| paper | d4-conf-atal-2015-0155 |
| authors | — |
| venue | ? |
| filed under | unclassified |
| judged by | gpt-6-luna / xhigh (triple__luna__xhigh__c6r1) |
| judge confidence | medium |
| authors would recognise it | yes |
other→P discrete result → its continuous mirror
Answer: Polynomial-time solvable; the exact computation belongs to \(\mathrm{FP}\). · discrete side: other · adjudication: proved
full proof & adversarial review →
Proposition 1
statement extracted from the paper’s text layer
Given shared options \(O\), criteria \(H\), and model family \(\Pi\), let \(T\) be a finite support of complete decision-maker types with binary rational masses \(\mu_t\) summing to one. Each type specifies a consistent preference base and an efficiently evaluable response policy; the instance also specifies a deterministic decision-aider move-selection policy so the dialogue is determined. Execute the paper's protocol under UN independently for each type. Compute each type's terminal state \(z_t\) (success or failure, and the recommendation if successful), the exact mass of each terminal state, and the average number of moves \(\bar{\ell}=\sum_{t\in T}\mu_t\ell_t\).
A high-multiplicity recommender population in which each type \(t\) carries mass \(\mu_t\) and has a complete preference base and response policy. The dialogue runs separately for each type under a specified decision-aider policy; the task is to compute terminal outcomes and aggregate communication workload, not to optimize moves.
The result bounds dialogue moves, not the computation time of each move, and the batch task also needs explicit decision-aider move selection to determine outcomes.
fatal: False
The mirror covers Proposition 1's interaction-length bound and Corollary 1's termination claim. It leaves Proposition 2's success and correctness guarantees, and the model-selection axioms, without population-level questions.
The strongest anchor is Proposition 1, proved in this paper. Under the Update Node rule (UN), it bounds a dialogue’s length by
\[
7|P|+2|O|(|O|-1)+1,
\]
where \(P\) is the decision maker’s preference-information base and \(O\) is the set of options. Corollary 1, also proved here, states that the protocol terminates under UN. This is an interaction-length result: the paper does not give a named P/NP classification or a running-time analysis for computing the dialogue. I would not present it as one.
A narrow continuous mirror is Mass Dialogue Workload. Its input is a shared option set \(O\), criteria and decision-model family as in the paper, and a distribution \(\mu\) over a finite support \(T\) of complete decision-maker types. Each type specifies a consistent preference base \(P_t\), the relevant preference model, and an efficiently evaluable policy for how that decision maker responds to every dialogue history. The masses are binary rationals summing to one. Run the paper’s protocol independently for each type. The requested solution is the terminal result for every support type—success or failure, recommendation if successful, and number of dialogue moves—plus the exact mass of each terminal outcome and the average moves per decision maker,
\[
\bar{\ell}=\sum_{t\in T}\mu_t\ell_t.
\]
The decision variable at each step is the dialogue move; the task is to compute the prescribed protocol’s outcomes and average communication burden, not to optimize its moves.
A plausible regime is a large recommender handling repeated, standardized choices—for example, employees selecting from the same benefits options under the same criteria. Many users may share the same complete preference information and feedback behavior, even though the number of users is enormous. Then \(\mu_t\) means the fraction of users following archetype \(t\), and \(\tau=|T|\) can be much smaller than the population. This respects the paper’s adaptive dialogue: each user still has an individual conversation, while the platform asks about aggregate workload and outcomes.
For each type, Proposition 1 supplies the bound
\[
\ell_t\le 7|P_t|+2|O|(|O|-1)+1.
\]
With the paper’s illustrative finite model family and efficient move computations, simulating one dialogue per support type and aggregating the rational masses is tractable in \(\tau\), the explicit type descriptions, and \(|O|\). The mirrored result is the bounded interaction workload across a high-multiplicity population. This is a modest Class A result, not a new optimization algorithm: Proposition 1 bounds dialogue moves, but says nothing by itself about the computation time of each move.
The mirror covers only the protocol’s termination and interaction-length guarantee. I would not extend it to the paper’s claims about whether the final recommendation is acceptable or whether the DA identifies the DM’s properties correctly; those are separate results, and population-wide versions would need their own question and anchor.
The weakest point is that the paper studies a dialogue with one DM, while this mirror batches many independent DMs. The type must include response behavior as well as preferences, so the support could be large in ordinary consumer recommendation. A repeated standardized workflow makes the population story plausible, and the paper itself motivates its protocol with recommender systems, but this remains an operational scale-up of the paper’s problem, not a central result about collective choice.
The strongest anchor is Proposition 1 and Corollary 1: under UN, every dialogue terminates, with at most \(7|P|+2|O|(|O|-1)+1\) moves. A population of users in a standardized recommender workflow is conceivable, and grouping users who share the same complete preference information and response behavior is legitimate high multiplicity. The problem is what the population adds to this result.
For a fixed dialogue policy, users of each type simply run separate copies of the protocol. The proposed average workload is then a weighted sum of individual dialogue lengths, and Proposition 1 gives the aggregate bound immediately:
\[
\bar{\ell}\le 7\sum_{t\in T}\mu_t|P_t|+2|O|(|O|-1)+1.
\]
Computing terminal-outcome masses likewise means simulating one dialogue per support type and adding its mass to the corresponding outcome. With explicit response policies and tie-breaking, the work is proportional to the number of types and the number of moves per type. The proportions are passive weights: there is no mass transfer, shared decision, or optimization over the population.
To make the batch question determinate, the input has to specify how each DM responds to dialogue histories, and how the DA chooses among the protocol’s permitted moves. With those policies specified, the aggregate calculation is bookkeeping around the individual executions. Without them, the protocol does not determine the requested outcomes. If one instead asks for worst-case behavior over legal responses, Proposition 1 already supplies the same per-dialogue bound; if one asks for expected behavior, one must add a stochastic response model. Any harder computation would then come from that added policy or probability model, rather than from continuizing the paper’s termination result.
A stronger population story might ask a service to allocate a limited dialogue budget across user groups, or to maximize the share who accept a recommendation. That could be an interesting computational problem, but the paper’s result says nothing about budget allocation or a shared objective. It would add the coupling that makes population-level computation substantive, and would therefore be a new resource-allocation problem built around the protocol, rather than a mirror of Proposition 1.
Proposition 2 does not rescue the mirror: its success and correctness guarantees also apply to each DM individually. For a population of types satisfying its assumptions, the guarantee lifts pointwise; calculating the mass of those types again requires no population-level algorithm. This is not an objection that the continuous answer is easy, or that high-multiplicity users are implausible. It is that the named result concerns the length of an isolated interaction, and its population version only averages isolated interactions. I would reject this as a worthwhile continuous-computational mirror under ChoCo’s aims. The negative case is not airtight: a platform operator might value aggregate workload as an operational metric, but that is a weaker case than a new computational question about a continuous society.
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.