| paper | d4-conf-ijcai-2021-0313 |
| 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 performance comparisons, but contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. The opponent’s collective-learning proposal is a new objective, not a continuous analogue anchored to a named result of this paper.
fails bit a — no named computational result to mirror
The paper’s empirical comparisons of plan generation and LTL formula learning have no eligible named computational anchor; the proposed collective-learning objective extends the paper’s setting.
I can’t make a paper-anchored case for a continuous mirror here. The paper contains no numbered theorem, lemma, corollary, or proposition asserting a computational result. Its contribution is an empirical study: it assembles existing planning and LTL-learning techniques and measures how well they perform. The reported findings—such as the gap between realistic and idealized plan generation—are not named results of the kind your anchor requirement calls for.
A population of users sharing temporal preferences might be a scenario to explore, but turning the study into an aggregate learning problem would need a new model and would not mirror one of this paper’s named results. There is therefore no honest anchor, and no continuous problem to state as its mirror under the requested standard.
The proponent names no continuous mirror, so there is no specific mirror to defeat. Their narrower point is right: this paper has no numbered complexity theorem. But that does not establish that its computational contribution has no population mirror. Section 6.3 reports a concrete algorithmic finding: learning succeeds much more often when example plans are balanced using the hidden target, while realistic plan generation performs markedly worse as the target formula grows. The paper identifies informative example generation as the bottleneck.
A natural population version keeps plans and LTL\(_f\) formulas discrete and makes the stakeholders continuous. Suppose a planning organization has many users, with common temporal preferences such as “eventually deliver the medicine” or “never leave this area unattended.” Let each preference formula be a type \(t\), and let \(\mu_t\) be the fraction of stakeholders with that type. For a plan \(\pi\), its approval mass is
\[
a_\mu(\pi)=\sum_t\mu_t\,\mathbf{1}[\pi\models\phi_t].
\]
The group’s target can be the majority approval pattern over feasible plans. The computational question is whether a compact formula \(\psi\) represents that pattern, or how to learn a formula with minimum weighted disagreement from annotated example plans. Its parameters include the planning task, support size \(\tau\), formula-size bound, and rational-mass bit length.
This is a credible high-multiplicity regime: large stakeholder groups often reuse a modest set of temporal requirements. A type records the complete preference relevant to the task; identities need not be retained when the aim is to produce a shared policy specification. The finite electorate with \(n_t\) users per type has \(\mu_t=n_t/n\), so the continuous model is its direct high-multiplicity form. The paper’s plan-generation and formula-learning machinery remains central. Its empirical bottleneck becomes a population question: can the generator find plans that expose the group’s majority boundary well enough to learn a compact collective preference?
That is a new group-learning objective, not a theorem already proved by this paper, and the paper’s single-user framing makes this case less direct than a mirror of a named complexity result. Still, the proponent has not shown that the authors would reject the population version: it retains their planning tasks, temporal preference language, example plans, and learning pipeline, while changing whose preference the system is meant to elicit. So the absence of a numbered theorem limits the strength of the mirror case; it does not support the universal claim that no worthwhile continuous mirror exists.
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.