Eliciting Honest Information from Authors Using Sequential Review*

The strongest honest positive case is that this paper admits a faithful continuous-population version, but it does not satisfy ChoCo’s strict anchor requirement: it contains no named theorem asserting NP-hardness, membership in P, W[1]-hardness, FPT, or another computational-complexity classification. Its named results are mechanism-theoretic. Thus this is a plausible mirror, but not a Gate-A green case under the programme’s rules.

The lead result would be Proposition 4.9, proved by the authors here (with the proof deferred to the full version): the memoryless coin-flip mechanism is truthful whenever \(P_{\mathrm{acc}}\) is increasing and the continuation probability \(\rho\) is increasing. The more general supporting result is Theorem 4.7, also proved here, which gives monotonicity conditions for truthfulness of the whole sequential-review framework. I would not add Theorem 6.1 as a formal anchor: it is a useful incentive-comparison theorem, but it is not a computational result either.

A concrete mirror is Threshold Sequential-Review Design\(_\infty\). An instance contains:

• a finite set \(T\) of author-portfolio types;
• a rational population distribution \(\mu\) over \(T\);
• for each \(t\in T\), a portfolio size \(n_t\), ordered paper qualities \(q_{t,1}\ge\cdots\ge q_{t,n_t}\), an additive nonnegative nondecreasing author reward function \(u_t\), and rational review-score distributions on a finite score grid \(R\);
• a review-burden budget \(b\).

A type is a complete mechanism-relevant author description: portfolio size, quality vector, reward function, and review-noise law. Mass \(\mu_t\) is the fraction of authors of that type. The action available to an author is exactly the paper-ranking report in the paper. The designer chooses two thresholds \(r_{\mathrm{rev}}\le r_{\mathrm{acc}}\). The first reported paper is reviewed; a later paper is reviewed precisely when the preceding score is at least \(r_{\mathrm{rev}}\); and a reviewed paper is accepted precisely when its score is at least \(r_{\mathrm{acc}}\).

The objective is to choose thresholds maximizing expected accepted quality per submitted paper,

\[ U(\mu)= \frac{\sum_t\mu_t\, \mathbb E[\sum_i q_{t,i}\mathbf 1\{\text{paper }i\text{ accepted}\}]} {\sum_t\mu_t n_t}, \]

subject to expected review burden per submitted paper being at most \(b\),

\[ B(\mu)= \frac{\sum_t\mu_t\, \mathbb E[\#\text{ reviewed papers}]} {\sum_t\mu_t n_t}\le b. \]

A solution is the pair of thresholds attaining the optimum, together with the truthful ranking policy for every type. Proposition 4.9 certifies truthfulness because threshold acceptance and continuation policies are monotone.

I would expect this particular problem to be in Class A. There are only \(O(|R|^2)\) threshold pairs, and for each pair and each type, expected acceptance quality and review burden can be computed by a dynamic program over the sequential review process. The population enters through weighted sums over \(\mu\), so its size does not cause a per-author enumeration. With rational finite-support inputs, this is polynomial in the explicit instance size.

This is not an artificial relaxation. Section 5 already studies the threshold sequential mechanism, compares conference utility against review burden, and in §5.3 optimizes mechanisms for an empirical population distribution. The mirror simply replaces that empirical distribution by an explicit high-multiplicity distribution \(\mu\), and replaces Monte Carlo evaluation by exact dynamic programming. The authors should recognize it as their own practical mechanism-design problem at population scale.

The regime would be a large standardized reviewing pipeline with many authors or content producers, bounded portfolio sizes, a finite quality rubric, and repeated portfolio profiles. For example, a conference or content platform may have millions of submissions generated by recurring producer classes: one- through four-item portfolios, common quality bands, common review-noise laws, and common “number accepted” utilities. Then \(N\gg |T|\), and a discrete population with \(N_t=N\mu_t\) authors is exactly the rational high-multiplicity realization of the continuous instance. The paper itself names content recommendation and other producer-review settings as broader applications.

The main weakness is that ordinary academic authors may not form a genuinely small set of exact types. The paper’s own data have an average of only about two submissions per author, and realized quality vectors can be highly idiosyncratic. The mirror is strongest in a standardized, high-volume regime rather than for a conventional conference electorate. Also, because truthfulness is an individual-contract property, the model must impose pointwise incentive compatibility for each type; defining truthfulness only through aggregate population utility would make unilateral deviations disappear in an atomless society.

So my positive conclusion is limited but real: Proposition 4.9 supports a faithful and computationally tractable population mirror, covering the paper’s threshold mechanism, utility objective, review burden, and truthful ranking action. But because the paper has no named computational-complexity anchor, it cannot receive a strict positive ChoCo verdict without relaxing that eligibility rule.