| paper | d4-conf-aaai-2026-4467 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The paper reports empirical findings about LLM allocations, but states no numbered computational result of its own that meets the anchor requirement. The MNW guarantee belongs to prior work, and extending the experiments to valuation-type distributions would be a new study rather than a mirror of an anchored result.
fails bit a — no named computational result to mirror
No result is mirrored; the paper's empirical findings are unanchored, and the cited MNW guarantee is prior work.
I find no qualifying anchor in the paper. The provided text contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. Its central finding—that LLMs tend to preserve social-welfare approximation at the expense of EF1 as the criteria conflict—is an empirical observation, not a named complexity result.
So there are no anchors to restate and no continuous problem to specify per anchor. A study of many agents grouped by valuation type could be a natural extension of the experiments, but it would not supply the required anchor-based case for a continuous mirror of this paper’s results.
The paper has no named computational result to mirror. Its contribution is empirical: across sampled finite instances, LLMs tend to retain high social-welfare approximation as EF1 approximation falls, and prompt changes have limited effects. The MNW guarantee cited as a baseline is prior work, not a result of this paper. The proponent is right that there is no anchor to restate as a continuous computational question.
The strongest candidate would be to represent agents by a distribution \(\mu\) over complete valuation types. But with finitely many indivisible goods, any allocation gives nonempty bundles to at most finitely many agents. A mass description either discards those zero-mass recipients—and with them the allocation the paper evaluates—or keeps pointwise constraints on them, which cannot be captured by the type masses alone. Scaling the goods with the population and tracking a distribution of bundles could make a sensible continuous fair-division problem. It would also change what the LLM is asked to output: the paper tests a model’s choice of one allocation for a finite instance, not its construction of a mass allocation.
Another option is to sample larger finite societies from \(\mu\) and average the LLM’s fairness and efficiency scores. That could extend the empirical study, but it would test how results scale with sample size; it would not create a complexity question about computation over a continuous society. Prompting the LLM with weighted valuation types instead would be a new experiment, since the paper’s results do not establish that its behavior is invariant to that compressed input.
So the negative case is strong within ChoCo’s stated scope: the best proposed continuizations lead to a new fair-division model or a larger empirical study, not a continuous mirror of a computational result in this paper. The universal claim that no related continuous study could ever be worthwhile is too broad to prove; a type-based large-market study might be worthwhile on its own. But that possibility does not supply the missing anchor.
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.