| paper | d4-conf-atal-2022-0107 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
Neither case identifies a numbered theorem, lemma, corollary, or proposition asserting a computational result. The feature-scoring algorithm and experimental findings do not meet that gate; optimizing population-level responses would add a computational problem the paper does not formulate.
fails bit a — no named computational result to mirror
No result is mirrored: the paper reports experimental effects of explanations and gives a heuristic for ranking features, but neither supplies a qualifying computational result.
The paper has no numbered Theorem, Lemma, Corollary, or Proposition asserting an algorithmic or complexity result. Algorithm 1 gives a heuristic for ranking feature based explanations, but it is not such a named result. The paper’s main findings are experimental: for example, it reports improved participant satisfaction and acceptance after explanations. Those findings cannot serve as the required computational anchors.
There is a plausible population regime behind the study: a large election with many voters but relatively few distinct ballot types. One could represent each participant by their ballot and their preference order over candidates, with \(\mu_t\) giving each type’s share. The voting features would then be computed from these masses. But asking which explanations maximize population acceptance or satisfaction would require a model of how each type responds to each explanation. This paper compares explanation methods experimentally; it does not formulate or analyze that optimization problem.
So there is no qualifying anchor, and I can’t state a per-anchor continuous problem without inventing one. A continuous study of explanation effectiveness may be worthwhile, but it would be a new computational question rather than a mirror of a named computational result in this paper.
The proponent raises no named result to defeat, and that is the strongest point against a mirror here. The paper’s main results are empirical: how explanations affect participants’ reported satisfaction and acceptance. Its algorithm for ranking feature based explanations is not a complexity result; it scans a finite set of candidates and features. The paper defines no computational problem whose complexity or algorithms it studies.
A large electorate with few ballot types is still a plausible setting. Replacing ballot counts by masses would give natural continuous versions of the voting features. But those masses alone do not determine how people react to an explanation. A stronger proposal would give each type a response model \(R_t(E)\), then choose explanations \(E\) to maximize average satisfaction, \(\sum_t \mu_t R_t(E)\). That is a coherent follow-on, and it connects to the paper’s suggestion of using machine learning to select explanations. But the paper supplies neither that response model nor an algorithmic selection problem. Defining them would add the central computational question rather than mirror a result in the paper.
The study’s purpose is to measure human responses under randomized presentation. A continuous distribution could express average responses across voter types, but that would make the population a statistical estimand; it would not by itself create the kind of computational problem the programme charts. The paper’s feature-scoring procedure can also take population-based feature values as input, but its computational task remains ranking a finite list of features.
So I would reject this paper as a continuous mirror under the programme’s current gate. The strongest adjacent direction is population-level optimization of explanations using a learned response model. That may be worthwhile in its own right, but the paper does not formulate it as a computational result. This is a strong case against mirroring the paper as written, not proof that no related continuous project could be valuable.
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.