| paper | Collective Robustness of Heterogeneous Decision-Makers Against Stubborn Individuals |
| authors | — |
| venue | AAMAS 2024 |
| filed under | unclassified |
| 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, so bit (a) fails. The proposed continuum optimization is a plausible extension, but the paper already begins with mean-field population proportions and does not provide a discrete computational object awaiting continuization. The opponent therefore correctly establishes that it cannot qualify under the anchor rule.
fails bit a — no named computational result to mirror
The candidate covers heterogeneous swarm composition, zealot robustness, regret, cognitive cost, and Bayes risk in equations (6)–(9) and Figures 3–4; it leaves aside the paper’s non-computational ODE analysis, homogeneous baselines, and finite-swarm or network extensions.
The strongest honest positive case is conditional—and weak under the programme’s anchor rule. The supplied paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a computational result. It gives mean-field ODEs in equations (1)–(7), defines performance metrics in (8)–(9), numerically integrates them, and reports fitted trends and optimal compositions in Figures 3–5. None is a named complexity or algorithmic result. Thus there is no qualifying anchor to quote; inventing one from Figure 4’s \(k^\ast\) would fail the stated standard.
That said, the paper is an unusually good source model for a new computational mirror. Its own central variable is already a population fraction: \(k\) is the mass of susceptible robots using the voter rule, with \(1-k\) using the majority rule. A natural regime is a mass-produced fleet with \(S\gg\tau\) well-mixed robots and only a few firmware cohorts: voter/direct-switch, voter/cross-inhibition, majority/direct-switch, and majority/cross-inhibition. Robots in one cohort share firmware, sampling effort \(G\), communication cost, and update rule. Zealots are additional high-multiplicity cohorts with fixed opinion \(A\) or \(B\). Rational masses \(\mu_t\) represent fractions of the fleet; clearing denominators recovers a finite fleet of cloned robots.
The lead candidate would be Robust Continuum Swarm-Composition.
An instance consists of a finite catalogue \(B\) of behavioural types, rational option quality \(q\), threshold \(\theta\), costs \(c_1,c_2\), initial state, and a finite rationally weighted set of attack scenarios \(a=(z_a,z_b)\), including wrong-addressing and denial-of-service attacks. A solution is a composition \(x\in\Delta(B)\), where \(x_b\) is the fraction of susceptible robots running type \(b\). For each attack scenario, the population state evolves according to the paper’s mean-field equations (1)–(7), with \(k\) replaced by the appropriate mass of voter-rule types. Define \(\tau_a(x)\) as the first time that total opinion \(A\) or \(B\) reaches \(\theta\), assign deadlocks a specified finite penalty, define \(\rho_a(x)\) exactly as in Section 4.1, and set \(\phi_a(x)=\tau_a(x)\sum_{b\in B}x_bG_b\). The problem is to output a composition minimizing
\[ \sum_a \lambda_a\bigl(c_1\phi_a(x)+c_2\rho_a(x)\bigr), \]
or decide whether this value is at most a given rational bound.
This is recognisably the paper’s question: it preserves its two opinions, zealot attacks, heterogeneous rules, cognitive cost, regret, Bayes-risk objective, and the design variable \(k\). The authors would probably recognise it as turning their search over \(k\) into an explicit optimization problem. It is genuinely population-continuous because the decision variable is the distribution of a large robot population over behavioural types, not merely a fractional outcome.
My expectation is mixed. A fixed-catalogue, bounded-horizon, \(\varepsilon\)-approximation version may be Class A: the vector field is explicit and low-dimensional, and the two-rule case has only one composition parameter. The literal \(t\to\infty\) version with equilibrium selection, deadlock detection, and globally optimal composition is more plausibly a Class-C/open problem, because nonlinear asymptotic dynamics and basin structure introduce difficulty not inherited from finite-population multiplicity. There is no hardness-transfer argument available from this paper.
The weakest point is decisive: this is an extension of the paper’s model, not a mirror of a named computational theorem. Moreover, the paper is already mean-field and analytic, so the continuous population is its starting point rather than a transformation of its stated problem. The proposed mirror covers only the heterogeneity/cost-performance results surrounding equations (6)–(9) and Figures 3–4; it does not cover the whole paper, and under a strict named-anchor requirement the correct conclusion is that the paper supplies no admissible positive anchor.
The negative case begins with the anchor rule: there is no qualifying anchor. The paper contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. Equations (1)–(7) are mean-field ODEs; equations (8)–(9) define performance metrics; Figures 3–5 report numerical sweeps and fitted trends. “Heterogeneous swarms can outperform homogeneous ones” and “there is an optimal \(k^\ast\)” are empirical modelling conclusions, not computational results to be mirrored.
The proposed Robust Continuum Swarm-Composition problem does not repair that defect. Its central variable \(x\) is already present in the paper in the simpler form \(k\): \(k\) is explicitly the fraction of susceptible robots using the voter rule, with \(1-k\) using the majority rule. The state variables \(A,B,U\) and the zealot parameters \(z_a,z_b\) are also already population proportions. Thus the paper has already performed the population-level mean-field passage. Replacing \(k\) by a distribution over a fixed catalogue of firmware cohorts is a generalisation of the model, not its continuization.
That generalisation is sensible, but it is a new nonlinear design or optimal-control problem. With the paper’s finite catalogue, the decision variable lies in a fixed-dimensional simplex, and the dynamics have one finite-dimensional ODE block per behavioural type. Clearing denominators merely gives a finite swarm whose relevant state is the vector of cohort counts. There is no high-multiplicity phenomenon to expose: no individual-level targeting, mass transfer, exponentially large type space, pricing problem, or discrete computational bottleneck that the continuous formulation clarifies.
The stronger versions do not help. One could let types include firmware, sampling effort, initial opinion, zealotry status, and communication cost; optimize over all compositions; average over attack scenarios; or impose worst-case attacks and initial conditions. Those are legitimate robustness-design questions, but their difficulty comes from choosing protocols, analysing nonlinear asymptotics, or solving robust control problems—not from replacing an individuated population by a distribution. Alternatively, one could restore finite-swarm stochasticity and optimize the probability of reaching the quorum. That would no longer be determined by \(\mu\) alone: finite-size fluctuations, network structure, stopping rules, and \(S\) become essential. A fixed positive zealot mass survives the mean-field limit, while a single stubborn robot has vanishing mass, so the continuum either reproduces the paper’s existing ODE model or changes the robustness question.
Nor do the paper’s three natural “anchors” become computational anchors after relabelling them. The cost–regret trade-off is already an aggregate objective; the wrong-addressing and denial-of-service analyses already use continuous parameters \(q,z_a,z_b\); and the optimal composition \(k^\ast\) is a numerical minimizer of an already-continuous mean-field model. Turning those plots into an optimization problem may produce a worthwhile robotics paper, but it does not produce a continuous mirror of a computational-social-choice result.
This is not the mistaken claim that continuity “does not help,” nor the claim that the mean-field analysis counts as prior computational work. The sharper point is that the paper supplies no discrete computational object awaiting continuization, while its population continuum is already the paper’s starting formalism. A broad research programme could certainly build a new continuum optimization problem around it. Under the stated anchor standard, however, that is an adjacent control/design programme, not a worthwhile ChoCo mirror.
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.