Pith. sign in

REVIEW 3 major objections 4 minor 73 references

PALMs: Using Multi Construct-Grounded Rationales for Modeling Population Preferences in LLMs

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 LLMs align to specific human populations when trained with synthetic rationales grounded in five psychological and cultural constructs, used as masked latent supervision during direct preference optimization, yielding

desk verdict A solid, honest empirical paper with a plausible effect, but the key ablation doesn't yet isolate construct grounding from rationale length and label leakage. read the letter →

arxiv 2608.01458 v1 pith:YILQOADE submitted 2026-08-02 cs.CL

Priyanka Dey , Brihi Joshi , Preyashi Poddar , Jieyu Zhao , Emilio Ferrara This is my paper
classification cs.CL
keywords populationalignmentpluralisticdirectpreferenceoptimizationconstruct-groundedrationalesculturallatentsupervisionpsychologicalconstructsLLMpersonalization
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

The paper sets out to show that a language model can be aligned to a human population—USA, India, Brazil, France, or Italy—not by feeding it demographics or survey answers, but by training it, during direct preference optimization (DPO), on synthetic reasoning chains grounded in five psychological and cultural constructs: personality traits, cultural dimensions, personal values, moral foundations, and primal world beliefs. The rationale tokens are masked from the training loss, so the model absorbs the construct-based reasoning as an inductive bias rather than learning to reproduce particular explanations. The resulting models beat demographic prompting, survey fine-tuning, and a culture-specialized baseline by an average of 8.59% relative across the five populations and four alignment dimensions (personality, values and beliefs, cultural norms, morality), and the gains transfer to personalized reward modeling, population simulation, and social reasoning without task-specific supervision. If the claim holds, established social-science constructs become a reusable scaffolding for pluralistic alignment, and behavioral fine-tuning looks like the weaker route to representing a population.

What carries the argument

The central object is the multi-construct rationale: a concatenation of five construct-specific sub-chains, $r = \langle r_{\text{PERS}}, r_{\text{CULT}}, r_{\text{VAL}}, r_{\text{MOR}}, r_{\text{BEL}} \rangle$, each generated label-conditionally (4–6 evaluative steps that must conclude with the given label) so that one chain reasons through personality, one through cultural dimensions, one through values, one through moral foundations, and one through primal world beliefs. Its load-bearing role is to be latent supervision: in the DPO objective the rationale tokens are masked from the loss, so the model never learns to reproduce a specific explanation; instead the rationale works as structur

What would settle it

Train the same pipeline in two controlled variants: (1) rationales whose construct labels are scrambled (a chain prompted as 'personality' but generated by the values prompt), and (2) rationales transplanted from a different population onto the persona. If alignment scores hold at the same level, construct content is not the operative mechanism. A complementary check: collect the rationales PALMs themselves generate at inference on held-out instances and annotate them for genuine construct coverage at scale—the paper's human validation covers 50 training-time rationales but does not measure wh

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that construct-grounded reasoning injected as masked latent supervision reshapes what a model knows about a population's preferences. For each training instance—a demographic persona, a question, and a chosen/rejected answer pair—a generator model produces a 4–6 step rationale through each of five construct lenses (OCEAN personality, Hofstede cultural dimensions, Schwartz values, moral foundations, primal world beliefs), with the conclusion forced to match the given preference label. The five chains are concatenated into one rationale, placed in the DPO context, and every rationale token is excluded from the loss so only the preferred-response token

Load-bearing premise

The load-bearing premise is that the label-conditioned synthetic rationales actually encode the five psychological and cultural constructs for each population, and that masking their tokens from the DPO loss prevents the model from exploiting them as a longer, label-consistent shortcut; if the rationales are only fluent post-hoc justifications, the reported gains would come from reasoning length or label leakage rather than construct grounding.

Editorial extensions

If this is right

  • If correct, fine-tuning directly on survey or behavioral responses is the wrong inductive signal for population alignment—it can collapse representational diversity—whereas construct-grounded preference optimization consistently improves it.
  • A single 8B model per country, trained only on preference pairs, doubles as a personalized reward model, a population simulator, and a social-reasoning model without task-specific data.
  • The ablations place the marginal value of construct grounding precisely: SFT < SFT+DPO < SFT+DPO+generic rationales < SFT+DPO+multi-construct rationales, so the five construct families contribute beyond generic intermediate reasoning.
  • The method's success implies psychological and cultural theory can serve as a structured inductive bias for alignment, offering a general path to pluralistic alignment that does not require per-task supervision.

Reading between the lines

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

  • A testable extension the paper leaves open: whether the same masked-rationale recipe works for any construct-defined subgroup—political, religious, regional, or professional—not just countries; if the mechanism is what the authors claim, the five constructs are one instance of a general scaffolding method.
  • Because PALMs generate their rationales at inference, their predictions are auditable in a way ordinary preference models are not; one could diagnose misaligned outputs by seeing which construct family the generated rationale leans on—an interpretability dividend the paper does not develop.
  • The English-only training pipeline (non-English instances translated) likely understates the method's ceiling: native-language construct rationales could carry stronger cultural signal, a testable prediction rather than a paper claim.
  • Scope the paper itself flags: all models share one base architecture and one training dataset, and morality is the dimension where intermediate models occasionally match PALMs—so how the recipe fares across architectures, datasets, and fuller construct taxonomies remains untested.
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 introduces Population Aligned Language Models (PALMs), country-specific LLMs trained with DPO on preference pairs augmented by synthetic rationales grounded in five psychological/cultural constructs (OCEAN, Hofstede, Schwartz values, moral foundations, primal world beliefs). The rationales are generated by Qwen2.5-7B-Instruct, conditioned on the ground-truth label, and used as context during DPO while being masked from the loss. The authors report that PALMs outperform demographic prompting, survey fine-tuning (CultureLLM), and generic-rationale controls across five countries on four alignment dimensions, with an average 8.59% relative improvement over the best baseline, and that the gains transfer to personalized reward modeling, population simulation, and SOCIALIQA. The paper includes human validation of the synthetic training rationales and publicly releases code and data.

Significance. If the core mechanism is real, the paper makes a useful contribution: it proposes a reusable inductive bias for pluralistic alignment, showing that training-time rationale supervision can improve population-level preference modeling and transfer to downstream tasks. The idea of masking rationale tokens from the DPO loss while allowing them to condition the answer is a plausible mechanism for injecting structured knowledge. The authors also provide human annotation of rationale quality, a meaningful check on the synthetic data, and the empirical scope (five countries, four dimensions, three downstream applications) is broad. However, the significance hinges on whether the reported gains come from construct grounding or from confounds such as rationale length and label conditioning; the current experimental design does not isolate those factors.

major comments (3)
  1. [§3.3 and Table 2] The comparison between SFT+DPO+Generic and PALMs is confounded by rationale length. The generic-rationale control receives a single 4–6 step chain (Appendix C, BASELINE), while PALMs receives five construct-specific chains of 4–6 steps each, i.e., roughly 5x more rationale tokens. The reported gap (e.g., USA overall 0.699 vs 0.726; India 0.676 vs 0.738) could therefore be due to more reasoning steps or more context, rather than the construct grounding per se. A length-matched control—e.g., five generic chains of 4–6 steps with no construct scaffolding, or a single construct chain of 20–30 steps—is needed to attribute the gain to the construct families. Without this ablation, the paper's central claim is underdetermined.
  2. [§3.4 and Appendix C] The rationales are generated by prompting with the ground-truth label and the instruction that 'the conclusion must match the label' (Appendix C). They are therefore post-hoc justifications, not independent evidence of psychological mechanisms. Masking rationale tokens from the DPO loss (§3.4) does not remove the label shortcut: the answer-token distribution is still conditioned on the rationale tokens in the context, so the model can use the label-consistent rationale as a latent hint. The generic control also uses label-conditioned rationales, so this alone does not distinguish construct grounding from label leakage. A test where the rationale is removed at inference, or where rationales are intentionally mismatched to the construct, would help establish that the mechanism is construct-driven.
  3. [Table 2 and §5.1] The claim that PALMs 'consistently outperform' baselines is too strong given the reported numbers. On France, the best baseline (Gemini-2.5-Flash) achieves overall 0.634±0.019 and PALMs achieves 0.647±0.011, which are within overlapping 95% CIs. On the morality dimension, Gemini beats PALMs in France (0.681 vs 0.650) and Brazil (0.726 vs 0.708). The aggregated 8.59% relative improvement is driven mostly by dimensions where baselines are weak (notably culture). The paper should report per-dimension significance tests or explicitly qualify which cells are not statistically distinguishable, rather than describing the results as consistent across populations and dimensions.
minor comments (4)
  1. [§4, Training Details] The sentence 'We train five separate models, one per target population...' is duplicated verbatim in the same paragraph. Please remove the repetition.
  2. [§5.4 heading] Typo: 'social intellgence' should be 'social intelligence'.
  3. [Appendix E] The human validation covers only the teacher-generated rationales used during training, not the rationales generated by the trained PALMs at inference. Since the paper claims that latent supervision shapes the model's own reasoning, a small human or automatic evaluation of the inference-time rationales (e.g., whether they actually mention the five constructs) would strengthen the evidence. Appendix F provides examples but no aggregate quality measure.
  4. [§3.2/Table 1] The assumption that the five construct families jointly 'span the latent space of preference-relevant variation' is stated without support. This is a strong assumption; the paper should acknowledge it more explicitly in Limitations and ideally test sensitivity to dropping one construct family at a time.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central claims are empirical comparisons against external human-distribution benchmarks, not quantities that reduce to the training rationales by construction.

full rationale

The paper's core claim is that construct-grounded rationales used as masked latent supervision during DPO improve population-level alignment. This is an empirical claim evaluated against held-out human distributions from external instruments (IPIP-120, CultureBank, Pew, WVS) and downstream benchmarks (PersonalizedRewardBench, PRISM, SIMBENCH, SOCIALIQA). I find no step where a predicted quantity is definitionally equal to an input, nor any fitted parameter that is renamed as a prediction. The rationales are indeed generated by prompting with the ground-truth label and the instruction that 'the conclusion must match the label' (Appendix C), and they are longer than the generic-rationale control (five 4-6 step chains vs. one 4-6 step chain). This creates a real experimental confound: the improvement attributed to 'construct grounding' is not cleanly isolated from rationale length or from label-consistent latent hints. However, a confound is not circularity. The paper does not define 'construct grounding' in terms of the output, nor does it fit a parameter to the evaluation data and then report that fit as a prediction. The comparison to the generic-rationale baseline is an honest, if imperfect, ablation; the claimed 8.59% improvement is a measured outcome against external human distributions, not a quantity forced by construction. Self-citations to Dey et al. (2025, 2026) and Joshi et al. (2025) occur in related work and in the choice of evaluation protocol, but they are not load-bearing in the sense of invoking an unverified uniqueness theorem or importing an ansatz that the paper then treats as external. The limitations section appropriately acknowledges the country-level proxy, construct choice, single dataset, and English-only training. Thus, while the mechanistic attribution to construct grounding is underdetermined by the reported ablations, the paper's main empirical findings are self-contained with respect to circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central claim rests on several domain assumptions (country as proxy, construct sufficiency, dataset validity) and on an ad hoc assumption that synthetic label-conditioned rationales faithfully encode the five constructs. No numeric constants are fitted to data to produce the headline results, but training hyperparameters and rationale lengths are hand-chosen. The heaviest burden is the unvalidated rationale-to-construct link, which is only spot-checked on 50 examples.

free parameters (3)
  • DPO temperature beta = 0.1
    Hand-chosen; no sweep or sensitivity analysis reported. Controls the strength of the preference optimization.
  • SFT and DPO epochs = 2 epochs each
    Hand-chosen; no early stopping or epoch sweep described.
  • Rationale length = 4-6 steps per construct; 5 constructs concatenated
    Set by prompt instructions. Multi-construct rationales are roughly five times longer than generic rationales, confounding the construct-grounding ablation.
assumptions (5)
  • domain assumption Country-level abstraction is a valid unit of population for preference alignment
    Adopted in Section 1 footnote and Section 7; the paper acknowledges it flattens regional, linguistic, and socioeconomic variation.
  • ad hoc to paper The five construct families jointly span the latent space of preference-relevant variation
    Section 3.2 selects OCEAN, Hofstede, Schwartz, moral foundations, and primal world beliefs as complementary lenses; no evidence is given that this set is sufficient or that other constructs would not change results.
  • ad hoc to paper Synthetic rationales generated by Qwen2.5-7B-Instruct faithfully instantiate the target constructs at scale
    Section 3.3 and Appendix C: rationales are prompted to match the given label. Appendix E validates only 50 instances with two annotators (mean Cohen's kappa 0.66), so construct fidelity is not established for the full training set.
  • domain assumption Community Alignment training preferences are ground-truth for each country and do not overlap the evaluation surveys
    Section 4: training uses Community Alignment while evaluation uses IPIP-120, CultureBank, PEW, and WVS; no overlap or leakage analysis is provided.
  • ad hoc to paper Masking rationale tokens from the DPO loss prevents the model from using the rationale as a shortcut to the answer label
    Section 3.4: the model still conditions on the full rationale text at train and test time, so the mechanism by which construct grounding (rather than label-consistent text) improves predictions is assumed, not demonstrated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of PALMs: Using Multi Construct-Grounded Rationales for Modeling Population Preferences in LLMs." pith.science (2026). https://pith.science/paper/YILQOADE

@misc{pith2026260801458,
  author       = {Pith},
  title        = {Pith review of: PALMs: Using Multi Construct-Grounded Rationales for Modeling Population Preferences in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YILQOADE}},
  note         = {Machine review of arXiv:2608.01458}
}
read the original abstract

Large language models are being extensively used to simulate individual user behavior, yet faithfully representing a population requires capturing the systematic variation in values, beliefs, and cultural norms that distinguish one group from another. We introduce Population Aligned Language Models (PALMs), a suite of models each aligned to specific populations, covering five countries: USA, India, Brazil, France and Italy. PALMs are created by synthesizing rationales grounded in psychological and cultural constructs and using these as latent supervision during preference tuning for population-specific alignment. Evaluated across four dimensions: personality, values and beliefs, cultural norms, and morality, PALMs consistently outperform baselines, including culture-specialized models, achieving an average of 8.59% relative improvement over the best baseline across all five populations. Notably, construct-grounded rationales outperform both demographic prompting and survey-based fine-tuning, suggesting that grounding preference learning in psychology and culture provides a richer inductive signal than surface-level response distributions. We further demonstrate strong generalization to downstream applications with- out task-specific supervision: outperforming best baselines by 5.19% in personalized reward modeling, 6.34% in population simulation, and showing strong transfer to social reasoning tasks. Datasets and code are available at: https://github.com/limenlp/PALMs.

Figures

Figures reproduced from arXiv: 2608.01458 by the authors.

Figure 1
Figure 1. Overview of the Population Aligned Language Models training and inference pipeline. We generate multi-construct post-hoc rationales grounded in C (see [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. During training, we prompt Qwen2.5-7B-Instruct to produce synthetic construct-grounded rationales for each preference instance. Each rationale decomposes the preference decision across five psychological and cultural construct families, providing structured label-conditioned supervision during DPO optimization. dimension, we select an established survey instru￾ment and measure alignment between model re￾sponse distr… view at source ↗
Figure 3
Figure 3. Pairwise preference prediction accuracy (%) [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Zero-shot performance (Accuracy % w/ 95% [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

73 extracted references · 33 canonical work pages

  1. [1]

    Can LLM s Express Personality Across Cultures? Introducing C ultural P ersonas for Evaluating Trait Alignment

    Dey, Priyanka and Bothra, Aayush and Khanter, Yugal and Zhao, Jieyu and Ferrara, Emilio. Can LLM s Express Personality Across Cultures? Introducing C ultural P ersonas for Evaluating Trait Alignment. Findings of the Association for Computational Linguistics: EMNLP 2025. 2025. doi:10.18653/v1/2025.findings-emnlp.1101

    doi:10.18653/v1/2025.findings-emnlp.1101 2025

  2. [2]

    Social IQ a: Commonsense Reasoning about Social Interactions

    Sap, Maarten and Rashkin, Hannah and Chen, Derek and Le Bras, Ronan and Choi, Yejin. Social IQ a: Commonsense Reasoning about Social Interactions. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP). 2019. doi:10.18653/v1/D19-1454

    doi:10.18653/v1/d19-1454 2019

  3. [3]

    arXiv preprint arXiv:2510.17516 , year=

    Simbench: Benchmarking the ability of large language models to simulate human behaviors , author=. arXiv preprint arXiv:2510.17516 , year=

    arXiv

  4. [4]

    Advances in Neural Information Processing Systems , volume=

    The PRISM alignment dataset: What participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models , author=. Advances in Neural Information Processing Systems , volume=

  5. [5]

    Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    Rewardbench: Evaluating reward models for language modeling , author=. Findings of the Association for Computational Linguistics: NAACL 2025 , pages=

    2025

  6. [6]

    arXiv preprint arXiv:2507.01352 , year=

    Skywork-reward-v2: Scaling preference data curation via human-ai synergy , author=. arXiv preprint arXiv:2507.01352 , year=

    arXiv

  7. [7]

    arXiv preprint arXiv:2604.07343 , year=

    Personalized RewardBench: Evaluating Reward Models with Human Aligned Personalization , author=. arXiv preprint arXiv:2604.07343 , year=

    arXiv

  8. [8]

    Journal of research in personality , volume=

    Measuring thirty facets of the Five Factor Model with a 120-item public domain inventory: Development of the IPIP-NEO-120 , author=. Journal of research in personality , volume=. 2014 , publisher=

    work page 2014

Show all 73 references
  1. [9]

    Spring 2023 Global Attitudes Survey , year =

    2023

  2. [10]

    The Fourteenth International Conference on Learning Representations , year=

    Cultivating Pluralism In Algorithmic Monoculture: The Community Alignment Dataset , author=. The Fourteenth International Conference on Learning Representations , year=

  3. [11]

    Journal of Cross-Cultural Psychology , volume=

    What kinds of value motives guide people in their moral attitudes? The role of personal and prescriptive values at the culture level and individual level , author=. Journal of Cross-Cultural Psychology , volume=. 2015 , publisher=

    2015

  4. [12]

    Annual review of psychology , volume=

    Personal values across cultures , author=. Annual review of psychology , volume=. 2022 , publisher=

    2022

  5. [13]

    Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    Training millions of personalized dialogue agents , author=. Proceedings of the 2018 conference on empirical methods in natural language processing , pages=

    2018

  6. [14]

    2006 , publisher=

    How the mind explains behavior: Folk explanations, meaning, and social interaction , author=. 2006 , publisher=

    2006

  7. [15]

    Folk psychology and the philosophy of mind , pages=

    Folk psychology and the explanation of human behavior 1 , author=. Folk psychology and the philosophy of mind , pages=. 2013 , publisher=

    2013

  8. [16]

    Proceedings of the National Academy of Sciences , volume=

    Using cognitive psychology to understand GPT-3 , author=. Proceedings of the National Academy of Sciences , volume=. 2023 , publisher=

    2023

  9. [17]

    , author=

    The structure of phenotypic personality traits. , author=. American psychologist , volume=. 1993 , publisher=

    1993

  10. [18]

    Advances in Experimental Social Psychology , year=

    Universals in the Content and Structure of Values: Theoretical Advances and Empirical Tests in 20 Countries , author=. Advances in Experimental Social Psychology , year=

  11. [19]

    Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Identifying the human values behind arguments , author=. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  12. [20]

    arXiv preprint arXiv:2601.14172 , year=

    Human values in a single sentence: Moral presence, hierarchies, and transformer ensembles on the schwartz continuum , author=. arXiv preprint arXiv:2601.14172 , year=

  13. [21]

    Daedalus , volume=

    Intuitive ethics: How innately prepared intuitions generate culturally variable virtues , author=. Daedalus , volume=. 2004 , publisher=

    2004

  14. [22]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    Moral foundations of large language models , author=. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing , pages=

    2024

  15. [23]

    arXiv preprint arXiv:2511.11790 , year=

    Differences in the Moral Foundations of Large Language Models , author=. arXiv preprint arXiv:2511.11790 , year=

  16. [24]

    2001 , publisher=

    Culture's consequences: Comparing values, behaviors, institutions and organizations across nations , author=. 2001 , publisher=

    2001

  17. [25]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Investigating cultural alignment of large language models , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  18. [26]

    Advances in Neural Information Processing Systems , editor=

    Chain of Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , editor=. 2022 , url=

    2022

  19. [27]

    2022 , url=

    Eric Zelikman and Yuhuai Wu and Jesse Mu and Noah Goodman , booktitle=. 2022 , url=

    2022

  20. [28]

    arXiv preprint arXiv:2112.00114 , year=

    Show your work: Scratchpads for intermediate computation with language models , author=. arXiv preprint arXiv:2112.00114 , year=

    arXiv

  21. [29]

    arXiv preprint arXiv:2501.12948 , year=

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning , author=. arXiv preprint arXiv:2501.12948 , year=

    arXiv

  22. [30]

    2024 , month = sep, howpublished =

    Learning to Reason with. 2024 , month = sep, howpublished =

    2024

  23. [31]

    International Conference on Learning Representations , volume=

    Let's verify step by step , author=. International Conference on Learning Representations , volume=

  24. [32]

    Advances in neural information processing systems , volume=

    Direct preference optimization: Your language model is secretly a reward model , author=. Advances in neural information processing systems , volume=

  25. [33]

    Advances in Neural Information Processing Systems , volume=

    Culturepark: Boosting cross-cultural understanding in large language models , author=. Advances in Neural Information Processing Systems , volume=

  26. [34]

    Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Gravity: A framework for personalized text generation via profile-grounded synthetic preferences , author=. Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

  27. [35]

    Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Suzhou, China

    Improving language model personas via rationalization with psychological scaffolds , author=. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing, Suzhou, China. Association for Computational Linguistics , year=

    2025

  28. [36]

    Nature Computational Science , volume=

    Generative language models exhibit social identity biases , author=. Nature Computational Science , volume=. 2025 , publisher=

    2025

  29. [37]

    Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    Aligning language models to user opinions , author=. Findings of the Association for Computational Linguistics: EMNLP 2023 , pages=

    2023

  30. [38]

    Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

    Personalizing dialogue agents via meta-learning , author=. Proceedings of the 57th annual meeting of the association for computational linguistics , pages=

  31. [39]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    Robustness and confounders in the demographic alignment of llms with human perceptions of offensiveness , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    2025

  32. [40]

    , author=

    Values as predictors of judgments and behaviors: The role of abstract and concrete mindsets. , author=. Journal of personality and social psychology , volume=. 2009 , publisher=

    2009

  33. [41]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Should you use llms to simulate opinions? quality checks for early-stage deliberation , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  34. [42]

    Sociodemographic Prompting is Not Yet an Effective Approach for Simulating Subjective Judgments with LLM s

    Sun, Huaman and Pei, Jiaxin and Choi, Minje and Jurgens, David. Sociodemographic Prompting is Not Yet an Effective Approach for Simulating Subjective Judgments with LLM s. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computat...

    2025 doi

  35. [43]

    Goldberg , title =

    Lewis R. Goldberg , title =. Personality Psychology in Europe , volume =. 1999 , editor =

    1999

  36. [44]

    Frontiers in Psychology , volume=

    Testing if primal world beliefs reflect experiences—or at least some experiences identified ad hoc , author=. Frontiers in Psychology , volume=. 2020 , publisher=

    2020

  37. [45]

    Advances in experimental social psychology , volume=

    Moral foundations theory: The pragmatic validity of moral pluralism , author=. Advances in experimental social psychology , volume=. 2013 , publisher=

    2013

  38. [46]

    Published as , year=

    The Big-Five trait taxonomy: History, measurement, and theoretical perspectives , author=. Published as , year=

  39. [47]

    Advances in Neural Information Processing Systems , volume=

    Questioning the survey responses of large language models , author=. Advances in Neural Information Processing Systems , volume=

  40. [48]

    Advances in Neural Information Processing Systems , volume=

    Culturellm: Incorporating cultural differences into large language models , author=. Advances in Neural Information Processing Systems , volume=

  41. [49]

    LLM Tropes: Revealing Fine-Grained Values and Opinions in Large Language Models

    Wright, Dustin and Arora, Arnav and Borenstein, Nadav and Yadav, Srishti and Belongie, Serge and Augenstein, Isabelle. LLM Tropes: Revealing Fine-Grained Values and Opinions in Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. ...

    2024 doi

  42. [50]

    Proceedings of the 40th International Conference on Machine Learning , articleno =

    Santurkar, Shibani and Durmus, Esin and Ladhak, Faisal and Lee, Cinoo and Liang, Percy and Hashimoto, Tatsunori , title =. Proceedings of the 40th International Conference on Machine Learning , articleno =. 2023 , publisher =

    2023

  43. [51]

    Proceedings of the 35th annual ACM symposium on user interface software and technology , pages=

    Social simulacra: Creating populated prototypes for social computing systems , author=. Proceedings of the 35th annual ACM symposium on user interface software and technology , pages=

  44. [52]

    Political Analysis , volume=

    Out of one, many: Using language models to simulate human samples , author=. Political Analysis , volume=. 2023 , publisher=

    2023

  45. [53]

    ArXiv , year=

    A Roadmap to Pluralistic Alignment , author=. ArXiv , year=

  46. [54]

    2022 , publisher =

    World Values Survey: Round Seven -- Country-Pooled Datafile Version 5.0 , author =. 2022 , publisher =. doi:10.14281/18241.20 , url =

    2022 doi

  47. [55]

    Conference on Empirical Methods in Natural Language Processing , year=

    CultureBank: An Online Community-Driven Knowledge Base Towards Culturally Aware Language Technologies , author=. Conference on Empirical Methods in Natural Language Processing , year=

  48. [56]

    Benchmarking distributional alignment of large language models , author=. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers) , pages=

    2025

  49. [57]

    International Conference on Learning Representations , volume=

    Bias runs deep: Implicit reasoning biases in persona-assigned llms , author=. International Conference on Learning Representations , volume=

  50. [58]

    Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions

    Suh, Joseph and Jahanparast, Erfan and Moon, Suhong and Kang, Minwoo and Chang, Serina. Language Model Fine-Tuning on Scaled Survey Data for Predicting Distributions of Public Opinions. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Vo...

    2025 doi

  51. [59]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    arXiv

  52. [60]

    arXiv preprint arXiv:2404.10199 , year=

    Culture-gen: Revealing global cultural perception in language models through natural language prompting , author=. arXiv preprint arXiv:2404.10199 , year=

    arXiv

  53. [61]

    An image speaks a thousand words, but can everyone listen? On image transcreation for cultural relevance

    Khanuja, Simran and Ramamoorthy, Sathyanarayanan and Song, Yueqi and Neubig, Graham. An image speaks a thousand words, but can everyone listen? On image transcreation for cultural relevance. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing...

    2024 doi

  54. [62]

    Organizational behavior and human decision processes , volume=

    Regulatory focus theory: Implications for the study of emotions at work , author=. Organizational behavior and human decision processes , volume=. 2001 , publisher=

    2001

  55. [63]

    Social work theories and methods , volume=

    Attachment theory , author=. Social work theories and methods , volume=. 2012 , publisher=

    2012

  56. [64]

    Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

    The ecological fallacy in annotation: Modeling human label variation goes beyond sociodemographics , author=. Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 2: Short Papers) , pages=

  57. [65]

    arXiv preprint arXiv:2306.16388 , year=

    Towards measuring the representation of subjective global opinions in language models , author=. arXiv preprint arXiv:2306.16388 , year=

    arXiv

  58. [66]

    arXiv preprint arXiv:2211.13069 , year=

    Cultural incongruencies in artificial intelligence , author=. arXiv preprint arXiv:2211.13069 , year=

    arXiv

  59. [67]

    Advances in Neural Information Processing Systems , volume=

    Dices dataset: Diversity in conversational ai evaluation for safety , author=. Advances in Neural Information Processing Systems , volume=

  60. [68]

    Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena , url =

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric and Zhang, Hao and Gonzalez, Joseph and Stoica, Ion , booktitle =. Judging LLM-as-a-Judge with MT-Bench and Cha...

  61. [69]

    arXiv preprint arXiv:2410.21276 , year=

    Gpt-4o system card , author=. arXiv preprint arXiv:2410.21276 , year=

    arXiv

  62. [70]

    arXiv preprint arXiv:2504.01127 , year=

    Can LLMs Grasp Implicit Cultural Values? Benchmarking LLMs' Cultural Intelligence with CQ-Bench , author=. arXiv preprint arXiv:2504.01127 , year=

  63. [71]

    , author=

    Lora: Low-rank adaptation of large language models. , author=. Iclr , volume=

  64. [72]

    arXiv preprint arXiv:2510.11408 , year=

    Valid survey simulations with limited human data: The roles of prompting, fine-tuning, and rectification , author=. arXiv preprint arXiv:2510.11408 , year=

  65. [73]

    P rime X : A Dataset of Worldview, Opinion, and Explanation

    Koncel-Kedziorski, Rik and Joshi, Brihi and Paek, Tim. P rime X : A Dataset of Worldview, Opinion, and Explanation. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.1256

    2025 doi

Pith tools

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