| paper | d4-conf-ijcai-2016-0497 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The proposed distribution over standardized belief-and-utility types gives a concrete, recognizable continuous problem, which the opponent concedes. But Proposition 5 proves replication invariance, while the weighted LP and its complexity are proposed new work rather than a computational result of this paper. The required named computational anchor is absent, so the grade is red.
fails bit a — no named computational result to mirror
The proposed question covers weighted generalized-belief computation and cautious-preference decisions for explicitly listed Omega; it does not computationally mirror the paper's other named propositions or Proposition 5's replication-invariance claim.
There is no named computational result in this paper that meets the requested anchor test: it states no numbered theorem, lemma, proposition, or corollary classifying a problem as tractable or hard, and it gives no numbered formulation that it computes with. The paper says generalized entailment can be solved by convex optimization, citing Potyka and Thimm (2015), but that claim is neither a numbered result here nor a complexity analysis of this paper’s group decision problem.
The closest relevant result is Proposition 5, proved in this paper (with proofs in its online appendix). It says that under the maximum norm, adding any number \(k\) of copies of an existing agent leaves the group belief \(B_G(\phi)\) unchanged for every formula \(\phi\); consequently, the group’s expected utilities are unchanged for every alternative. That is useful evidence that multiplicity is a meaningful dimension of their model, but it is not a computational anchor under the criteria given.
A plausible *new* continuous problem would be Continuous Probabilistic Belief-Merging Decision. Let an instance give a finite set of possible worlds \(\Omega\), alternatives and criteria, public integrity constraints, and a rational distribution \(\mu\) over a finite support of agent types. Each type \(t\) specifies a complete probabilistic knowledge base \(K_t\) and utility vector \(u_t\); its mass is the fraction of the population with exactly that profile. For a probability distribution \(P\) over \(\Omega\), let \(v_{t,j}(P)\) be the positive violation of the \(j\)-th conditional in \(K_t\). Among distributions satisfying the public constraints, minimize
\[
\sum_t \mu_t\sum_{j\in K_t} v_{t,j}(P).
\]
Use the minimizing distributions to compute the lower and upper group-belief bounds, then the group expected-utility intervals. The task is to return whether a specified alternative is cautiously preferred to another, meaning its lower expected utility is at least the other’s upper expected utility. With explicitly listed worlds and the \(1\)-norm, the violation minimization is an LP; computing the requested bounds is a tractability question for this precise continuous version.
A credible regime would be a large public consultation or organization whose participants share a modest number \(\tau\) of standardized belief-and-utility profiles. Here population size \(n\) is large relative to \(\tau\), and mass represents the fraction of participants in each profile. This preserves the paper’s core object: aggregating beliefs and utilities to derive group preferences. It is not a case for continuizing outcomes or preferences alone.
My strongest honest case is therefore a promising mirror question, but not an anchored positive case under your rules. The weakness is substantive: the paper’s named results are properties of the aggregation framework, not computational results about computing its decisions. The proposed LP tractability claim would need to be established as new work, and a succinct relational representation could change its complexity substantially.
The strongest case against is about anchoring. This paper states no named computational result about the complexity of its group-decision problem. Proposition 5 is a replication invariance result under the maximum norm, not a complexity result. It shows that adding copies has no effect in that case; it does not establish an algorithmic question the paper leaves open. So the proponent has no result-based anchor under the stated test.
Their proposed weighted problem is a reasonable computational question, but its first version is very direct: with explicitly listed possible worlds and the \(1\)-norm, minimizing weighted violations is an LP, and the belief bounds can be optimized over its solution set. For rational masses, this is exactly the finite model with agents repeated in the corresponding proportions, up to scaling the objective. That makes the contribution a high-multiplicity formulation and complexity analysis of probabilistic inference, rather than a new continuous phenomenon. The mirror’s practical case also depends on large groups sharing exact knowledge bases and utility vectors; coarsening those rich profiles would no longer preserve the paper’s notion of a type.
That attack does not defeat the stronger version. A large organization or consultation could elicit standardized belief-and-utility profiles, with many participants per profile. Then the support size \(\tau\) can be modest, the weighted LP is a useful high-multiplicity computation, and an explicit-world result is a legitimate tractable mirror. The paper’s model supplies the ingredients even though it supplies no complexity theorem. So the honest negative case is narrow: the paper is a poor *anchored* candidate, and the most obvious mirror may be algorithmically straightforward. I can’t make the universal claim that no worthwhile continuous mirror exists; the proposed standardized-profile regime survives.
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.