| paper | d4-conf-aaai-2023-0300 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | high |
| authors would recognise it | no |
The paper states no numbered computational result to anchor a mirror, and the proponent explicitly concedes this. The proposed cohort preference estimation is a plausible follow-on, but it changes the paper’s individual reward-learning task into a population-level aggregation problem.
fails bit a — no named computational result to mirror
No named computational result is covered; the proposed cohort preference estimation is a new population-level task, while the paper studies learning one stakeholder’s reward function.
The strongest honest positive case is an adjacent extension, not a mirror anchored in a result this paper proves. The paper contains no numbered theorem, lemma, corollary, or proposition asserting a complexity or algorithmic classification. Equations (1) and (2) define the reward and choice models; the paper also describes an active query-selection method and reports simulations and a 33-person study. None is a named complexity result, so there are no qualifying anchors to quote or restate.
A plausible continuous problem would treat a large health-policy consultation as a population of stakeholders. A type \(\theta=(w,c)\) specifies a complete reward function, including its response model; \(\mu_\theta\) is the fraction of stakeholders of that type. For each query \(q\), consisting of two equal-length allocation trajectories, type \(\theta\) chooses the first with the paper’s Bradley–Terry probability \(p_{\theta q}\). Given \(N_q\) responses to query \(q\), of which \(k_q\) favor the first trajectory, estimate the type distribution by finding
\[ \hat{\mu}\in\Delta(\Theta) \quad\text{maximizing}\quad \sum_q \left[ k_q\log\!\left(\sum_{\theta}\mu_\theta p_{\theta q}\right) +(N_q-k_q)\log\!\left(1-\sum_{\theta}\mu_\theta p_{\theta q}\right) \right]. \]
Here \(\Theta\) is a finite set of possible reward types, and the solution is the estimated distribution \(\hat{\mu}\). The regime would be a large constituency with many stakeholders sharing a relatively small number of complete reward types. With rational likelihood inputs bounded away from zero and one, this is a concave optimization problem, so an approximate version looks tractable. The connection to the paper is its trajectory comparisons and reward-based preference model.
The weak point is substantial: the paper learns one stakeholder’s reward function and explicitly leaves aggregation of multi-dimensional preferences outside its scope. A large population with few shared reward types is plausible, but the paper’s small, diverse sample does not establish that regime. So this is a defensible direction for follow-on work, but it is not a continuous mirror of a named result in this paper; under the stated anchor requirement, there are zero anchors.
There are no anchors to carry over: the paper states no theorem, lemma, or corollary about the complexity of a computational problem. It gives a reward model, describes active query selection over a pre-generated set, and reports simulations and a small study. So there is no computational result here for a continuous mirror to restate.
The proposed population estimate is a new task. It changes the paper’s target from learning one stakeholder’s preference to estimating a distribution of preferences across stakeholders. Its likelihood also treats each answer as if it came from a fresh draw of a stakeholder type. In the paper, a stakeholder keeps the same preferences across an adaptive sequence of questions. A coherent cohort model would instead group respondents by their answer histories and assign each type a probability to each history. With a finite set of candidate types and a fixed query protocol, maximizing that history likelihood is concave over the type-mass simplex. The number of respondents then affects the counts and their bit length; the main dimensions are the candidate types and possible histories. A continuous type space raises identifiability questions: finite histories generally reveal only finitely many aspects of the distribution. Imposing assumptions to recover more would define a new model.
A better version could jointly choose queries to learn the cohort distribution, or use that distribution to rank allocation policies. Either could be worthwhile, but both add an aggregation objective and a population-level query design problem that this paper leaves outside its scope. The paper’s stated query task is to infer an individual’s reward, so these extensions do not supply a mirror of one of its computational results.
I cannot honestly make the universal negative airtight. The allocation side is a plausible counterexample: the paper already represents group states and actions as proportions, assumes deterministic transitions for a large population, and simulates a synthetic cohort. Replacing that cohort with masses over demographic profiles and asking how to optimize allocation trajectories could produce a worthwhile continuous-control problem. But that would be a new computational question built from the paper’s model, not a mirror of a result it proves. The negative case is strongest on the absence of anchors and the mismatch in the proposed preference-mixture task; it is weakest against that fluid allocation extension.
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.