| paper | d4-conf-aaai-2025-1653 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | unclear |
The numbered theorems establish axiomatic failures or guarantees, while the polynomial-time LP implementation of LCPO is stated outside a numbered computational result. The paper therefore has no named computational result to anchor a ChoCo mirror, so bit (a) fails.
fails bit a — no named computational result to mirror
No result is covered: the numbered results are axiomatic, and the polynomial-time implementation of LCPO is an unnumbered observation.
The paper has no named computational result that qualifies as an anchor under your criteria. Its numbered theorems concern axiomatic properties: for example, Theorem 3.1 proves failures of PMC and PO for certain loss-based rules, and Theorem 4.3 proves axiomatic guarantees for LCPO. Neither states a complexity result.
The text does mention a polynomial-time implementation of LCPO and an exponential brute-force algorithm for linear Kemeny; it also says ordinary Kemeny is generally NP-hard. But these are not numbered results about the paper’s computational problems, so I can’t honestly use them as anchors.
A continuous population of RLHF preference types is easy to imagine, but without an eligible named computational result there is no anchor-based mirror question to put to a prover.
The proponent identifies the decisive problem: there is no named computational result here to mirror. The numbered results establish axiomatic claims—such as the failures of loss-based rules in Theorem 3.1 and LCPO’s guarantees in Theorem 4.3—not complexity bounds. The paper’s polynomial-time LP implementation for LCPO is an algorithmic observation attached to an axiomatic result. Its NP-hardness statement concerns ordinary Kemeny; for linear Kemeny, the paper gives brute-force enumeration but no complexity classification.
A continuous population of RLHF annotators is quite plausible: types could be complete preference rankings, with a distribution over them. But using that distribution to restate or test PO, PMC, or the other axioms would continuize an axiomatic question, outside ChoCo’s stated scope. One could instead study the complexity of computing LCPO or linear Kemeny from a supported distribution. That might be worthwhile research, but it would introduce a new computational question rather than mirror a computational result this paper proved. Its potential value cannot rescue the missing anchor.
So the strongest honest negative case is narrow but firm: this paper offers no worthwhile *mirror of its results* under the programme’s anchor requirement. It does not establish that every future computational study of continuous RLHF preference populations would be uninteresting; that broader claim would go beyond the evidence.
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.