| paper | d4-conf-aaai-2019-0160 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
Neither Theorem 1 nor Theorem 3 states an algorithmic or complexity result. Mass-DBT Seed Design is a statable population-level extension, but its added optimization choices cannot supply the missing computational character of either anchor.
fails bit a — no named computational result to mirror
Theorem 1 covers pointwise DUSIC and positive rewards under proficiency and payment conditions; Theorem 3 covers per-worker fairness. Neither gives a runtime, hardness bound, or population-level design algorithm; the proposed seed-design optimization is new.
The strongest positive setting is a very large platform repeatedly collecting answers to the same family of microtasks—for example, binary labels for text items. Let a worker type \(\theta\) include the complete proficiency matrix \(A_\theta\), effort cost, and every other worker-specific parameter the mechanism uses. A society is a finite distribution \(\mu\) over such types, with \(\mu_\theta\) the fraction of workers of type \(\theta\). This is plausible for large, recurring crowds with a small number of genuinely shared worker profiles, though less so for a one-off crowd with highly individualized abilities.
The nearest named result is Theorem 1, proved in this paper, with its proofs placed in the supplement. It says that if
\[
\beta > \frac{C^E}{\left(\sum_{g\in[K]} A_i[g,g]\right)-1}
\]
and \(A_i[g,g] > A_i[g',g]\) for every \(g'\ne g\), then the Deep Bayesian Trust mechanism is dominant uniform strategy incentive compatible for worker \(i\), and truthful play gives that worker strictly positive expected reward. Theorem 3, also proved here, says that the mechanism is fair under the paper’s definition: expected reward is directly proportional to a worker’s reported-answer accuracy and independent of the strategy and proficiency of the random peer. These results make a continuous-population incentive-design question plausible. They are not computational-complexity results, however. Neither says that a problem is in P, is hard, or has some other complexity classification. The paper gives no named computational result of the kind the requested anchors require.
If a new computational question motivated by those theorems is still useful, I would formulate Mass-DBT Seed Design as follows. An instance gives \(K\), a fully mixed rational task prior \(P\), a finite list of worker types with rational masses, proficiency matrices and effort costs, a gold oracle, allowable peer links with shared-task capacities, required task coverage for each type, and a bound on gold-task work and per-worker payments. The question is whether there is a rational assignment of worker mass across peer links and rounds, together with a payment scale \(\beta\), that meets the coverage and resource bounds, makes each type’s trustworthiness identifiable from its assigned shared-task distributions, and gives truthful uniform strategy as a strict best response for every type covered by Theorem 1. The output is the assignment and payment policy, with witnesses for identification and the incentive conditions.
For explicitly listed, fixed types and precomputed allowable links, I would expect this design question to reduce to a finite flow or linear optimization problem. That is a conjecture about this new problem, not a conclusion of the paper. A simpler exercise—apply the paper’s existing mechanism separately to each type and check its theorem conditions—would aggregate naturally, but would add little computational content.
That is the soft spot, and it is decisive under the requested anchor rule: Mass-DBT Seed Design is an extension, not a problem whose computational complexity is established by a named result here. The paper’s own asymptotic step is the number of shared tasks per worker-peer pair tending to infinity; it is not a continuization of worker population size. So the honest positive case is that the setting supports a plausible population-level mechanism-design problem, not that this paper already supplies a qualifying anchor for a continuization programme.
The clearest objection is that Theorems 1 and 3 are not computational results. Theorem 1 proves a dominant-strategy and positive-reward guarantee under conditions on a worker’s proficiency and payment; Theorem 3 proves the mechanism’s fairness property. Neither establishes an algorithm, runtime, hardness result, or computational optimization problem. So neither supplies the kind of result a continuous-computational mirror is meant to chart.
The best population version of Theorem 1 would give workers’ static characteristics—proficiency matrices and effort costs—to types, then require the theorem’s conditions for almost every type in a distribution \(\mu\). But the conditions remain pointwise. For a fixed \(\beta\), one checks them type by type; choosing a common \(\beta\) amounts to taking the largest required threshold over the covered types. That can be a sensible design exercise, but it is not a computational question established by this paper. The more ambitious “Mass-DBT Seed Design” adds assignment, link, coverage, and resource decisions. Those may lead to a worthwhile new problem, but the paper neither defines nor solves it. Moreover, mass over proficiency types does not by itself ensure identifiable trustworthiness: that depends on workers’ strategic reports and on the shared-task data used to infer them. The paper’s limiting assumption is that each worker-peer pair shares infinitely many tasks; increasing the worker population does not supply those observations.
Theorem 3 fares no better as an anchor. It gives a per-worker statement about expected reward and accuracy, independent of the peer’s strategy and proficiency. Integrating that statement over \(\mu\) gives an average of individual guarantees; it creates no population-level decision problem. A planner could optimize payments or coverage subject to fairness constraints, but that would again add the computational problem from outside the paper.
There is a plausible high-multiplicity setting here: large platforms may repeatedly use crowds with recurring worker profiles. So I would not claim that crowdsourcing admits no useful population model. The stronger, narrower conclusion is that this paper offers no named computational result for such a model to mirror. Any worthwhile type-mass optimization would be a new extension, and the paper’s own asymptotic axis is shared tasks per worker-peer pair, not population size. The case against is strong as a reason to exclude this paper as a continuization mirror; it is weaker as a claim that no adjacent research direction could be worthwhile.
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.