Mitigating Skewed Bidding for Conference Paper Assignment

The paper numbers only empirical hypotheses and reports no named computational result of its own. The proposed Mass-Balanced Presentation Design is a plausible high-multiplicity extension, but it assumes a response law and changes per-paper orphan outcomes into aggregate shortage; it cannot supply the missing computational anchor.

The objection that survived

The proposed model replaces the paper’s measured per-paper orphan phenomenon with an assumed response function and aggregate bid mass, while leaving stochastic dependence, budgets, order, and endogenous demand unspecified.

What the mirror covers

No named computational result is covered. The exploratory model targets order, demand signals, reviewer types, and bid balancing, but leaves the experiments, statistical hypotheses, and cited algorithms outside any valid mirror.

Open questions for a prover

The case FOR (proponent)

The strongest honest case has an important limitation: this paper contains no qualifying computational anchor. Its only numbered named statements are Hypothesis 1 (the order effect) and Hypothesis 2 (the demand effect). It contains no Theorem, Lemma, Proposition, or Corollary asserting membership in \(P\), NP-hardness, parameterized complexity, or an algorithmic result. The computational and mechanism-design results it mentions from Fiez et al. [17] and Meir et al. [33] are cited prior work, not results proved in this paper. Therefore, strictly under the programme’s rules, there are zero eligible anchors and hence no legitimate “one continuous problem per anchor” to report.

There is nevertheless a plausible subject-matter mirror, though it cannot be presented as a mirror of a named computational result from this paper. The best candidate is a high-multiplicity conference-reviewing regime: a very large recurring conference has thousands of reviewers and papers, but reviewers fall into finitely many complete types. A type records a reviewer’s relevance scores for paper types, order sensitivity, demand sensitivity, bidding budget, and any other parameter affecting bidding. Paper types record topical features and demand targets. The mass \(\mu_t\) is the fraction of reviewers of type \(t\), while \(\nu_p\) is the fraction of papers of type \(p\). This is plausible for large conferences with repeated topical cohorts and a finite taxonomy of reviewer profiles, although it is not a good description of the paper’s small AAMAS field sample or its Mechanical Turk subjects.

A precise exploratory problem would be:

Mass-Balanced Presentation Design. An instance consists of finite reviewer types \(T\), paper types \(P\), rational masses \(\mu_t\) and \(\nu_p\), a required reviewer-to-paper ratio \(r\), a finite set of allowable demand signals, and a specified response function
\[ q_t(p\mid \pi,z_p)\in[0,1], \]
giving the probability that reviewer type \(t\) bids positively on paper type \(p\) when the platform uses presentation order \(\pi\) and signal \(z_p\). The response function may include relevance, primacy, and iPrice sensitivity exactly as studied in the paper. The platform chooses an order \(\pi\) and signals \(z_p\). Its expected bid mass on paper type \(p\) is
\[ b_p=\sum_{t\in T}\mu_t q_t(p\mid \pi,z_p). \]
The objective is to minimize aggregate under-demand,
\[ \sum_{p\in P}\max\{0,r\nu_p-b_p\}, \]
or, in the decision version, determine whether this value is at most a rational threshold \(K\).

This is recognisably the paper’s problem: it optimizes the order and demand information that the paper experimentally manipulates, with “under-demanded” and “orphan” papers as the objective. It is not a mirror obtained by deleting the behavioural content until an LP appears. The authors would likely recognise it as a computational extension of their platform-design question.

For a response law that is additive in relevance, rank, and demand signal, I would expect a Class A result: aggregation over reviewer mass makes expected bid totals linear, and order choices may admit an assignment or column-generation formulation. The interesting technical issue is precisely the pricing problem over possible orders. If demand signals are endogenous—updated from the resulting bid distribution—or if response to an order is nonseparable, continuum-specific hardness becomes plausible. Further questions include whether one should minimize shortage or the probability of an orphan at finite population size, whether a small but positive mass of demand-sensitive reviewers is enough to improve coverage, and how to impose reviewer welfare or fairness constraints.

The weakest point is also decisive: the paper’s main contribution is empirical, not computational. Its findings concern heterogeneous, possibly unstable human responses, whereas the proposed mirror requires a known finite response function \(q_t\). Moreover, the paper reports that a relatively small subgroup drives much of the improvement. If that subgroup has vanishing rather than positive population mass, continuization erases precisely the phenomenon of interest. Thus the mirror is substantively plausible, but this paper supplies no named computational theorem on which a programme-compliant positive case can rest.

The case AGAINST (opponent, writing after the proponent)

The negative case is unusually strong at the source level: this paper has no qualifying computational anchor. Hypotheses 1 and 2 are empirical claims about order and demand sensitivity, not theorems about an algorithm, optimization problem, complexity class, approximation ratio, or parameterized procedure. The field experiment, Mechanical Turk experiments, statistical tests, and practical recommendations are likewise not computational results. Results from Fiez et al. and Meir et al. are cited prior work, not results established by this paper. Under ChoCo’s stated screening rule, there are therefore zero eligible anchors. That alone prevents a programme-compliant positive verdict.

The proposed Mass-Balanced Presentation Design does not repair this defect; it creates a new problem inspired by the paper. Its central input, the response function \(q_t(p\mid \pi,z_p)\), already contains the empirical phenomenon the paper studies. If \(q_t\) is known, the behavioral contribution has been assumed away and the optimization concerns a stipulated response table. If it is unknown and must be estimated, the problem becomes one of statistical learning, model uncertainty, or mechanism design under behavioral uncertainty—none of which is a computational result of this paper.

The proposed objective also does not preserve the paper’s measured outcome. The paper counts individual papers receiving fewer than \(r\) bids or no bids. The quantity
\[ \sum_{p\in P}\max\{0,r\nu_p-b_p\} \]
only measures an aggregate shortage for paper types. It cannot distinguish evenly spread demand from all bids concentrating on a few papers of the same type. Expected bid mass is not expected orphan count: the latter depends on the per-paper distribution and correlation of bids, reviewer budgets, and the finite number of reviewers. A faithful limit would need a stochastic per-paper model—perhaps binomial, Poisson, or a more complicated correlated process—not merely the mean \(b_p\). If the model instead treats paper types as divisible mass, it has changed the orphan-paper predicate into a flow deficit.

The order effect creates a second structural problem. Primacy is a function of a paper’s position in a list, while the proposed paper type \(p\) is otherwise supposed to describe interchangeable papers. To preserve the effect, the model must attach a rank or location to each paper mass. Then the platform is choosing a measure-level ordering, not a finite permutation of paper types. Alternatively, one can split paper types by every relevant position, but that makes the positions part of the type space and removes the multiplicity compression the mirror is supposed to provide. The same tension appears for reviewer relevance: finite reviewer types require identical relevance profiles over the paper population, whereas the experiments concern highly paper-specific interests.

The demand signal is not exogenous either. In the field experiment, iPrices are derived from current demand and can change when reviewers return; reviewers also face coupled bidding budgets. A correct model must therefore specify timing, feedback, strategic response, and the joint distribution of bids, rather than independently assigning probabilities \(q_t\). Adding those ingredients could produce an interesting mean-field mechanism-design or equilibrium problem, but it would be a substantial re-modelling, not a continuous mirror of a named result here.

A better high-multiplicity scenario is imaginable: recurring conferences with topical reviewer cohorts, repeated paper categories, jointly scaled reviewer and paper populations, and a calibrated behavioral law. That could be a worthwhile new ChoCo problem. But it would be worthwhile because it introduces a new computational model, not because this paper supplies a computational theorem to continuize. The paper’s actual contribution remains an empirical claim about heterogeneous and partly unstable human behavior. Once that behavior is fixed as a finite response law, the empirical object has disappeared; once it is retained, the proposed finite-type optimization is under-specified.

Thus the honest verdict is red for this paper as a ChoCo mirror source. The universal claim that no related continuous population model could ever be worthwhile would be too strong—the proposed extension is not absurd. But no such model is a valid mirror of this paper’s own computational result, because the paper contains no qualifying computational result to mirror, and the proponent’s construction changes the empirical per-paper phenomenon into an assumed aggregate policy model.

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.