| paper | d4-conf-wine-2025-0026 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | yes |
The paper contains no numbered theorem, lemma, corollary, or proposition establishing a computational result; its interaction savings are experimental, and its query bound is stated outside a named result. The continuous allocation LP is plausible and recognizable as a population version of CECA winner determination, but it is an unanchored subroutine rather than a qualifying result of this paper.
fails bit a — no named computational result to mirror
The candidate mirror covers CECA winner determination; it leaves proxy-driven elicitation, LLM query design, and the reported interaction savings unmirrored.
I find no qualifying anchor in this paper: it has no numbered theorem, lemma, corollary, or proposition establishing a complexity or query-complexity result. Section 6.1 reports an empirical finding—75% efficiency after about ten interactions for \(\omega_{\mathrm{xor}}\), four for \(\omega_{\mathrm{vd2}}\), and two for \(\omega_{\mathrm{nvd}}\)—but those are experimental measurements, not a named guarantee. Algorithms 1–3 describe procedures, and Equation (3) gives a winner-determination formulation; neither asserts a complexity classification. The termination guarantee is attributed to prior work [18]. So there is no result of the required kind to restate, and no anchor-backed continuous problem to supply.
The strongest *unanchored extension* I see is Continuous XOR Winner Determination. Its instance has \(n\) goods, types \(t\in T\) with rational masses \(\mu_t\) summing to one, rational per-capita supplies \(q_j\) for each good, and an explicit finite XOR bid list \((S_{t\ell},v_{t\ell})\) for each type, including the empty bundle. The question is to find a welfare-maximizing allocation of mass. Let \(x_{t\ell}\) be the mass of type \(t\) receiving the integral bundle \(S_{t\ell}\). A solution maximizes
\[
\sum_{t,\ell} v_{t\ell}x_{t\ell}
\]
subject to \(\sum_\ell x_{t\ell}=\mu_t\) for every type and \(\sum_{t,\ell:\,g_j\in S_{t\ell}}x_{t\ell}\le q_j\) for every good. This is an LP with size polynomial in the number of types, listed bids, goods, and input bit length, so I expect this formulation to be tractable.
A plausible regime is a large procurement or resale market with many buyers choosing among bundles of standardized goods. Buyers in the same type have identical preferences and other relevant parameters; there may be millions of buyers but relatively few types. Inventory must scale with the population, as the per-capita supplies \(q_j\) specify. Each buyer still receives an integral bundle; \(x_{t\ell}\) records how much of a type receives each bundle. This preserves the paper’s XOR valuations and welfare-maximizing auction objective, and its authors could recognize it as a high-multiplicity version of the winner-determination step used by CECA.
That is the scope of the positive case: the auction’s allocation problem, not the paper’s central claim about LLM proxies reducing elicitation interactions. The mirror raises further questions about learning the listed bids from type-level value or demand queries, defining a query model for natural-language questions, and rounding mass allocations to finite populations while bounding welfare loss.
My weakest point is the source and model gap: the LP assumes explicit XOR bids, while the paper’s main contribution concerns proxies learning hidden preferences, often through natural language. Sharing a type makes one type-level interaction stand for many people, which changes the elicitation setting. The paper’s experiments neither establish the LP result as a theorem nor show that the interaction savings survive that change.
The strongest negative case is that this paper is a poor target for ChoCo, not that no continuous mirror could be worthwhile. Its headline is an experimental comparison: LLM proxies reach a given welfare level with fewer person–proxy interactions. It gives no named theorem classifying a computational problem. Section 3 does state a query bound for the DNF learner, but presents it as an adaptation of earlier work [18], not as a result about the paper’s LLM proxies. So there is little formal complexity content here to mirror.
The proposed Continuous XOR Winner Determination problem is well-formed: given explicit XOR bids for each type, allocating mass to integral bundles under per-capita supply constraints is an LP. But that formulation assumes the preferences the paper works to elicit are already known and listed. It therefore mirrors the auction’s allocation subroutine, not the paper’s main contribution. Its tractability also follows immediately from the chosen representation: with a polynomial-size bid list, the LP is polynomial-size; with an exponentially large list, the input already hides the hard part. That may be a useful baseline, but it is a thin result to motivate a continuization project around this paper.
A stronger mirror would include elicitation: a distribution over bidder types, type-level value and demand queries, per-capita supplies, and a welfare target. Here the model faces a tradeoff. If people of a type give identical answers to every query—including natural-language questions—then one representative can stand for the whole mass and the problem reduces to learning each of the \(\tau\) types. The population size brings little new complexity to the paper’s query comparison. If people share valuations but answer natural-language questions differently, a type-level transcript cannot safely stand in for them; the type must include that response behavior, and its support may grow with the number of people. Replacing language questions with standardized value and demand queries avoids that problem, but removes the paper’s distinctive proxy design.
That is a strong argument about fit and priority: the clean continuous allocation problem strips away elicitation, while a faithful end-to-end version either compresses away the individual communication burden or needs many types to preserve it. It also explains why the paper’s reported interaction savings do not directly carry over to a mass society.
It does not, however, defeat the proposed LP as a valid population-continuous question. A market with many buyers of a few known types and inventory scaled with the population is plausible, and the resulting allocation problem matches the programme’s scope. I would argue that it is not yet a compelling mirror of this paper’s main result, but I cannot honestly support the universal claim that no scenario yields a worthwhile continuous mirror.
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.