| paper | d4-conf-ijcai-2021-0184 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The proposed mixture-MAP task is concrete and preserves the paper’s broad aim of inferring task labels while modeling worker bias. But the paper has no numbered theorem, lemma, corollary, or proposition asserting a computational result, and the proponent explicitly acknowledges that it supplies no eligible anchor; this formulation is a follow-on, not a continuization of a result in the paper.
fails bit a — no named computational result to mirror
The candidate follows Section 3’s latent label-generation and inference model; it does not mirror Sections 4–5’s empirical and simulation findings as computational guarantees.
There is a plausible population mirror of the paper’s label aggregation problem. In a large, repeated crowdsourcing campaign, let worker type \(t\) be a stable behavioral profile \((c_t,p_t)\): the worker’s value and degree of confirmation bias. Let \(\mu_t\) be that type’s fraction of the worker population, with \(N\gg\tau\). For task \(j\), keep the paper’s task value \(s_j\), true label \(z_j\), global bias parameter \(a\), and prior \(\pi\). A type-\(t\) worker reports label \(0\) with probability
\[
q_{tj}(z_j)=\exp\!\left(-a\left[(1-p_t)(s_j-c_t)^2+p_tz_j\right]\right).
\]
Thus mass describes how common each worker profile is, while labels remain binary.
A candidate problem, Continuous Confirmation-Bias Aggregation, would take \(M\), a bound \(\tau\), and the observed population distribution \(\nu\) over workers’ complete response vectors \(r\in\{0,1\}^M\). It would seek the task labels \(z\), the type masses and profiles, the task values, and \(a,\pi\) that maximize the model’s joint log-likelihood:
\[
\sum_r \nu_r\log\!\left(\sum_t\mu_t\prod_{j=1}^{M}
q_{tj}(z_j)^{1-r_j}\bigl(1-q_{tj}(z_j)\bigr)^{r_j}\right)
+\sum_{j=1}^{M}\left((1-z_j)\log\pi+z_j\log(1-\pi)\right).
\]
This preserves the paper’s central inference question—recovering ground-truth labels while accounting for workers’ confirmation bias—and keeps the task outcomes discrete. The plausible regime is a large pool completing standardized tasks, with workers clustering into relatively few stable behavioral profiles. The mirror covers the model and inference procedure in Section 3, not the empirical findings as computational complexity results.
The weakness is substantial: the paper estimates separate latent parameters for individual workers, and neither its MTurk study nor its simulations establish that real workers fall into a small number of shared profiles. That regime is plausible for large repeated campaigns, but it is an additional modeling assumption. The continuous mixture-MAP question above is therefore a candidate for a follow-on complexity study, not a result this paper establishes.
In particular, the paper contains no numbered Theorem, Lemma, Corollary, or Proposition asserting a computational complexity result. It presents an EM algorithm and reports empirical and simulation comparisons; those are not eligible anchors under the requested standard. So I cannot honestly provide an anchor-based positive case, assign a tractability or hardness-transfer classification to an anchor, or claim that the paper’s authors proved a complexity result for this mirror.
The paper has no named computational result to continuize. Section 3 gives a probabilistic model and an EM procedure; Sections 4–5 report accuracy comparisons. The paper proves no runtime, optimality, approximation, or hardness result for label aggregation. So the proponent’s mixture-MAP problem is a proposed follow-on, not a mirror of a computational result in this paper.
The proposed problem also shifts where the population appears. ChoCo’s central model takes a distribution over complete agent types as input. Here the observed input is a distribution over response vectors, while worker masses and profiles \((c_t,p_t)\) are hidden variables in the optimization. Supplying the profiles and the other model parameters would make label inference separate task by task. Keeping them hidden preserves the paper’s harder inference problem, but turns the mirror into latent mixture estimation with a nonconvex likelihood. The paper’s local EM procedure neither specifies how to find a global optimum nor establishes that the model parameters and true labels are identifiable. And response vectors can have exponentially many possible values as the number of tasks grows, so their empirical distribution need not be a compact representation of a large worker population.
That is the strongest case against treating this paper as a ChoCo mirror: it supplies no complexity anchor, and the proposed candidate moves the computational challenge into a new latent-variable problem rather than showing what continuization does to a stated result. But the universal negative remains weak. A large, recurring crowd with a small number of stable behavioral profiles is plausible, and studying exact inference from that population could be worthwhile. The honest conclusion is that this paper does not establish a mirror worth prioritizing—not that no worthwhile mirror can be formulated in any scenario.
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.