{"id":"83ed63e5-8f7b-48d9-be88-c245c2170061","arxiv_id":"2507.20728","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"The paper defines a society's value system as a shared grounding plus clustered group preferences, and introduces a deep-clustering method to learn it from pairwise choice data.","lead":"This paper formalizes a society's value system as a shared value-grounded model plus a set of distinct group value systems, and learns it from pairwise preferences. The approach is tested on train route choices, where the learned groups correlate with trip purposes such as business and shopping.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Value-system weights are not identifiable from ordinal pairwise preferences unless the grounding output scales are fixed; the reported cluster weights in Table 2 are therefore not, by themselves, evidence of time/cost/comfort priorities.","rationale":"The reader's verdict is CONDITIONAL, and I agree with it, but the conditions should be widened. The central contribution is not merely clustering preferences but obtaining interpretable group value systems, and those value systems are displayed as weight vectors. If those weights cannot be separated from the arbitrary scale of the learned groundings, the interpretation is not supported by coherence, representativeness, or the reported cluster weights. The reader's proposed condition—elicit or independently verify the grounding—would not by itself resolve this, because the identifiability problem remains even with a valid grounding dataset. A simple post-hoc contribution analysis using the existing model would settle the issue: if effective contributions match the raw weights, the concern is rebutted; otherwise Table 2's qualitative conclusions are not warranted. Other limitations, such as L always equaling Lmax and the absence of held-out validation, are real but less central to the paper's interpretability claim. I therefore keep the conditional verdict, with the additional requirement that the authors demonstrate, via normalization or contribution analysis, that the learned value-system weights are meaningful relative to the learned grounding scales.","tokens_in":19253,"tokens_out":11821,"duration_ms":155071,"concrete_test":"Using the best Lmax=3 trained model from the released code, freeze G_θ and W_l and compute, for each cluster l and value v, the effective contribution E_{l,v} = W_{l,v} · mean_{(e,e′)∈D^l_VS} |G_v(e) − G_v(e′)|, then normalize E_l over v. Compare the resulting effective weights with the raw W_l in Table 2 across all ten seeds. If the effective ordering differs (e.g., Cluster 2 is not primarily time-driven, or Cluster 1 is not primarily comfort-driven), the reported value-system interpretations are artifacts of the arbitrary per-value output scales. As a direct robustness check, re-run training with each G_v output standardized to unit variance over the route set before the final scalarization and check whether the final W_l and cluster interpretations remain stable.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing issue is not only that the grounding labels in §6 are hand-constructed; it is that, even if the grounding were valid, the learned per-value alignment functions A_vi are constrained only by ordinal preference labels (Defs. 1–2 and DV in §4). The paper reports only normalized cluster weights W_l in Table 2 and interprets them as value priorities (e.g., Cluster 2 as 70% time + 26% comfort), but the actual contribution of value v inside the value-system function W_l·G_V is W_{l,v}·(A_v(e)−A_v(e′)), not W_{l,v} alone. Since each A_v is a separate neural network with its own arbitrary output scale (§5–6) and is never anchored to a meaningful unit (unlike fixing A_time = −travel_time or A_cost = −cost), the same pairwise preferences over routes can be fit by many (G,W) pairs. The reported coherence of 1.0 reflects ordinal agreement, not a recovered scale. Consequently the cluster labels in Table 2, and the shopping/business interpretation, are not established by the reported weights. This concern is independent of the reader's valid point that the grounding dataset in §6 is constructed by the authors rather than elicited from agents.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formalizes the problem of learning the value system of a society as a shared grounding function together with a clustered set of agent value systems. It defines coherence, representativeness, and conciseness, casts the learning problem as a bi-level optimization, and proposes a deep-clustering heuristic that combines an EM-style algorithm with evolutionary exploration. The method is evaluated on a Swiss train-route choice dataset with 388 agents, yielding three clusters interpreted as comfort-oriented, time/comfort-oriented, and cost-oriented value systems, whose memberships correlate with trip purposes such as business and shopping.","tokens_in":1175,"tokens_out":1145,"duration_ms":59334,"significance":"The formal framework is a useful step toward explicit, interpretable models of societal value diversity, and the paper ships source code plus supplementary baseline experiments, including an ablation of the Lagrange-multiplier mechanism. If the identifiability and validation gaps below are closed, the contribution would be of real interest to value-aware AI and computational social choice. As it stands, the central empirical claim that the method recovers a socially shared grounding and a concise set of true group value systems is not yet supported.","major_comments":[{"comment":"The reported cluster weights W_l are not identifiable from the ordinal preference data. The Bradley-Terry probability in Eq. (1) depends only on differences A_v(e)-A_v(e'), and each alignment network A_v has an arbitrary output scale that is never anchored to a meaningful unit (unlike fixing A_time = -travel time or A_cost = -cost). Consequently, the same pairwise choices can be fitted by many (G_V, W) pairs, and the interpretation of Cluster 2 as '70% time and 26% comfort' is not established by the experiments. Please anchor the alignment scales, report scale-invariant quantities, or explicitly characterize the equivalence class of weights that yield the same ordinal predictions.","section":"Section 4, Eq. (1), Definition 4, and Table 2"},{"comment":"The grounding coherence of 1.0 is a fit success, not evidence of a socially agreed grounding. The dataset DV is constructed by the authors from the same route features used as network inputs: the paper states 'we presume that comfort depends on headway and interchanges' and builds DV by comparing instances using those definitions. The grounding network is trained on these self-generated labels and then evaluated on them. The Section 7 claim that clusters 'reflected their travel intentionality' also rests on correlations between the hand-built grounding and context features, so it requires external validation (e.g., elicited value judgments from agents or a held-out set of route features) before it can support the paper's central claim.","section":"Section 6, grounding dataset construction"},{"comment":"The number of clusters is never learned: the paper reports that 'the number of clusters (L) found always matched Lmax.' Since L is supplied as a hyperparameter, the conciseness objective does not select a concise society model in practice, and the decision to call L=3 'the best configuration' is justified only by a subjective reading of the representativeness-conciseness trade-off (especially since Lmax=2 gives the best Dunn Index). Please provide an explicit model-selection rule based on the Dunn Index or another criterion and show that it would choose L=3.","section":"Section 6, Figure 1 and Table 2"}],"minor_comments":[{"comment":"There is a typo: 'Avi(e) = Avi(e)' should read 'Avi(e) = Avi(e prime)'.","section":"Equation (1)"},{"comment":"The return statement returns beta_t, but the best solution in memory is obtained by GetBestSolution; the returned assignment should be the one associated with the selected solution.","section":"Algorithm 2, line 13"},{"comment":"The notation for standard deviations is inconsistent (e.g., 'pm(0.03, 0.08 0.11)' is missing a comma), and the Income column reports means without stating currency units or standard deviations.","section":"Table 2"},{"comment":"The phrase 'a family of |J| >= L >= 1 value systems' is confusing because the family has L members, not |J|; it should be rephrased as 'a family of L >= 1 value systems'.","section":"Definition 6"}],"recommendation":"major_revision","confidential_remarks":"The identifiability problem is the most load-bearing issue and is independent of the hand-constructed grounding concern; it should be resolved before acceptance. The paper is otherwise well-written, and the availability of code and supplementary baselines is a strength. There is also a scope question for the journal: the empirical demonstration is a single dataset, so the generalizability of the method remains untested."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, here's my read of arXiv:2507.20728.\n\nThe thing to know is that the formalization is the contribution, not the experimental result. Defining a society's value system as a shared grounding plus a clustered set of value systems, with coherence, representativeness, and conciseness properties, is a clean conceptual step beyond the single-aggregation view. The bi-level optimization (grounding first, then clusters) is well motivated, and the algorithm—MaxMin-RLHF-style EM with an evolutionary outer loop—is a reasonable adaptation. The paper is honest about several limitations; the supplementary material even proves why the bi-level approach beats naive sequential fitting, and it shares code.\n\nNow the soft spots. The most load-bearing one: the learned value system weights W_l in Table 2 are not identifiable. The value system function is W_l·G_V, and each A_v is a freely learned neural network with arbitrary output scale. The Bradley-Terry preferences only depend on differences A_v(e) - A_v(e′). So you can rescale each A_v and absorb the factor into W_l without changing any prediction. The paper never anchors these functions to meaningful units, which means the reported weights like (0.70 time, 0.26 comfort) for cluster 2 are not evidence that those agents prioritize time. The shopping/business interpretation built on those weights doesn't follow from the data. This is not a minor caveat; it undercuts the qualitative validation.\n\nThe other issue is the grounding. The authors build DV from their own route-feature definitions of time, cost, and comfort (comfort from headway/interchanges), then show the network reproduces those labels with coherence 1.0. That's a fitting success, not evidence that the method learned what the society actually values. The abstract's phrase 'learns socially shared value groundings' overstates this.\n\nA third, acknowledged weakness: L always equals Lmax, so the conciseness objective never actually selects a smaller number of clusters. This is stated as a limitation, but it means the 'conciseness' part of the learning problem is not demonstrated.\n\nThese are not fatal to the framework. The formal definitions are solid and the algorithm works as a clustering procedure. But the paper's central claim—that the method discovers a society's value systems from observed preferences—is not supported by the current evaluation. What's needed: either fix the identifiability problem (e.g., anchor A_v to fixed features with known scales, or report scale-invariant distances), elicit or independently verify the grounding, and validate the cluster-value associations on held-out data or with synthetic datasets where ground truth is known.\n\nWho should read this: anyone working on value alignment, multi-objective preference learning, or computational social choice. It deserves a serious referee—the formalization is worth engaging with—but the evaluation needs major revision before the empirical claims can be trusted. Send it to review, but with the expectation that the authors will have to rework the experiments.\n\nRecommendation: accept for peer review, and push for substantive revision on the identifiability and grounding-validation fronts.\n\nLet me know if you want to chat about it.","headline":"A genuinely new formalization of societal value systems, but the empirical evaluation overreaches: the learned cluster weights are not identifiable from ordinal preferences, and the hand-built grounding makes coherence 1.0 uninformative.","tokens_in":20094,"tokens_out":3933,"would_cite":true,"duration_ms":46660,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a society's value system is best represented as a shared grounding plus a small set of group value systems, and presents a deep-clustering method that learns both from qualitative pairwise preferences, validated on…","keywords":["value alignment","value learning","value grounding","social value systems","preference-based clustering","Bradley-Terry model","deep clustering","travel choice modelling"],"falsifier":"Re-run the pipeline with grounding labels elicited directly from travellers (asking them to compare routes for each value separately) instead of the authors' comfort rule; if coherence drops substantially or if the clusters no longer track trip purpose, the learned social value system is an artifact of the hand-crafted grounding. A second check: hold out a subset of travellers, learn the value systems on the rest, and test whether the assignment of held-out agents to value systems predicts their stated trip intentions above chance.","tokens_in":18947,"feed_emoji":"🚆","tokens_out":7350,"duration_ms":83656,"temperature":0.7,"pith_summary":"This paper tries to establish that the value system of a society is better modelled as the combination of a socially shared value grounding—a computational semantics for value labels such as 'time efficiency' or 'comfort'—together with a small set of group-level value systems, rather than as one aggregated preference model. It formalizes this claim as a bi-level optimization problem with three desiderata: grounding coherence, representativeness, and conciseness, and solves it approximately with a heuristic deep-clustering algorithm built on an EM-style clustering loop with evolutionary exploration. Evaluated on 3,492 real pairwise train-route choices from 388 Swiss travellers, the method recovers three value systems—a large comfort-oriented group, a time-efficiency group, and a cost-efficiency group—whose memberships correlate with stated trip purposes such as business and shopping. If correct, this gives value-aware AI a way to obtain interpretable, group-diverse representations of stakeholder values from data that is easy to collect.","feed_headline":"Train choices reveal three value systems: comfort, time, cost","feed_subtitle":"A deep-clustering method learns shared value meanings and group value systems from 3,492 pairwise route choices.","key_machinery":"The load-bearing object is the grounding function $G_V$, a set of per-value alignment functions learned by neural networks from pairwise preference datasets, with each agent's value system represented as a positive linear weighting $W_l \\cdot G_V^\\top(e)$ over those alignments. Preferences are modelled with the Bradley-Terry likelihood, and the learning problem is cast as a bi-level optimization: first maximize grounding coherence (via a Lagrange-multiplier constrained loss), then maximize a Dunn-index-like trade-off between conciseness (minimal pairwise discordance between value systems) and representativeness (how well each agent's stated preferences are matched by its assigned value system). The algorithm alternates an EM-style assignment of agents to value systems with gradient updates, wrapped in an evolutionary memory that mutates and refines candidate solutions to escape local optima.","core_discovery":"On its own terms, the paper claims that observing only qualitative pairwise comparisons between alternatives—'I prefer this route over that one'—is enough to learn (a) a socially agreed grounding function that assigns each value label a continuous alignment score over entities, and (b) a set of L value systems, each a linear weighting of those alignments, together with an assignment of every agent to one value system. These two pieces, the grounding plus clustered value systems, constitute its definition of the value system of a society. The empirical discovery in the Swiss route-choice use case is that the learned clusters are not arbitrary preference groups: the largest cluster weights comfort, a second weights time efficiency, and the third weights cost, and the memberships align with stated intentions (business travellers in the time cluster, shopping travellers in the cost cluster) even though those intentions were never used in training.","pith_inferences":["The hand-crafted comfort grounding is the main open risk; an immediate test is to elicit per-value comparisons from travellers and check whether the grounding network still reaches coherence 1.","Because each cluster is a linear weighting over shared values, a deployed system could personalize its advice by cluster without retraining, e.g., showing cost-focused travellers cheaper options and time-focused travellers faster ones; the paper does not spell this out.","In multicultural or institutionally mixed settings the single socially-agreed grounding may not exist; the authors flag this, and a natural extension is to learn several groundings and test which one each cluster uses.","The conciseness measure rewards separated clusters, but the EM assignment greedily favours representativeness and the number of clusters always hit Lmax; an alternative scoring metric such as Davies-Bouldin or Ray-Turi is suggested by the authors as future work and could change the reported L = 3 choice."],"forward_implications":["A value-aware system can be built from pairwise preferences alone, with no numerical ratings per alternative per value, because the grounding networks supply the alignment scales and the clusters supply the group weights.","In the route-choice case, the recovered clusters carry external meaning: the time-oriented cluster is 142.7% above average in business trips and the cost-oriented cluster is 92.3% above average in shopping trips, so learned value systems can label groups by real-world intentions.","With L = 3 clusters the method reaches 84.5% representativeness versus 80.7% for a single society-wide value system, showing that the pluralistic representation buys accuracy without losing a concise explanation.","The same formalism transfers to any domain where options can be compared pairwise (policy options, products, services), since the grounding function only needs features of the compared entities."],"supporting_citations":[{"why":"Supplies the EM-style preference clustering algorithm that the method adapts for joint agent assignment and value-system training.","marker":"[3]"},{"why":"Supplies the Bradley-Terry pairwise preference model and cross-entropy loss used for both grounding and value-system fitting.","marker":"[4]"},{"why":"Supplies the Lagrange-multiplier ascent method that enforces grounding coherence before value-system optimization.","marker":"[5]"},{"why":"Supplies the Dunn Index that defines the trade-off between conciseness and representativeness in the optimization objective.","marker":"[8]"},{"why":"Supplies the Gaussian perturbation mutation strategy used to explore new value systems in the evolutionary outer loop.","marker":"[9]"},{"why":"Provides the prior grounding-and-value-system representation and route-choice modelling framework that this paper extends to societies.","marker":"[12]"},{"why":"Supplies the rank-proportional selection rule used to choose candidate solutions from the memory.","marker":"[13]"},{"why":"Provides the pluralism argument that societies hold multiple coexisting value systems rather than one aggregate.","marker":"[17]"},{"why":"Provides the real-world dataset of 388 agents and 3,492 pairwise train-route choices used in the evaluation.","marker":"[43]"}],"fun_headline_variants":["Pairwise choices train AI to find society's value systems","Deep clustering reveals three value systems from travel preferences","Preferences alone teach machines the value systems of societies","Route choices expose comfort, time, and cost value systems"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the authors' hand-made definitions of the three values—time efficiency from travel time, cost efficiency from cost, and comfort from headway and interchanges—are the meanings travellers actually had in mind; if those semantics are wrong, the perfect grounding coherence is a fit to self-generated labels rather than evidence about real societal values.","fun_headline_variants_meta":{"raw":{"variants":["Pairwise choices train AI to find society's value systems","Deep clustering reveals three value systems from travel preferences","Preferences alone teach machines the value systems of societies","Route choices expose comfort, time, and cost value systems"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000162,"raw_usage":{"total_tokens":1212,"prompt_tokens":887,"completion_tokens":325,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":503,"completion_tokens_details":{"reasoning_tokens":262}},"tokens_in":503,"tokens_out":325,"duration_ms":4077,"temperature":1.0,"reasoning_tokens":262,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T13:19:00.502942+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the pipeline with grounding labels elicited directly from travellers (asking them to compare routes for each value separately) instead of the authors' comfort rule; if coherence drops substantially or if the clusters no longer track trip purpose, the learned social value system is an artifact of the hand-crafted grounding. A second check: hold out a subset of travellers, learn the value systems on the rest, and test whether the assignment of held-out agents to value systems predicts their stated trip intentions above chance.","supporting_citations":[{"cited_title":"Chakraborty, J","cited_arxiv_id":null,"evidence_quote":"Supplies the EM-style preference clustering algorithm that the method adapts for joint agent assignment and value-system training."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Bradley-Terry pairwise preference model and cross-entropy loss used for both grounding and value-system fitting."},{"cited_title":"Cotter, H","cited_arxiv_id":null,"evidence_quote":"Supplies the Lagrange-multiplier ascent method that enforces grounding coherence before value-system optimization."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Dunn Index that defines the trade-off between conciseness and representativeness in the optimization objective."},{"cited_title":"Holgado-Sánchez, J","cited_arxiv_id":null,"evidence_quote":"Provides the prior grounding-and-value-system representation and route-choice modelling framework that this paper extends to societies."},{"cited_title":"Jannoud, Y","cited_arxiv_id":null,"evidence_quote":"Supplies the rank-proportional selection rule used to choose candidate solutions from the memory."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the pluralism argument that societies hold multiple coexisting value systems rather than one aggregate."},{"cited_title":"Vrtic and K","cited_arxiv_id":null,"evidence_quote":"Provides the real-world dataset of 388 agents and 3,492 pairwise train-route choices used in the evaluation."}],"review_version":1}