| paper | Learning to Manipulate Under Limited Information |
| authors | Wesley H. Holliday, Alexander Kristoffersen, Eric Pacuit |
| venue | AAAI 2025 |
| filed under | voting · manipulation |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting an algorithmic or complexity result. Its only relevant complexity statement is an unnumbered citation to prior work, so bit (a) fails under the explicit rubric. The proposed high-multiplicity Nanson coalition problem is coherent, but cannot anchor a mirror of this paper.
fails bit a — no named computational result to mirror
The proposed full-information mass-bloc problem drops the paper’s defining limited-information learning setup and cardinal expected-utility objective; restoring those requires an additional prior and utility-type model.
fatal: False
The proposed mirror covers Nanson strategic manipulation over a high-multiplicity population, but leaves the paper’s neural-network learnability, limited-information policies, model-size comparisons, probability-model experiments, and other voting methods untouched.
Strictly, this paper supplies no compliant anchor. It contains no numbered Theorem, Lemma, Corollary, or Proposition establishing a computational classification. Its closest claim is the unnumbered sentence attributing NP-hardness of Nanson manipulation to Narodytska, Walsh, and Xia (2011); that result is cited from elsewhere, not proved or numbered here. The paper’s own results are experimental claims about neural-network profitability and model size. Thus, under the stated rubric, no paper-anchored positive verdict is formally available.
The strongest charitable case uses that Nanson claim as a non-compliant anchor. Consider High-Multiplicity Nanson Mass Manipulation. Let \(C\) be the candidates and let \(T\subseteq\mathcal R(C)\) be the finite set of complete ballot types. A society is a rational mass vector \(\mu\in\mathbb Q_{\ge0}^{T}\) summing to one. A designated homogeneous bloc has type \(t^\star\), mass \(\alpha\le\mu_{t^\star}\), and may replace its sincere ballot \(t^\star\) by any ranking \(r\). The reported profile is
\[ \mu'=\mu+\alpha(e_r-e_{t^\star}). \]
Apply strict Nanson to \(\mu'\), using real-valued Borda scores and average scores at every elimination round. Given a target candidate \(c^\star\), decide whether some ranking \(r\) makes \(c^\star\) a Nanson winner, and output such an \(r\) when one exists.
This is a genuine continuous-population question: \(N\gg\tau\) members of a party, professional association, or large electorate share a finite set of ballot types, while a coordinated caucus controls an \(\alpha\)-fraction of the mass. The paper’s authors should recognize the same voting rule, strategic ballot, and desired-winner objective. The change from one voter to a positive-mass bloc is necessary: an individual has zero influence in an atomless society. It is also consistent with the paper’s own suggestion of coalition manipulation as future work.
I would expect this problem to be Class B when the number of candidates is variable. A discrete Nanson manipulation instance embeds by taking \(\mu_t=n_t/N\) and \(\alpha=1/N\). To make the population genuinely high-multiplicity, duplicate every voter type \(K\) times; all Borda scores scale by \(K\), so Nanson’s elimination sequence is unchanged, while the electorate can be arbitrarily larger than its type catalogue. The hardness lives in the candidate agenda, not in population multiplicity. With fixed \(m\), however, testing all \(m!\) reports is straightforward, so that restricted regime is tractable.
The paper-faithful limited-information extension is also natural: give the bloc only a plurality-score vector, majority matrix, margin matrix, or sincere-winner set, together with a prior over possible mass vectors, and ask for a policy mapping the observed signal to a report maximizing expected normalized utility. Further questions include the minimum \(\alpha\) needed to elect \(c^\star\), approximation under incomplete aggregate information, and whether Borda, Minimax, Split Cycle, and Stable Voting have different mass-manipulation thresholds.
The weakest point is decisive: this is a coalition/high-multiplicity extension of an unnumbered result cited from another paper, not a numbered computational result proved by these authors. Moreover, their sampled utility profiles almost surely contain no repeated complete cardinal types. So this is a defensible research mirror, but not a fully rubric-compliant positive case for this particular paper.
The negative case is decisive at the source gate: this paper supplies no compliant anchor. It contains no numbered theorem, lemma, corollary, or proposition giving a computational result. Its own contribution is experimental: neural-network profitability, model size, and comparisons across probability models. The only complexity statement is the unnumbered attribution that Nanson manipulation is NP-hard, a result proved in Narodytska, Walsh, and Xia (2011), not in this paper. Under the stated rubric, that cannot anchor a mirror of this paper.
Granting the proponent’s charitable use of Nanson, their proposed problem is mathematically coherent but changes the paper’s computational object in several fundamental ways.
The paper studies one voter who changes one ranking, with limited information about the other voters and an objective defined by that voter’s cardinal expected utility. In an atomless population, that individual has zero mass and cannot change any aggregate score. A nontrivial limit therefore requires replacing the individual by a positive-mass bloc. That is not merely replacing counts by masses: it changes unilateral manipulation into coalition manipulation. The duplication argument establishes that this coalition problem has a sensible high-multiplicity regime, but it does not make it the paper’s single-voter problem. It is precisely the coalition extension the authors list as future work.
The proposed Nanson problem also removes the paper’s distinctive limited-information component. With the complete mass vector \(\mu\) supplied, the manipulator knows the entire relevant profile and simply searches for a report making a target win. That is ordinary full-information manipulation, not learning to manipulate under limited information. To restore the paper’s subject, one must give the bloc only a signal such as a majority matrix and specify a prior over possible mass vectors. But then the central computational issue is the representation and integration of that prior, not continuization itself. A finite parametric prior, an oracle for posterior probabilities, and an explicit list of possible mass vectors produce materially different problems. The proposed “paper-faithful” version is therefore an additional Bayesian decision-theoretic model, not a direct population mirror.
There is a further type-space mismatch. The paper’s probability models generate cardinal utilities from continuous distributions, and the payoff is normalized cardinal expected utility. Almost surely, every voter has a distinct utility vector. Treating only rankings as types preserves Nanson’s winner calculation but discards the utility information that defines profitability and the learning objective. Treating cardinal utility vectors as part of the type makes the type space uncountable rather than the finite \(T\) required by the programme. Discretizing utilities or restricting them to finitely many classes would be a new model.
The limiting regime also destroys the paper’s empirical phenomenon. Under the paper’s iid models, aggregate plurality scores and margins concentrate as the population grows. The finite-election effects the paper emphasizes—pivotality, parity, ties, and a single voter’s ability to change the winner—either disappear or become measure-zero boundary events. Keeping a positive bloc mass preserves a nontrivial problem only by changing the strategic actor; scaling the bloc down with population size instead makes its gain vanish or requires a separate fluctuation-limit analysis.
The strongest concession to the proponent is that their mass-Nanson problem is not nonsensical. Rational-clone scaling preserves Borda averages and Nanson’s elimination sequence, and it could be a worthwhile study of coalition manipulation in large electorates. But it is an extension inspired by one cited background result, not a continuous mirror of a named computational result in this paper. The paper’s actual contribution—learnability of profitable manipulation under limited information—has no formal computational anchor to continuize. On the programme’s explicit screening rules, that is enough for a negative verdict.
The adversarial triple: the proponent anchors on up to three named results; the opponent sees that case and must defeat every anchor; the judge decides which case convinced it. These are the pipeline’s own outputs, generated by tools/triple_run.py — no human edited them. The paper’s own text is not reproduced here beyond the quoted statement above.