| paper | Fostering Collective Action in Complex Societies Using Community-Based Agents |
| authors | Jonathan Skaggs, Michael Richards, Melissa Morris, Michael A. Goodrich, Jacob W. Crandall |
| venue | IJCAI 2024 |
| filed under | coalition · hedonic |
| judged by | gpt-5.6-luna / xhigh (triple__luna__xhigh__c2r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper contains no numbered theorem, lemma, corollary, or proposition asserting complexity or solvability of a computational problem. Algorithm 1 is a behavioural procedure, while the CAB–CAT results are simulations and experiments. The proposed continuous CAT-resilience problem is a new re-modeling rather than a mirror of a qualifying paper result.
fails bit a — no named computational result to mirror
No qualifying computational result is covered. The proposed question addresses only CAT resilience in Sections 5.1–5.2 and leaves CAB learning, community detection, human studies, and empirical comparisons aside.
The strongest honest answer is that this paper has no qualifying anchor. It contains no numbered Theorem, Lemma, Corollary, or Proposition, and it makes no named claim of NP-hardness, polynomial-time solvability, W[1]-hardness, or FPT status. “Algorithm 1” is a procedure description, not a computational-complexity result. The claims in Sections 4–5 and Table 3 are simulation and user-study findings, not formal named results. Thus I cannot truthfully quote an anchor or claim that a continuous problem mirrors one of the paper’s named computational theorems.
The closest positive case comes from Section 5.2’s empirical finding that hand-coded CAB agents can collectively withstand CAT agents, whereas evolved CAB policies often fail. This is not an admissible anchor under the brief, but it suggests a credible continuous problem.
My lead candidate would be Continuum CAT-Resilience. Consider a large online society whose users fall into finitely many complete behavioral types \(T=T_C\cup T_A\). A type records initial popularity, CAB or CAT policy parameters, observability, token budget, and interaction block. The society is given by rational masses \(\mu_t\), with \(n\gg |T|\). This could represent millions of users instantiated from a small number of recurring behavioral templates, rather than eight individually distinct players.
For each type \(t\), let \(k_t^\tau\) be the fraction of tokens kept in round \(\tau\), and let \(g_{u,t}^\tau,a_{u,t}^\tau\) be the per-recipient densities of tokens given to or taken from type \(u\). The type-symmetric feasibility condition is \(k_t^\tau+\sum_{u\ne t}\mu_u(g_{u,t}^\tau+a_{u,t}^\tau)=1\). Aggregate attack strength on type \(u\) is computed by replacing the player-level sums in Equation (4) with mass-weighted sums over types; popularity and influence then evolve by the mass-weighted versions of Equations (1)–(3). CAB types follow fixed parameterized CAB policies, while CAT types may choose admissible target and attack allocations.
Given \(T\), \(\mu\), initial popularities and influences, the JHG constants, a horizon \(H\), a terminal threshold \(\lambda\), and cooperative CAB policies, the problem asks for the smallest adversarial mass \(\eta\) such that CAT types can force some cooperative type’s terminal popularity below \(\lambda\). Equivalently, its decision form asks whether there exists a type-level CAT policy and attack-flow sequence with total CAT mass at most \(\eta\) that achieves this. A solution consists of \(\eta\), the type-level attack policy, and the resulting certificate trajectory. Setting \(\lambda=0\) gives the direct analogue of the paper’s “overpowering” experiment.
This is recognisably the paper’s question: it preserves popularity-dependent power, positive and negative token transfers, defensive keeping, collective retaliation, CAT coalitions, and the distinction between group formation and threat mitigation. It covers only the CAT-resilience part of Sections 5.1–5.2, not the entire CAB-learning programme.
I would tentatively expect the fixed-policy, fixed-horizon, finite-type version to have tractable subcases: once the population is compressed to finitely many blocks, it becomes a finite-dimensional piecewise nonlinear optimization problem. The general policy-synthesis version is more likely to be continuum-specifically hard, because it combines mass allocation, attack-defense ratios, state-dependent policy choices, and type-level community structure. If arbitrary individual influence rows are retained, however, the construction ceases to be a genuine high-multiplicity mirror and drifts back toward discrete network-game hardness.
The weakest point is decisive: the original paper’s core object is relational and identity-sensitive. Its influence matrix is \(n\times n\), and CAB uses individual-level observations and Louvain communities. Collapsing users into finitely many types replaces that network with a block model and may remove exactly the community dynamics the authors care about. Moreover, the paper studies societies of roughly ten players, not an established high-multiplicity regime. So this is a plausible new research problem inspired by the paper, but not a continuous mirror anchored in one of its named computational results.
The proponent’s concession is the decisive negative point: this paper has no qualifying computational anchor. It contains no theorem, lemma, corollary, or proposition asserting the complexity or solvability of a computational problem. Algorithm 1 is a behavioural policy, not an algorithm solving a formally posed optimization or decision problem. The CAB–human experiments and the CAT experiments are empirical comparisons, not named computational results. Consequently, there is no paper result whose continuous high-multiplicity counterpart ChoCo could classify.
The proposed Continuum CAT-Resilience problem is therefore a new problem inspired by Section 5, not a mirror of it. The paper never asks for the minimum adversarial mass, gives no terminal threshold, and does not formulate policy synthesis or an optimization over CAT strategies. Its experiment fixes eight CAB agents and two CATs, then compares training procedures. Recasting that comparison as a mass-minimization problem may be interesting, but it cannot serve as the requested anchor.
More fundamentally, the proposed aggregation does not preserve the JHG’s central object. A type in the ChoCo sense must capture everything relevant to an agent’s behaviour. In JHG, that includes an individual’s influence row, incoming and outgoing token history, visibility relations, popularity, and position in the evolving network. Two agents with the same CAB parameters are not the same type if they occupy different relational positions. If those positions are included in the type, almost every agent becomes its own type and the high-multiplicity gain disappears. If they are omitted, the model is no longer the paper’s JHG.
A finite block model or graphon variant avoids that problem only by changing the subject. It would study a mean-field or block-structured network game in which community formation, retaliation, and Louvain partitions are replaced by aggregate interaction patterns. That may be a legitimate new model, but it is not a continuous mirror of the paper’s identity-sensitive dynamics. Allowing a continuum of locations or network coordinates preserves the relational structure, but then agents are individuated by those coordinates rather than represented by finitely many repeated types; the problem becomes a continuum network-control model, with a different input representation and computational theory.
Thus the proponent’s own “weakest point” is not merely a technical caveat. It is the fundamental obstruction: preserving multiplicity destroys the network structure, while preserving the network structure destroys multiplicity. The proposed finite-dimensional nonlinear control problem could still be worthwhile on its own, but nothing in this paper establishes it as a ChoCo problem, and its conjectured tractability or hardness cannot repair that missing connection.
The absolute claim that no imaginative re-modelling could ever produce a worthwhile continuous question is not provable; a mean-field JHG-inspired control problem is conceivable. But the strongest honest verdict is negative: this paper supplies no computational result to mirror, no credible high-multiplicity regime for its defining dynamics, and no proposed continuous formulation that remains recognisably the same problem.
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.