| paper | Dynamic Coalition Structure Detection in Natural-Language-based Interactions |
| authors | — |
| venue | AAMAS 2025 |
| 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 hardness, tractability, parameterized complexity, or an approximation guarantee. Problem 1 is only a definition, while the reported F1, MRR, and Brier results are empirical evaluations. The proposed mirror aggregates repeated Diplomacy games and forecasts, changing the paper’s object rather than continuizing a named computational result.
fails bit a — no named computational result to mirror
The proposed construction would cover only the paper’s empirical coalition-prediction task; it leaves agreement detection, strategic-value estimation, and the learned models without computational results.
The strongest honest case is narrow: a credible population mirror exists for the paper’s formal prediction task, but the paper supplies no eligible computational anchor.
There is no Theorem, Lemma, Corollary, or Proposition in the paper asserting tractability, hardness, or parameterized complexity. “Problem 1” is a problem definition, not such a result. Definitions 1–3 and Tables 1–2 likewise do not qualify. The statement that enumerative approaches are “intractable” is informal and attributed to cited work [4], not proved here. Thus, strictly under the programme’s rules, this paper has no named computational result to anchor.
The best conditional mirror would be Continuum Dynamic Coalition Prediction. Consider a platform running a very large number of simultaneous Diplomacy-like negotiation games, or many repeated strategic coordination cells in an operational setting. The individual games still have discrete players, units, actions, and agreements; the population is continuous only across the many player-rounds or game instances.
A type \( \theta \) is a complete player-round context: the player’s role, board state, units, bounded semantic dialogue history, candidate agreements, policy distribution, subjective belief about the opponent, and value model. Two agents have the same type only when all information used by the prediction method is identical. Let \( \mu_\theta \) be the fraction of active player-rounds of type \( \theta \), with \( \sum_\theta \mu_\theta=1 \). A matching kernel \( K_{\theta,\theta'} \) records the mass of interactions between types \( \theta \) and \( \theta' \), with its marginals given by \( \mu \).
For a candidate agreement \( \alpha=(u_1,u_2,a_1,a_2) \), the type-level input supplies the paper’s agreement-detection score and strategic values. The latter combine as \( r_{\theta,\theta',\alpha}=(V_\theta(\alpha)V_\theta^{\theta'}(\alpha))(V_{\theta'}(\alpha)V_{\theta'}^{\theta}(\alpha)) \), mirroring the paper’s \( wt_i(\alpha)wt_j(\alpha) \). After calibration, let \( p_{\theta,\theta',\alpha} \) be the predicted probability that this agreement is both detected and honored.
The decision variable is \( x_{\theta,\theta',\alpha} \), the mass of the population whose coalition graph contains that agreement. The problem is: given \( \mu \), \( K \), the finite agreement sets, and the type-level detection and rationalizability scores, compute the weighted type-level coalition structure minimizing expected Brier loss, \( \sum_{\theta,\theta',\alpha}K_{\theta,\theta'}\mathbb E[(x_{\theta,\theta',\alpha}/K_{\theta,\theta'}-Y_{\theta,\theta',\alpha})^2] \), where \(Y\) indicates whether the agreement is actually honored. Its solution is the aggregate forecast \(x_{\theta,\theta',\alpha}=K_{\theta,\theta'}p_{\theta,\theta',\alpha}\), together with the ranking of agreements induced by the rationalizability score.
This is genuinely a population continuization: an individual agreement remains discrete, while \(x\) records the fraction of many agents or game instances exhibiting that edge. It preserves the paper’s two central objects—natural-language agreement detection and subjective-rationalizability scoring—rather than replacing them with fractional outcomes. The authors would likely recognize it as the high-multiplicity deployment version of their Problem 1, especially because they explicitly describe the method as generalizable to other dialogue-based multi-agent environments.
I would expect this narrowly defined problem to be Class A once the type-level neural and game-theoretic scores are supplied: aggregation requires time polynomial in \( |\Theta| \), the number of interaction types, and the number of agreement schemas. A more ambitious version that enforces consistency among conflicting agreements would become a configuration LP or matching problem, with a meaningful pricing question over compatible local coalition structures.
The weakest point is substantial. This mirror is an ensemble of many repeated games, not a single seven-player Diplomacy board. Ordinary human transcripts may also be nearly all distinct, making \( |\Theta| \) comparable to the population size. The high-multiplicity regime is most convincing for templated or automated agents, repeated operational scenarios, or a platform with a finite library of semantic dialogue states. Moreover, the paper gives no complexity theorem for computing \(V_i\), \(V_i^j\), or the language-model detector itself. Therefore this is a plausible conditional mirror of Problem 1, not a theorem-backed continuization result of the paper.
The proponent’s strongest point is also the decisive objection: this paper has no eligible computational anchor. “Problem 1” is a prediction task, not a theorem, lemma, corollary, or complexity result. The claim that enumeration is “intractable” is informal and attributed to prior work. The actual results are empirical F1, MRR, and Brier-score comparisons for learned predictors. There is therefore no named discrete computational result whose high-multiplicity relaxation could contribute to ChoCo’s complexity landscape.
Even granting a broader remit, the proposed mirror is not a continuous version of the paper’s problem. Diplomacy itself has exactly seven players. The paper predicts the coalition graph in one particular game state from one particular history. Treating many independent games as a population changes the object: it becomes aggregate calibration across replicated experiments, not coalition detection in a society.
The proposed type also exposes the problem. A complete type must include the board state, unit configuration, private dialogue, action history, subjective beliefs, candidate agreements, and model outputs. With natural-language histories, exact types are generally nearly all distinct, or even drawn from an unbounded space. If one compresses them into a finite semantic library, one loses information that the paper treats as central. If one instead constructs many exact replicas, the high multiplicity is experimentally imposed rather than a meaningful population regime of the original interaction.
The proposed optimization is computationally empty once its hard inputs are supplied. Given \(p_{\theta,\theta',\alpha}\), the expected Brier objective is minimized coordinatewise by
\[ x_{\theta,\theta',\alpha} = K_{\theta,\theta'}p_{\theta,\theta',\alpha}, \]
and producing the ranking is sorting. The substantive work—language understanding, subjective-belief estimation, and strategic value computation—remains the original black-box LLM and reinforcement-learning pipeline. It is not an LP whose pricing problem is exposed or simplified by continuization.
A more careful mirror could impose consistency among agreements by optimizing over distributions of complete coalition graphs rather than independent edge marginals. That would avoid impossible combinations of agreements, but it confirms the mismatch: each feasible graph still belongs to one finite Diplomacy instance, and the aggregate merely mixes independent instances. The problem is then structured prediction or probabilistic game analysis, not continuous computational social choice. A pairwise matching kernel \(K\) is also insufficient to preserve the higher-order correlations among seven players, multiple units, and a shared board state.
A deployment platform with many repeated Diplomacy-like games could certainly benefit from population-level forecasting. That is a plausible machine-learning application, but it is not a worthwhile ChoCo mirror unless one adds an entirely new model in which a continuum of agents jointly interacts through aggregate state. Such a mean-field game might be interesting, but it would be a new research problem, not a continuization of any computational result in this paper.
Thus the negative case is strong within the programme’s stated scope: there is no eligible anchor, no genuine population regime in the paper’s object, and the suggested aggregate objective is either trivial or shifts the problem elsewhere. The universal claim against every imaginable re-modelling is not mathematically provable—one can always construct a useful ensemble model—but that possibility does not rescue this paper as a 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.