pair supports an equilibrium claim. A high return need not imply response optimality; a best response to a declared flow need not generate that flow; and passive trajectories may not expose either failure. We instantiate the broader Claim-Aware Protocol Auditing (CAPA) pattern for MFGs as MF-CAPA, a post-hoc auditor that separates response optimality from population consistency, restricts evidence to executable trajectory interventions, and returns CERTIFY, REFUTE, or ABSTAIN. The central distinction is that auditing a fixed claim need not identify the game. For finite quotient models, we define a claim-conditioned trajectory-KL characteristic time, prove that audit characteristic time can be arbitrarily smaller than the time required for model identification, and derive an expected-time lower bound for decisively terminating audits and an almost-sure upper rate. An anytime mixture confidence sequence controls adaptive stopping, while the intervention interface and sound semantic bounders determine whether a verdict is supportable. In actual sequential trials, passive aliases yield only 3/200 and 1/200 decisive stops (3 and 0 wrong), versus 200/200 for each active condition. Across two separately family-wise-calibrated suites of 60 final-candidate/control cells using exact best-response probes and exact mean-field transitions, every decisive audit agrees with exact two-axis truth and all 90 single-axis/wrong-flow controls are refuted; a separate 360-cell study across three native MFGLib learners again yields no wrong decisive decision. Structured numerical relaxations illustrate robust-box verification, distinct from auditing an unknown true model. MF-CAPA provides an evidence contract for deciding which learned equilibrium declarations the available interventions support; it audits supplied candidates rather than introducing a new equilibrium solver.

Primary area:
reinforcement learning
Keywords:
mean-field games, equilibrium auditing, sequential experimental design, trajectory interventions, claim-conditioned inference

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_rf3pOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

MF-CAPA audits a supplied MFE claim (pi, L) on two axes: the response gap against L and the consistency of L with Gamma(pi). It returns CERTIFY, REFUTE or ABSTAIN with risk delta, using adaptive trajectory experiments. For a finite realizable model class, the paper shows that on-policy data cannot detect unplayed deviations, that auditing can be much cheaper than identification, and gives a change-of-measure lower bound and an anytime-valid mixture-likelihood Track-and-Stop. Experiments cover a nine-model routing MFG, small MFGLib games with neural policy-flow candidates, and LP-based bounders.

Critical strengths

Treating a learned equilibrium as a claim to be audited, with abstention, is a sensible framing. Separating response optimality from flow consistency is correct.

The statistical treatment is careful and the limitations are stated honestly.

Critical weaknesses

The theory is a specialization of known pure-exploration results, and the mean-field structure plays no role. The "broken equivalences" are simply the definition of an MFE as a fixed point. Prop. 4.1 is immediate. Prop. 4.2 follows because identification minimizes over a superset. Thm. 4.3 is Garivier-Kaufmann. Thm. 4.4 is Ville plus Track-and-Stop. Multiple-answer pure exploration (Degenne-Koolen) already covers "only the verdict must separate." Nothing in Section 4 uses the population coupling. The intervention study is explicitly population-independent, i.e., an MDP. Standard solvers induce L = Gamma(pi) by construction, so the flow axis does no work for them.

The useful setting is excluded and the tested setting does not need the method. The theory needs a finite, realizable model class. The experiments are known-model simulator audits, where G and C are computed exactly by dynamic programming, so no sequential audit is needed. With an unknown model, no guarantee is given. The environments are tiny (3 states, a 9x9 grid, 9 models). The 57x gain is against a strawman identity design. The gain over uniform allocation is about 2x, and over a fixed maximin design 1.03-1.10x for certification. All validated formal backends abstain.

Minor points: "CAPA" is never defined or cited; if it is the authors' prior work, this raises an anonymity question. The jargon is heavy and the citation style is nonstandard. No finite-population (N-player) analysis.

Overall: Correct but incremental. Known sequential-testing theory is relabeled for MFGs without using the mean-field coupling, and it is demonstrated in settings where the audited quantities are computable exactly.

Questions for authors

Give a realistic setting where G and C cannot be computed exactly, yet a finite realizable model class is available.

Which result in Section 4 uses mean-field structure?

Which widely used solvers declare an L that is not Gamma(pi)?

What happens under misspecification, when the true model is outside the class?

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

This paper proposes MF-CAPA, a post-hoc method for auditing learned mean-field equilibrium (MFE) claims. The method separately tests response optimality and flow consistency of a policy–flow pair, using executable trajectory interventions and statistical tests to return CERTIFY, REFUTE, or ABSTAIN. Theoretically, the paper defines a claim-conditioned trajectory-KL characteristic time, proves an expected-time lower bound, and establishes an almost-sure asymptotic stopping bound. Experiments evaluate the method on synthetic MFGs, learned policy–flow candidates, and native MFGLib learners, with additional structured numerical relaxation studies.

Critical strengths
  1. The method only needs to distinguish models that change the final conclusion, without having to identify all model details. The corresponding trajectory-KL characteristic time characterizes the audit difficulty and explains why testing for specific claims can be more efficient than identifying the entire model.

  2. The paper controls decision error under adaptive experiment selection using an anytime-valid mixture confidence sequence and allows ABSTAIN when evidence is insufficient. This is more rigorous than making a deterministic equilibrium judgment based solely on a limited sample.

Critical weaknesses
  1. Certification relies heavily on pre-specified model information rather than trajectory evidence alone. Algorithm 1 requires known model verdicts and trajectory likelihoods for a pre-specified model class, while the learned neural-candidate experiments rely on an exact dynamic-programming best-response probe. In practice, trajectory samples can provide evidence of profitable deviations, but cannot by themselves rule out all other profitable deviations. Although the paper discusses structured semantic bounders, it does not demonstrate end-to-end response-gap certification from unknown-dynamics trajectory data alone. Therefore, the reported CERTIFY outcomes for learned candidates still depend on exact best-response computation.

  2. Counterfactual interventions based on a declared population flow LL are not necessarily executable. Response Optimality is defined relative to a declared population flow LL. However, if L≠ΓM(π)L\neq\Gamma_M(\pi), executing a representative-agent deviation in the actual population does not necessarily produce the counterfactual trajectory defined under the fixed flow LL. The paper assumes that these trajectory interventions are executable, but does not explain how this assumption can be satisfied when $ L$ differs from the realized population flow. Consequently, the core auditing protocol may implicitly assume additional capabilities, such as the ability to control the population distribution or access to a precise simulator.

Questions for authors
  1. Without known model verdicts or an exact dynamic-programming best-response probe, under what assumptions can MF-CAPA obtain a global response-gap upper bound sufficient for CERTIFY from trajectory data?

  2. When $ L\neq\Gamma_M(\pi) $, how can a representative-agent deviation against the declared flow LL be executed in a real system without a simulator that can hold LL fixed?

  3. For a true model in the omitted boundary strip (0<ρM(c)<γ0<\rho_M(c)<\gamma), can Algorithm 1 return CERTIFY or REFUTE rather than ABSTAIN? If so, what guarantee applies to these decisions?

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

After the LLM helped me understand, here is my summary: The authors are interested in Mean Field games. In the context of multi agent RL with a very large amount of agent, if all agent follow the same policies, and these policies are independent to which other agent the agent is interacting with (anonymity), then one can collapse all the policies of the other agent into a global distribution L. Then, from the point of view of a given agent, with a policy π\pi, they interact with L, that is of course dependent of the policy π\pi, that each agent follows. The Mean Field equilibrium is a pair of a policy π∗\pi* and a crowd L* (resulted from each agent following π∗\pi*), in which:

  • π∗\pi* is optimal given L*
  • L* is indeed generated by π∗\pi*

The authors address how to verify whether a candidate pair (π\pi, L) produced by a black-box solver actually satisfies both equilibrium conditions in an unknown environment using as few sample rollouts as possible. They claim that:

  1. auditing both conditions jointly is strictly necessary because neural solvers often output a policy π\pi that is optimal for a predicted crowd LL, even when executing π\pi fails to generate L;
  2. passive on-policy rollouts of π\pi cannot verify optimality due to counterfactual blindness on unplayed actions, making active interventions (rolling out alternative deviation policies) mandatory
  3. when the environment dynamics or rewards aren't fully known, you don't need to waste thousands of episodes learning the entire world model just to verify (π\pi, )L—you can stop running rollouts as soon as all remaining plausible models of the environment agree on the same Yes/No answer ("it is an equilibrium" or "it isn't"), (or safely output "I don't know" (abstain) if your budget runs out first).
Critical strengths

The paper seems sound and once digested, and understood thanks to an LLM, the problem is very interesting.

Critical weaknesses

The critical weakness is clarity. I am not sure I have ever read a paper for which I needed so much pre-work, self background checking, and multiple passes to understand its content. The paper should be completely rewritten to lose the (likely LLM generated) complex phrasing and make it accessible. The authors should provide an abstract much more in the direction of what I have written in summary (can be greatly improved for an abstract though).

No science is good if no one can understand it, and in the current times where LLM agents generate mathematical proofs, the job of scientist should more than ever make accessibility greater.

I invite the authors to rewrite the paper to make it accessible.

Another example of this are the figures. Both Figure 1 a and b do not provide any insight, and I am not sure if such figures are really necessary in the main text.

While the content of the paper can be valuable, it is so inaccessible that I invite the authors to iterate on its clarity, to a point that any AI researcher could extract its value.

If I didn't have to review this paper, I would never have decided read it after reading the title, or after reading the abstract. Your research deserves to be read, please make it accessible.

Questions for authors

None.

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

Related papers

Discussion (0)

Sign in to comment.