Pith. sign in

REVIEW 3 major objections 4 minor 48 references

Learning the Value Systems of Societies from Preferences

T0 review · 3 major / 4 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read 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…

desk verdict 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. read the letter →

arxiv 2507.20728 v1 pith:VPVZYCJ7 submitted 2025-07-28 cs.AI cs.CYcs.LG

Andrés Holgado-Sánchez , Holger Billhardt , Sascha Ossowski , Sara Degli-Esposti This is my paper
classification cs.AIcs.CYcs.LG
keywords valuealignmentlearninggroundingsocialsystemspreference-basedclusteringBradley-Terrymodeldeeptravelchoicemodelling
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

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.

What carries the argument

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.

What would settle it

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.

Watch

Extended reading notes

Core claim

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.

Load-bearing premise

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.

Editorial extensions

If this is right

  • 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.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • 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.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 4 minor

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.

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 (3)
  1. [Section 4, Eq. (1), Definition 4, and Table 2] 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.
  2. [Section 6, grounding dataset construction] 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.
  3. [Section 6, Figure 1 and Table 2] 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.
minor comments (4)
  1. [Equation (1)] There is a typo: 'Avi(e) = Avi(e)' should read 'Avi(e) = Avi(e prime)'.
  2. [Algorithm 2, line 13] 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.
  3. [Table 2] 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.
  4. [Definition 6] 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'.

Circularity Check

1 steps flagged · score 4.0 of 10

Grounding coherence in the travel use case is a fit to author-defined labels built from the same route features; the value-system clustering and intention correlation remain independent.

  1. self definitional [Section 6, 'Evaluation' (grounding dataset construction; grounding network input)]
    "We assume that the route choices were guided by three values: time efficiency, cost efficiency, and comfort.While the groundings for time and cost efficiency are based on travel time and cost, respectively, we presume that comfort depends on headway and interchanges: if a route has both lower headway and fewer interchanges, we consider it more comfortable. In cases where only one of the features is better, we assume no preference and let the model estimate comfort alignment freely."

    The grounding dataset D_V is not elicited from agents; it is generated by the authors from the same four route features (time, cost, headway, interchanges) that the grounding network consumes. A time-efficiency label is just 'lower travel time' and a comfort label is 'lower headway and fewer interchanges'. The network is then trained to reproduce these labels and reported to reach coherence 1.0. The perfect coherence is therefore a supervised fit to a self-generated target: the same feature values define both input and label, so any sufficiently expressive monotone model should recover the mapping.

full rationale

The derivation chain is mostly open. The formalization (Definitions 1-8) and the EM/evolutionary clustering algorithm are self-contained: value-system weights and cluster assignments are trained on the real pairwise route choices D_VS, and the association between clusters and travel intentions (shopping, business) uses context features that were not training inputs, so that correlation is emergent rather than constructed. I find one significant self-definitional step in the evaluation: the grounding dataset D_V is built by the authors from the same route attributes that the grounding network observes; therefore the reported grounding coherence of 1.0 is agreement with a self-generated label, not evidence of a recovered socially-agreed grounding. The paper does not invoke a uniqueness theorem, and the citation to the authors' prior work [12] for the socially-agreed grounding assumption is an acknowledged modeling premise (with a limitation caveat), not an external result that forces the output. The non-identifiability of W_l without fixing the scale of A_v is a correctness/interpretability risk rather than a circular step. Overall, the central value-system clustering and its external intention correlation retain independent content, so the circularity is partial and localized to the grounding-coherence demonstration. Score 4.

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

The central claim rests on the socially-agreed grounding assumption and on the authors' hand-built definitions of what each value means in the route domain. These are not learned from independent evidence. The cluster weights and the conciseness-representativeness trade-off are fitted to the same data, and the conciseness goal is not achieved (L=Lmax).

free parameters (4)
  • Lmax (maximum number of clusters) = 2 to 12; best configuration L=3 chosen by authors
    The algorithm always returns L=Lmax clusters (Section 6: 'the number of clusters (L) found always matched Lmax'); the conciseness objective does not actually reduce the cluster count, so Lmax is a hand-selected parameter.
  • Lagrange multiplier hyperparameters (lambda0, alpha_lambda, gamma_lambda) = 0.01, 0.005, 1e-4
    Chosen to enforce the coherence constraint; the ablation in supplementary Tables 5-6 shows that without the ascent the method trades coherence for representativeness.
  • Evolutionary parameters (epsilon0, sm, pm, N) = 0.0-0.4, 0.1-0.3, 0.1, 4-8
    Exploration-exploitation tuning of Algorithm 2 (supplementary Table 3); the authors state that for higher Lmax, memory size, learning rates and iterations had a positive effect.
  • Cluster value system weights W_l = e.g., (0.02, 0.05, 0.92), (0.70, 0.04, 0.26), (0.05, 0.89, 0.06) for L=3
    These are the fitted outputs used to interpret clusters; the qualitative claims about business vs. shopping are read directly from them.
assumptions (4)
  • domain assumption A socially-agreed grounding function exists for the society (single shared semantics of values).
    Section 4: 'within human societies and certain application domains there exist typically a socially-agreed grounding.' The whole bi-level formulation depends on this.
  • ad hoc to paper The value labels (time, cost, comfort) and their grounding in route features are correct and sufficient.
    Section 6: comfort is 'presumed' to depend on headway and interchanges; the grounding dataset DV is constructed from these assumptions rather than elicited from agents.
  • domain assumption Value systems are linear scalarizations of value alignments.
    Definition 4 restricts fj to positive normalized linear weights; the supplementary counterexample shows this is restrictive (some coherent groundings cannot be linearly represented).
  • domain assumption Pairwise preferences follow the Bradley-Terry model.
    Eq. (1) uses the Bradley-Terry model for preferences; standard in RLHF but not tested against alternative preference models for this dataset.
invented entities (2)
  • Socially-agreed grounding function GV
    purpose: Represents the consensus meaning of each value label across the society (Definitions 1-2, Section 4).
    No falsifiable handle outside the paper: GV is fitted to labels constructed by the authors from route features, and no independent test of social agreement is provided.
  • Society value system VS_J,L,beta_V
    purpose: Formalizes a society's values as a family of L group value systems plus an assignment function beta (Definition 6).
    The central formal construct of the paper; empirical support is limited to one train-choice dataset, and the correlation with trip intentions is a post-hoc reading of the fitted clusters.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning the Value Systems of Societies from Preferences." pith.science (2026). https://pith.science/paper/VPVZYCJ7

@misc{pith2026250720728,
  author       = {Pith},
  title        = {Pith review of: Learning the Value Systems of Societies from Preferences},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VPVZYCJ7}},
  note         = {Machine review of arXiv:2507.20728}
}
read the original abstract

Aligning AI systems with human values and the value-based preferences of various stakeholders (their value systems) is key in ethical AI. In value-aware AI systems, decision-making draws upon explicit computational representations of individual values (groundings) and their aggregation into value systems. As these are notoriously difficult to elicit and calibrate manually, value learning approaches aim to automatically derive computational models of an agent's values and value system from demonstrations of human behaviour. Nonetheless, social science and humanities literature suggest that it is more adequate to conceive the value system of a society as a set of value systems of different groups, rather than as the simple aggregation of individual value systems. Accordingly, here we formalize the problem of learning the value systems of societies and propose a method to address it based on heuristic deep clustering. The method learns socially shared value groundings and a set of diverse value systems representing a given society by observing qualitative value-based preferences from a sample of agents. We evaluate the proposal in a use case with real data about travelling decisions.

Figures

Figures reproduced from arXiv: 2507.20728 by the authors.

Figure 1
Figure 1. Normalized Dunn Index (scaled down by the maximum found), representativeness, and conciseness for experiments with Lmax ranging from 2 to 12. Each point shows the average and standard error over 10 seeds. other values. The second-largest cluster (∼88 agents) conveys a mix of comfort and time efficiency (26% and 70%, respectively), while the smallest group prioritizes cost (>89%). Both smaller clusters achieve around… view at source ↗
Figure 2
Figure 2. Learning curves for Dunn Index, representativeness, conciseness and grounding coherence of the best found clustering at each iteration (in terms of, first, coherence, and then Dunn Index): averages and standard errors from ten experiments ran with Lmax = 3 and different seeds. We finish with a qualitative analysis for this case, in the right side of [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 38 canonical work pages

  1. [1]

    Anderson and S

    M. Anderson and S. L. Anderson. Geneth: A general ethical dilemma analyzer. Paladyn, 9:337–357, 2 2018. doi: 10.1515/PJBR-2018-0024

    work page doi:10.1515/pjbr-2018-0024 2018

  2. [2]

    Aydo ˘gan, O

    R. Aydo ˘gan, O. Kafali, F. Arslan, C. M. Jonker, and M. P. Singh. Nova: Value-based negotiation of norms.ACM Trans. Intell. Syst. Technol., 12 (4), Aug. 2021. ISSN 2157-6904. doi: 10.1145/3465054

    doi:10.1145/3465054 2021

  3. [3]

    Chakraborty, J

    S. Chakraborty, J. Qiu, H. Yuan, A. Koppel, D. Manocha, F. Huang, A. Bedi, and M. Wang. MaxMin-RLHF: Alignment with diverse human preferences. In Proc. 41st Int. Conf. on Machine Learning , volume 235 of Proceedings of Machine Learning Research, pages 6116–6135. PMLR, 21–27 Jul 2024

    work page 2024

  4. [4]

    P. F. Christiano, J. Leike, T. B. Brown, M. Martic, S. Legg, and D. Amodei. Deep reinforcement learning from human preferences. In Proc. NIPS’17, page 4302–4310, 2017

    work page 2017

  5. [5]

    Cotter, H

    A. Cotter, H. Jiang, and K. Sridharan. Two-player games for efficient non-convex constrained optimization. In A. Garivier and S. Kale, edi- tors, Proceedings of the 30th International Conference on Algorithmic Learning Theory, volume 98 of Proceedings of Machine Learning Re- search, pages 300–332. PMLR, 22–24 Mar 2019

    work page 2019

  6. [6]

    D. L. Davies and D. W. Bouldin. A cluster separation measure. IEEE Transactions on Pattern Analysis and Machine Intelligence, PAMI-1(2): 224–227, 1979. doi: 10.1109/TPAMI.1979.4766909

    arXiv 1979

  7. [7]

    De Giorgis, A

    S. De Giorgis, A. Gangemi, and R. Damiano. Basic human values and moral foundations theory in valuenet ontology. In O. Corcho, L. Hollink, O. Kutz, N. Troquard, and F. J. Ekaputra, editors, Knowl- edge Engineering and Knowledge Management, pages 3–18. Springer,

    work page

  8. [8]

    J. C. Dunn. Well-separated clusters and optimal fuzzy parti- tions. Journal of Cybernetics , 4(1):95–104, 1974. doi: 10.1080/ 01969727408546059

    work page 1974

Show all 48 references
  1. [9]

    D. Fogel. Using evolutionary programing to create neural networks that are capable of playing tic-tac-toe. In IEEE International Conference on Neural Networks, pages 875–880 vol.2, 1993. doi: 10.1109/ICNN. 1993.298673

    1993

  2. [10]

    Grenfell

    M. Grenfell. Pierre Bourdieu: key concepts. Routledge, 2014

    2014

  3. [11]

    J. Haidt. The new synthesis in moral psychology. science, 316(5827): 998–1002, 2007

    2007

  4. [12]

    Holgado-Sánchez, J

    A. Holgado-Sánchez, J. Bajo, H. Billhardt, S. Ossowski, and J. Arias. Value learning for value-aligned route choice modeling via inverse re- inforcement learning. In N. Osman and L. Steels, editors, Value Engi- neering in Artificial Intelligence , pages 40–60, Cham, 2025. Spri...

    2025 doi

  5. [13]

    Jannoud, Y

    I. Jannoud, Y . Jaradat, M. Z. Masoud, A. Manasrah, and M. Alia. The role of genetic algorithm selection operators in extending wsn stability period: A comparative study. Electronics, 11(1), 2022. doi: 10.3390/ electronics11010028

    2022

  6. [14]

    Karanik, H

    M. Karanik, H. Billhardt, A. Fernández, and S. Ossowski. On the rel- evance of value system structure for automated value-aligned decision- making. In Proceedings of the 39th ACM/SIGAPP Symposium on Ap- plied Computing, pages 679–686. Association for Computing Machin- ery, 202...

    2024

  7. [15]

    Kishikawa and S

    D. Kishikawa and S. Arai. Multi-Objective Deep Inverse Reinforcement Learning through Direct Weights and Rewards Estimation. 2022 61st Annual Conference of the Society of Instrument and Control Engineers of Japan, SICE 2022, pages 122–127, 2022. doi: 10.23919/SICE56594. 2022.9905799

    2022

  8. [16]

    Leike, D

    J. Leike, D. Krueger, T. Everitt, M. Martic, V . Maini, and S. Legg. Scal- able agent alignment via reward modeling: a research direction. ArXiv, abs/1811.07871, 2018

    2018 arXiv

  9. [17]

    R. X. Lera-Leri, E. Liscio, F. Bistaffa, C. M. Jonker, M. Lopez-Sanchez, P. K. Murukannaiah, J. A. Rodríguez-Aguilar, and F. Salas-Molina. Ag- gregating value systems for decision support. Knowledge-Based Sys- tems, 287:111453, 2024. doi: 10.1016/j.knosys.2024.111453

    2024

  10. [18]

    X. Li, R. Zhou, Z. C. Lipton, and L. Leqi. Personalized language mod- eling from personalized human feedback, 2024. URL https://arxiv.org/ abs/2402.05133

    2024 arXiv

  11. [19]

    Liscio, M

    E. Liscio, M. van der Meer, L. C. Siebert, C. M. Jonker, and P. K. Mu- rukannaiah. What values should an agent align with?: An empirical comparison of general and context-specific values. Autonomous Agents and Multi-Agent Systems, 36, 2022. doi: 10.1007/s10458-022-09550-0

    2022 doi

  12. [20]

    Loreggia, N

    A. Loreggia, N. Mattei, F. Rossi, and K. B. Venable. Metric learning for value alignment. In CEUR Workshop Proceedings, volume 2419, 2019

    2019

  13. [21]

    López-García

    A. López-García. A proposal for selecting the most value-aligned pref- erences in decision-making using agreement solutions. In Proc. Int. Conf. on Agents and Artificial Intelligence, page 461 – 470, 2024. doi: 10.5220/0012586300003636

    2024 doi

  14. [22]

    Macedo and L

    P. Macedo and L. M. Camarinha-Matos. A qualitative approach to as- sess the alignment of value systems in collaborative enterprises net- works. Computers and Industrial Engineering , 64:412 – 424, 2013. doi: 10.1016/j.cie.2012.09.019

    2013 doi

  15. [23]

    Montes and C

    N. Montes and C. Sierra. Synthesis and properties of optimally value- aligned normative systems. Journal of Artificial Intelligence Research, 74:1739–1774, 2022. doi: 10.1613/jair.1.13487

    2022 doi

  16. [24]

    Montes, N

    N. Montes, N. Osman, C. Sierra, and M. Slavkovik. Value engineering for autonomous agents. CoRR, abs/2302.08759, 2023. doi: 10.48550/ arXiv.2302.08759

    2023 doi arXiv

  17. [25]

    N. Mu, Y . Luan, and Q. S. Jia. Preference-based Multi-Objective Re- inforcement Learning with Explicit Reward Modeling. Proceedings - 2024 China Automation Congress, CAC 2024, pages 4874–4879, 2024. doi: 10.1109/CAC63892.2024.10865310

    2024

  18. [26]

    A. Y . Ng and S. J. Russell. Algorithms for inverse reinforcement learn- ing. In Proceedings of the Seventeenth International Conference on Machine Learning, ICML ’00, page 663–670, San Francisco, CA, USA,

  19. [27]

    F. Nielsen. On a generalization of the jensen–shannon divergence and the jensen–shannon centroid. Entropy, 22(2), 2020. doi: 10.3390/ e22020221

    2020

  20. [28]

    Osman and M

    N. Osman and M. d’Inverno. A computational framework of human values. In Proc. AAMAS’24, pages 1531–1539, 2024

    2024

  21. [29]

    Peschl, A

    M. Peschl, A. Zgonnikov, F. A. Oliehoek, and L. C. Siebert. Moral: Aligning ai with human norms through multi-objective reinforced ac- tive learning. In Proc. Int. Joint Conf. on Autonomous Agents and Mul- tiagent Systems, AAMAS, volume 2, page 1038 – 1046, 2022

    2022

  22. [30]

    L. Qiu, Y . Zhao, J. Li, P. Lu, B. Peng, J. Gao, and S.-C. Zhu. Valuenet: A new dataset for human value driven dialogue system. In Proceed- ings of the 36th AAAI Conference on Artificial Intelligence, AAAI 2022, volume 36, page 11183 – 11191, 2022

    2022

  23. [31]

    Ray and R

    S. Ray and R. Turi. Determination of number of clusters in k-means clustering and application in colour image segmentation. In 4th Inter- national Conference on Advances in Pattern Recognition and Digital Techniques (ICAPRDT’99), pages 137 – 143, India, 2000. Narosa Pub- lishi...

    2000

  24. [32]

    Rodriguez-Soto, M

    M. Rodriguez-Soto, M. Serramia, M. Lopez-Sanchez, and J. A. Rodriguez-Aguilar. Instilling moral value alignment by means of multi- objective reinforcement learning. Ethics and Information Technology, 24:9, 3 2022. ISSN 1388-1957. doi: 10.1007/s10676-022-09635-0

    2022 doi

  25. [33]

    Rodriguez-Soto, N

    M. Rodriguez-Soto, N. Osman, C. Sierra, N. Montes, J. Mar- tinez Roldan, R. Cintas Garcia, C. Farriols Danes, M. Garcia Re- tortillo, and S. Minguez Maso. User study design for identifying the semantics of bioethical principles. In Value Engineering in Arti- ficial Intelligenc...

    2025

  26. [34]

    S. Russell. Artificial intelligence and the problem of control. In H. Werthner, E. Prem, E. A. Lee, and C. Ghezzi, editors, Perspectives on Digital Humanism, pages 19–24. Springer, 2022

    2022

  27. [35]

    S. H. Schwartz. Schwartz value survey. Journal of Cross-Cultural Psy- chology, 2005

    2005

  28. [36]

    Serramia, M

    M. Serramia, M. Lopez-Sanchez, J. A. Rodriguez-Aguilar, M. Ro- driguez, M. Wooldridge, J. Morales, and C. Ansotegui. Moral values in norm decision making. IFAAMAS, 9, 2018

    2018

  29. [37]

    Serramia, M

    M. Serramia, M. Lopez-Sanchez, and J. A. Rodriguez-Aguilar. A quali- tative approach to composing value-aligned norm systems. In Proceed- ings of the 19th International Conference on Autonomous Agents and MultiAgent Systems, pages 1233–1241. IFAAMAS, 2020

    2020

  30. [38]

    Shahin, W

    M. Shahin, W. Hussain, A. Nurwidyantoro, H. Perera, R. Shams, J. Grundy, and J. Whittle. Operationalizing human values in software engineering: A survey. IEEE Access, 10:75269 – 75295, 2022. doi: 10.1109/ACCESS.2022.3190975

    2022

  31. [39]

    L. C. Siebert, E. Liscio, P. K. Murukannaiah, L. Kaptein, S. Spruit, J. V . D. Hoven, and C. Jonker. Estimating value preferences in a hybrid participatory system. Frontiers in Artificial Intelligence and Applica- tions, 354:114 – 127, 2022. doi: 10.3233/FAIA220193

    2022 doi

  32. [40]

    N. Soares. The value learning problem. Artificial Intelligence Safety and Security, 2018

    2018

  33. [41]

    T. R. Sumers, R. D. Hawkins, M. K. Ho, T. L. Griffiths, and D. Hadfield- Menell. How to talk so ai will learn: Instructions, descriptions, and autonomy. In Advances in Neural Information Processing Systems, vol- ume 35, 2022

    2022

  34. [42]

    Van Moffaert, M

    K. Van Moffaert, M. Drugan, and A. Nowe. Scalarized multi-objective reinforcement learning: novel design techniques. In 2013 IEEE Sympo- sium on Adaptive Dynamic Programming and Reinforcement Learning (ADPRL), pages 191–199, 2013. doi: 10.1109/ADPRL.2013.6615007

    2013

  35. [43]

    Vrtic and K

    M. Vrtic and K. W. Axhausen. The impact of tilting trains in switzer- land. a route choice model of regional- and long distance public trans- port trips. Report, Zurich, 2002-06. 82nd Annual Meeting of the Trans- portation Research Board

    2002

  36. [44]

    Watson, T

    E. Watson, T. Viana, S. Zhang, B. Sturgeon, and L. Petersson. Towards an end-to-end personal fine-tuning framework for ai value alignment. Electronics (Switzerland), 13, 2024. doi: 10.3390/electronics13204044

    2024 doi

  37. [45]

    C. F. J. Wu. On the Convergence Properties of the EM Algorithm. The Annals of Statistics, 11(1):95 – 103, 1983

    1983

  38. [46]

    the bi-level optimization setup is needed instead of first estimating a coherent grounding and then trying to learn a social value system

    T. Zhi-Xuan, M. Carroll, M. Franklin, and H. Ashton. Beyond prefer- ences in ai alignment. Philosophical Studies, pages 1–51, 2024. Supplementary Material for: Learning the V alue Systems of Societies from Preferences (ECAI 2025 paper id: M6755) Source Code Source code is avai...

    2024

  39. [2000]

    ISBN 1558607072

    Morgan Kaufmann Publishers Inc. ISBN 1558607072

  40. [2022]

    doi: 10.1007/978-3-031-17105-5_1

    ISBN 978-3-031-17105-5. doi: 10.1007/978-3-031-17105-5_1

    doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.