| paper | d4-conf-atal-2013-0218 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
Theorem 1 is a named computational result, but the paper already models responders as IID draws from task-specific population distributions. The proposed masses and participation weights simply induce those same distributions, while the guarantee depends on their aggregate means and does not use the type structure. This is a \(b\)-already case: the proposed mirror restates the paper’s existing model.
fails bit b-already — the paper is already native to the continuum
Theorem 1 applies to every IID response law \(X_k\) satisfying the margin assumption, so introducing \(\mu_s\) and \(a_{s,k}\) only reparameterizes a law already covered by the paper; the proponent notes this overlap but gives no distinct population-dependent computational question.
fatal: True
The proposed type formulation restates Theorem 1’s error guarantee; it leaves Theorem 2, Corollary 3, Theorem 4, and the experimental comparisons unmirrored.
The strongest mirror is a continuous worker population behind the paper’s budgeted classification problem. Imagine a large open marketplace processing many binary microtasks. There may be hundreds of thousands of workers but only a modest number \(\tau\) of recurring behavioral profiles: workers with the same task-specific response tendencies and chances of entering each task’s responder pool. A type \(s\) has mass \(\mu_s\), with \(\sum_s\mu_s=1\). For task \(k\), let \(a_{s,k}\) be its type-specific participation weight and \(p_{s,k}\) the probability that a type-\(s\) worker reports \(1\). A requested response draws a worker type independently from the task’s population, with probability proportional to \(\mu_s a_{s,k}\), and then draws that worker’s Bernoulli response. This is a distribution over the workforce, not a lottery over outcomes or a noise model over a finite profile.
Continuous CrowdBudget Guarantee is the anchored problem. Its instance consists of \(K\) binary tasks, a budget \(B\), rational task costs \(c_k>0\), and a continuous society \(\mu\) over \(\tau\) worker types with the response and participation parameters above. Each task has a hidden truth \(y_k\in\{0,1\}\). Write \(\bar p_k\) for the mean response to task \(k\) induced by the type masses and participation weights, and promise \(|\bar p_k-y_k|<1/2\). The decision variable is an integer number \(n_k\geq 0\) of independently sampled responses requested for each task, with \(\sum_k c_kn_k\leq B\); the estimate is majority vote. The underlying objective is to minimize the sum of the tasks’ error probabilities. The guarantee version asks for a feasible allocation, computed from the public allocation data \(K,B,(c_k)\), whose error obeys the bound below for every promised worker society. As in the paper, the allocation rule need not know the hidden truths or worker accuracies.
My lead anchor is Theorem 1 (Main result 1), proved in this paper. It states that, for \(B\geq\sum_j c_j\), CrowdBudget’s expected total estimation error is at most
\[
K\exp\!\left(
-\frac{2B d_{\min}^{\,2}}
{c_{\max}^{\,2}\sum_{j=1}^{K}1/c_j}
\right),
\]
where \(c_{\max}=\max_j c_j\) and \(d_{\min}=\min_j|\bar p_j-\tfrac12|\). In the continuous mirror, the type masses and task participation weights induce each task’s responder distribution; its mean is exactly the \(\mu_k\) used in the paper’s analysis. Independent sampling and the same margin assumption therefore give the same concentration argument. CrowdBudget’s allocation depends on \(B\) and the task costs, so the guarantee problem has a direct constructive answer and is tractable. This is a guarantee for a feasible allocation, not a proof that it finds the exact minimum-error allocation. Exact optimization is a natural further question, and I would not infer its complexity from Theorem 1.
The authors would plausibly recognize this as their problem: the mirror retains task-specific response costs, the open crowd, IID responders, and majority aggregation. It makes explicit that many workers can share a behavioral type and that the workforce is summarized by type masses. This scenario fits platforms with large responder pools and repeated microtasks, where a small set of skill and participation profiles is plausible. It does not require selecting named workers or assigning different prices to individual identities.
The paper proves Theorem 1 here; it is not a result cited from earlier work. It also has named results on PAC error bounds and a uniform-cost specialization, but I would keep this case to Theorem 1: it supplies the cleanest anchor. The paper contains no named NP-hardness or complexity-class result to anchor. Its result is an algorithmic performance guarantee, so the expected direction for this guarantee problem is tractability. Further questions include exact allocation optimization, task-specific type mixtures with correlated responses, and whether type structure can yield guarantees stronger than one that depends only on aggregate response means.
The scope is deliberately narrow: it mirrors CrowdBudget’s binary, IID, margin-assumption result. It does not claim to continuize all the paper’s fusion methods, experiments, or future work.
The soft spot is substantial. The paper already describes responders using task-specific distributions \(X_k\), and Theorem 1 depends only on each task’s aggregate mean response, not on the number or structure of worker types. So the proposed type-mass society may look like an explicit high-multiplicity description of a distribution the paper already uses, rather than a new source of computational structure. That weakens the novelty of this particular mirror. The positive case rests on the distinction between a sampling law over individual users and an instance represented directly by masses of behaviorally identical worker types.
Theorem 1 is the proponent’s only anchor, and it is a real algorithmic result: it gives CrowdBudget an error guarantee. So the case against should not rest on saying the paper lacks a computational result. The problem is that its proposed continuous version is already inside the paper’s model.
For task \(k\), the proposed type masses and participation weights induce a responder law
\[
X_k(s)=\frac{\mu_s a_{s,k}}{\sum_r \mu_r a_{r,k}}.
\]
Because each response is sampled independently, the response is Bernoulli with mean
\[
q_k=\sum_s X_k(s)p_{s,k}.
\]
That is exactly the aggregate mean used in the paper’s concentration argument. The paper already models responders as IID draws from an unknown task-specific distribution \(X_k\), with user-specific response distributions \(D(u,k)\). Its assumptions and Theorem 1 cover the proposed worker populations, even though the theorem does not list type masses as input. CrowdBudget ignores those masses and guarantees the same bound for every response law satisfying the margin assumption. Making the law explicit therefore restates a case the theorem already covers; it adds no population-dependent algorithmic question.
A stronger proposal would provide the type data and ask for the exact minimum-error allocation. But the paper already formulates that objective in Equation (5). Once the worker parameters are given, the IID model reduces each task to its mean \(q_k\); its error as a function of \(n_k\) is a binomial tail. The resulting optimization is over the \(K\) task allocations and their costs. It may be a worthwhile allocation problem, but the support size \(\tau\) and the worker masses disappear after computing the means. Any computational challenge lives in the task-allocation problem, not in continuizing the workforce.
Worker masses could affect decisions if workers had finite capacities, could be assigned to particular tasks, or had to be coordinated across tasks. That would introduce type-level allocation variables and capacity constraints. The paper explicitly assumes IID responders from an open crowd and defers explicit user assignment to future work; those additions change the model that Theorem 1 analyzes. They could support a worthwhile separate project, but the proponent’s anchor does not establish it as a continuous mirror of CrowdBudget. The negative case is strongest on that boundary; it cannot rule out every useful future model of typed, capacity-limited workers.
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.