| paper | d4-conf-ijcai-2020-0127 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper contains no numbered theorem, lemma, corollary, or proposition asserting an algorithmic, approximation, hardness, or solver-backed formulation result. Its MAP/ELBO derivation and empirical accuracy tables do not meet the stated computational-result gate, so no result can anchor a continuous mirror.
fails bit a — no named computational result to mirror
No result qualifies as an anchor; the paper's MAP objective, ELBO derivation, gradient-ascent training, empirical accuracy comparisons, and ablations remain without a computational mirror.
I can’t make a valid positive case under the anchor requirement: the paper contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. Its named contribution is the CLA model, and its results are empirical accuracy comparisons and ablation studies. The MAP objective and ELBO are formulations, not named complexity results.
A plausible follow-up would treat annotators as a population: each type is a complete annotator-response profile, its mass is its fraction of the annotator pool, and the task is to infer object labels from type-level response frequencies and object features. In a regime with many annotators but few distinct profiles, that could be a sensible high-multiplicity model. But it is a proposed extension, not a mirror anchored in a named result of this paper. I therefore have no qualifying continuous problem or expected complexity classification to offer.
The proponent’s strongest point is also the decisive one: this paper states no computational complexity result to mirror. Its equations define CLA’s MAP objective and derive an ELBO; the paper then trains the model by gradient ascent and reports empirical accuracy and ablations. It gives no theorem about the complexity, exactness, or approximation of an algorithmic problem. The accuracy comparisons are results, but they are not computational results in the programme’s sense.
The best proposed rescue is a population of annotators grouped by complete behavioral type: annotators with the same response mechanism, expertise, and other relevant properties. That could be a sensible high-multiplicity regime, and I would not reject it on the grounds that type-level modeling is inherently distorting. But asking for the complexity of inference in that regime would define a new type-aggregated crowd-labeling problem. The paper neither formulates that population model nor claims a computational result about it. A continuous population of objects is a less faithful rescue: CLA predicts a label for each indexed object using its features, so replacing those objects with mass changes the target of inference.
So the negative case is strong against a *mirror of this paper’s results*, but weaker as an assertion that no worthwhile adjacent project could be built from its ideas. A continuous annotator model might be worth studying; it does not supply an anchor in this paper.
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.