pith. sign in

arxiv: 2605.10843 · v2 · pith:BU3UEWO5new · submitted 2026-05-11 · 💻 cs.CL · cs.AI· cs.CY

Training-Free Cultural Alignment of Large Language Models via Persona Disagreement

Huynh Trung Kiet , Dao Sy Duy Minh , Tuan Nguyen , Chi-Nguyen Tran , Phu-Hoa Pham , Nguyen Lam Phu Quy , The Anh Han , Long Tran-Thanh This is my paper

Pith reviewed 2026-05-20 22:15 UTC · model grok-4.3

classification 💻 cs.CL cs.AIcs.CY
keywords cultural alignmentlarge language modelsinference-time steeringpersona agentsdisagreement signalworld values surveyblack-box modelsmoral preferences
0
0 comments X

The pith

Disagreement among survey-grounded persona agents steers black-box LLMs toward cultural alignment at inference time.

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

Large language models often embed cultural preferences that fail to match the diversity of global users. This paper identifies within-country sociodemographic disagreement, rather than consensus, as the main signal for correction. It introduces an inference-time procedure that builds country-specific panels of World Values Survey personas and turns their disagreements into a bounded logit adjustment. The procedure requires no weight updates, no white-box access, and only public data. A reader would care because it supplies a practical route to serve varied moral preferences without the cost of per-country fine-tuning.

Core claim

The paper establishes that within-country sociodemographic disagreement, not consensus, is the primary steering signal for cultural alignment. DISCA instantiates each country as a panel of World-Values-Survey-grounded persona agents and converts their disagreement into a bounded, loss-averse logit correction. Across 20 countries and 7 open-weight backbones ranging from 2B to 70B parameters, this reduces cultural misalignment on MultiTP by 10-24% on the six backbones of 3.8B parameters or larger and by 2-7% on open-ended scenarios, all without changing any model weights.

What carries the argument

DISCA, the mechanism that extracts a bounded logit correction from disagreement among a panel of World-Values-Survey-grounded persona agents to steer model outputs.

If this is right

  • Cultural alignment becomes possible for commercial black-box APIs that expose only text outputs.
  • No per-country preference datasets or fine-tuning budgets are required.
  • The same procedure works across model scales from 3.8B upward and across 20 countries.
  • Open-ended generation tasks receive measurable alignment gains of 2-7%.
  • Inference-time calibration offers a scalable route to address the long tail of global moral preferences.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The disagreement signal could be reused to align models on other value dimensions such as political or ethical stances.
  • Real-time location or context cues could trigger country-specific persona panels inside deployed assistants.
  • The approach might reduce the engineering burden of maintaining separate regional model variants.
  • Direct API experiments on closed-source models would test whether the black-box gains observed on open weights generalize.

Load-bearing premise

Disagreement among World-Values-Survey-grounded persona agents constitutes a reliable, primary steering signal that can be converted into an effective bounded logit correction for cultural alignment in black-box models.

What would settle it

Apply DISCA to a held-out cultural benchmark unrelated to the World Values Survey and measure whether misalignment scores on that benchmark fall relative to the unadjusted baseline.

Figures

Figures reproduced from arXiv: 2605.10843 by Chi-Nguyen Tran, Dao Sy Duy Minh, Huynh Trung Kiet, Long Tran-Thanh, Nguyen Lam Phu Quy, Phu-Hoa Pham, The Anh Han, Tuan Nguyen.

Figure 1
Figure 1. Figure 1: DISCA overview. Stage 1 builds WVS-grounded persona prompts for a trolley scenario in country c; Stage 2 runs a frozen large language model (LLM) on the base prompt and each persona, aggregates persona-level signals in logit space, and applies Prospect-Theory importance sampling (PT–IS) together with a dual-pass reliability gate to obtain the final sparing probability. Pseudocode and the six MultiTP attrib… view at source ↗
Figure 2
Figure 2. Figure 2: Per-dimension DISCA improvement across the seven headline backbones. Each cell is the macro-averaged (over 20 countries) reduction in per-dimension MPR error: ∆ = |vanilla − human| − |DISCA − human|. Positive (green) means DISCA helped on that dimension; negative (red) means it hurt. Utilitarianism, Species, and Social Value are the dimensions where DISCA delivers the largest gains, consistent with these b… view at source ↗
Figure 3
Figure 3. Figure 3: Geometric story: DISCA pulls model AMCE vectors toward the human cluster. 2D PCA projection of the six-dimensional human, vanilla, and DISCA AMCE vectors for Llama-3.3- 70B across all 20 countries (joint fit, two components capture 93.2% of the variance). Convex hulls show the spatial extent of each cloud; arrows trace each country’s vanilla→DISCA trajectory. All 20 of 20 country points end closer to the h… view at source ↗
Figure 4
Figure 4. Figure 4: Geographic distribution of DISCA gain. Each marker is one of the 20 paper countries placed at its longitude/latitude; marker size is proportional to |∆MIS| and color encodes sign (green = DISCA helped, red = hurt). Aggregated across the seven headline backbones, 19 of 20 countries see a positive mean gain, distributed across the Americas, East and Southeast Asia, and Eastern Europe; the largest single-coun… view at source ↗
Figure 5
Figure 5. Figure 5: Cost-vs-quality frontier on the headline 7 models. Per-scenario DISCA latency (log scale) vs. mean DISCA MIS. Marker size is proportional to parameter count; color encodes ∆MIS (greener = larger DISCA gain). Phi-4 (14B, ∆ = +0.108) lies bottom-left: Pareto-dominant over Llama-3.3-70B in both latency and alignment. A16 Relationship to Persona-Dependent LLM Alignment Kim et al. [2025] is the closest prior wo… view at source ↗
read the original abstract

Large language models increasingly mediate decisions that turn on moral judgement, yet a growing body of evidence shows that their implicit preferences are not culturally neutral. Existing cultural alignment methods either require per-country preference data and fine-tuning budgets or assume white-box access to model internals that commercial APIs do not expose. In this work, we focus on this realistic black-box, public-data-only regime and observe that within-country sociodemographic disagreement, not consensus, is the primary steering signal. We introduce DISCA (Disagreement-Informed Steering for Cultural Alignment), an inference-time method that instantiates each country as a panel of World-Values-Survey-grounded persona agents and converts their disagreement into a bounded, loss-averse logit correction. Across 20 countries and 7 open-weight backbones (2B--70B), DISCA reduces cultural misalignment on MultiTP by 10--24% on the six backbones >=3.8B, and 2--7% on open-ended scenarios, without changing any weights. Our results suggest that inference-time calibration is a scalable alternative to fine-tuning for serving the long tail of global moral preferences.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The paper claims that within-country sociodemographic disagreement (rather than consensus) supplies the primary steering signal for cultural alignment. It introduces DISCA, an inference-time black-box method that instantiates each of 20 countries as a panel of World-Values-Survey-grounded personas, converts their disagreement into a bounded loss-averse logit correction, and reports 10-24% reduction in cultural misalignment on the MultiTP benchmark for the six backbones >=3.8B (plus 2-7% on open-ended scenarios) across seven open-weight models from 2B to 70B, all without weight updates or per-country fine-tuning data.

Significance. If the central claim holds after the required controls, the work would demonstrate a practical, training-free route to culturally adaptive LLM behavior that relies only on public survey data and black-box access. This is potentially significant for serving long-tail global preferences at inference time and for shifting emphasis from consensus-based to disagreement-based steering signals in alignment research.

major comments (2)
  1. [Abstract and §3] Abstract and §3 (method description): the claim that 'within-country sociodemographic disagreement, not consensus, is the primary steering signal' is load-bearing for the entire contribution, yet the manuscript provides no ablation that isolates the disagreement-derived logit correction from a plain average or consensus logit over the identical persona panel. Without such a control (or a non-disagreement baseline), the reported 10-24% MultiTP gains could be explained by multi-persona prompting alone.
  2. [§4] §4 (experiments): baseline construction, statistical significance tests, and prompt-sensitivity controls are not described for the MultiTP results across 20 countries and seven backbones. These details are required to establish that the percentage improvements are robust rather than artifacts of prompt formulation or evaluation protocol.
minor comments (2)
  1. [§3] The precise mathematical form of the bounded logit correction and the loss-aversion factor should be stated explicitly (including any free parameters) rather than summarized at a high level.
  2. Figure or table captions should clarify the exact persona count per country and the number of disagreement samples used to compute the correction.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major point below and describe the revisions we will incorporate to strengthen the work.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (method description): the claim that 'within-country sociodemographic disagreement, not consensus, is the primary steering signal' is load-bearing for the entire contribution, yet the manuscript provides no ablation that isolates the disagreement-derived logit correction from a plain average or consensus logit over the identical persona panel. Without such a control (or a non-disagreement baseline), the reported 10-24% MultiTP gains could be explained by multi-persona prompting alone.

    Authors: We agree that an explicit ablation isolating the disagreement-based logit correction from a consensus or average logit over the same persona panel is necessary to substantiate the central claim. The current manuscript motivates the disagreement signal from the World Values Survey data patterns but does not include this control. In the revised version we will add a direct comparison: DISCA versus a baseline that applies the mean logit across the identical persona panel without the bounded disagreement adjustment. This will quantify the incremental benefit attributable to disagreement modeling rather than multi-persona prompting alone. revision: yes

  2. Referee: [§4] §4 (experiments): baseline construction, statistical significance tests, and prompt-sensitivity controls are not described for the MultiTP results across 20 countries and seven backbones. These details are required to establish that the percentage improvements are robust rather than artifacts of prompt formulation or evaluation protocol.

    Authors: We acknowledge that the experimental section would benefit from additional methodological transparency. In the revision we will expand §4 to specify: (i) the exact construction of all baselines, including how the no-alignment and multi-persona controls were prompted and decoded; (ii) statistical significance testing (paired tests across countries with appropriate multiple-comparison correction); and (iii) prompt-sensitivity results obtained by re-running the evaluation with two additional prompt templates and reporting the range of observed improvements. These additions will demonstrate that the 10–24 % gains are not artifacts of a single prompt or evaluation choice. revision: yes

Circularity Check

0 steps flagged

No circularity: method grounded in external WVS data with independent personas

full rationale

The paper defines DISCA using World Values Survey-grounded personas whose disagreement is converted into a logit correction at inference time. This steering signal is constructed from external survey data and sociodemographic categories chosen independently of any target LLM outputs or fitted parameters. No equations or steps reduce the claimed misalignment reduction to a self-definition, a renamed fit, or a self-citation chain; the reported gains are presented as empirical outcomes of the proposed procedure rather than tautological consequences of its inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The approach rests on external survey data and the modeling choice that disagreement among personas is the dominant useful signal; no new physical entities are postulated.

free parameters (1)
  • logit correction bounds and loss-aversion factor
    The correction is described as bounded and loss-averse, implying tunable limits whose exact values are not stated in the abstract.
axioms (1)
  • domain assumption Within-country sociodemographic disagreement, not consensus, is the primary steering signal for cultural alignment.
    Explicitly stated as the central observation motivating the method.

pith-pipeline@v0.9.0 · 5760 in / 1237 out tokens · 45162 ms · 2026-05-20T22:15:55.178685+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

43 extracted references · 43 canonical work pages · 8 internal anchors

  1. [1]

    M. S. Z. b. Ahmad and K. Takemoto. Large-scale moral machine experiment on large language models. PLOS ONE, 20 0 (5): 0 e0322776, 2025. doi:10.1371/journal.pone.0322776. URL https://doi.org/10.1371/journal.pone.0322776

    work page doi:10.1371/journal.pone.0322776 2025

  2. [2]

    Arditi, O

    A. Arditi, O. Obeso, A. Syed, D. Paleka, N. Panickssery, W. Gurnee, and N. Nanda. Refusal in language models is mediated by a single direction. In Advances in Neural Information Processing Systems, volume 37, pages 136037--136083, 2025. doi:10.52202/079017-4322. URL https://proceedings.neurips.cc/paper_files/paper/2024/hash/f545448535dfde4f9786555403ab7c4...

    work page doi:10.52202/079017-4322 2025

  3. [3]

    Atari, M

    M. Atari, M. J. Xue, P. S. Park, D. E. Blasi, and J. Henrich. Which humans? PsyArXiv preprint, 2023. doi:10.31234/osf.io/5b26t. https://osf.io/preprints/psyarxiv/5b26t

    work page doi:10.31234/osf.io/5b26t 2023

  4. [4]

    E. Awad, S. Dsouza, R. Kim, J. Schulz, J. Henrich, A. Shariff, J.-F. Bonnefon, and I. Rahwan. The moral machine experiment. Nature, 563 0 (7729): 0 59--64, 2018

    work page 2018

  5. [5]

    Chand, F

    S. Chand, F. Baca, and E. Ferrara. No free lunch in language model bias mitigation? T argeted bias reduction can exacerbate unmitigated LLM biases. AI, 7 0 (1): 0 24, 2026. doi:10.3390/ai7010024

    work page doi:10.3390/ai7010024 2026

  6. [6]

    R. Chen, W. Chai, Z. Yang, X. Zhang, Z. Wang, T. Quek, J. T. Zhou, S. Poria, and Z. Liu. D iff PO : Diffusion-styled preference optimization for inference time alignment of large language models. In Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 18910--18925, Vienna, Austria, July 202...

    work page doi:10.18653/v1/2025.acl-long.926 2025

  7. [7]

    Y. Du, S. Li, A. Torralba, J. B. Tenenbaum, and I. Mordatch. Improving factuality and reasoning in language models through multiagent debate. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 11733--11763. PMLR, 2024. URL https://proceedings.mlr.press/v235/du24e.html. arX...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  8. [8]

    KTO: Model Alignment as Prospect Theoretic Optimization

    K. Ethayarajh, W. Xu, N. Muennighoff, D. Jurafsky, and D. Kiela. KTO : Model alignment as prospect theoretic optimization. In International Conference on Machine Learning, 2024. arXiv:2402.01306

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    Gal and Z

    Y. Gal and Z. Ghahramani. Dropout as a B ayesian approximation: Representing model uncertainty in deep learning. In Proceedings of the 33rd International Conference on Machine Learning, volume 48 of Proceedings of Machine Learning Research, pages 1050--1059. PMLR, 2016. URL https://proceedings.mlr.press/v48/gal16.html

    work page 2016

  10. [10]

    C. M. Greco, L. La Cava, and A. Tagarelli. Culturally grounded personas in large language models: Characterization and alignment with socio-psychological value frameworks. arXiv preprint arXiv:2601.22396, 2026

    work page arXiv 2026

  11. [11]

    C. Guo, G. Pleiss, Y. Sun, and K. Q. Weinberger. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning, volume 70 of Proceedings of Machine Learning Research, pages 1321--1330, 2017. URL https://proceedings.mlr.press/v70/guo17a.html. arXiv:1706.04599

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  12. [12]

    Haerpfer, R

    C. Haerpfer, R. Inglehart, A. Moreno, C. Welzel, K. Kizilova, J. Diez-Medrano, M. Lagos, P. Norris, E. Ponarin, and B. Puranen. World Values Survey : Round seven -- country-pooled datafile. Madrid, Spain & Vienna, Austria: JD Systems Institute & WVSA Secretariat, 2020. https://doi.org/10.14281/18241.20

    work page doi:10.14281/18241.20 2020

  13. [13]

    Henrich, S

    J. Henrich, S. J. Heine, and A. Norenzayan. The weirdest people in the world? Behavioral and Brain Sciences, 33 0 (2-3): 0 61--83, 2010. doi:10.1017/S0140525X0999152X. URL https://www.cambridge.org/core/journals/behavioral-and-brain-sciences/article/abs/weirdest-people-inthe-world/BF84F7517D56AFF7B7EB58411A554C17

    work page doi:10.1017/s0140525x0999152x 2010

  14. [14]

    Inglehart and C

    R. Inglehart and C. Welzel. Modernization, Cultural Change, and Democracy: The Human Development Sequence. Cambridge University Press, 2005. URL https://social.hse.ru/data/2012/11/03/1249193128/inglehart_welzel.pdf

    work page 2005

  15. [15]

    Z. Jin, M. Kleiman-Weiner, G. Piatti, S. Levine, J. Liu, F. G. Adauto, F. Ortu, A. Strausz, M. Sachan, R. Mihalcea, Y. Choi, and B. Sch \"o lkopf. Language Model Alignment in Multilingual Trolley Problems . In International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=VEqPDZIDAh. arXiv:2407.02273

    work page arXiv 2025

  16. [16]

    Kahneman and A

    D. Kahneman and A. Tversky. Prospect theory: An analysis of decision under risk. Econometrica, 47 0 (2): 0 263--291, 1979

    work page 1979

  17. [17]

    Kalai and M

    E. Kalai and M. Smorodinsky. Other solutions to N ash's bargaining problem. Econometrica, 43 0 (3): 0 513--518, 1975. doi:10.2307/1914280

    work page doi:10.2307/1914280 1975

  18. [18]

    A. Khan, S. Casper, and D. Hadfield-Menell. Randomness, not representation: The unreliability of evaluating cultural alignment in LLMs . In Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency, pages 2151--2165. Association for Computing Machinery, 2025. doi:10.1145/3715275.3732147. URL https://dl.acm.org/doi/10.1145/371527...

    work page doi:10.1145/3715275.3732147 2025

  19. [19]

    Args: Alignment as reward-guided search

    M. Khanov, J. Burapacheep, and Y. Li. ARGS : Alignment as reward-guided search. In International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=shgx0eqdw6. arXiv:2402.01694

    work page arXiv 2024

  20. [20]

    J. Kim, J. Kwon, L. F. Vecchietti, A. Oh, and M. Cha. Exploring persona-dependent llm alignment for the moral machine experiment. arXiv preprint arXiv:2504.10886, 2025. doi:10.48550/arXiv.2504.10886. URL https://arxiv.org/abs/2504.10886

    work page doi:10.48550/arxiv.2504.10886 2025

  21. [21]

    H. R. Kirk, A. Whitefield, P. R \"o ttger, A. Bean, K. Margatina, J. Ciro, R. Mosquera, M. Bartolo, A. Williams, H. He, B. Vidgen, and S. A. Hale. The PRISM alignment dataset: What participatory, representative and individualised human feedback reveals about the subjective and multicultural alignment of large language models. arXiv preprint arXiv:2404.160...

    work page doi:10.48550/arxiv.2404.16019 2024

  22. [22]

    J. Kwon, L. F. Vecchietti, S. Park, and M. Cha. Dropouts in confidence: Moral uncertainty in human- LLM alignment. In Proceedings of the AAAI Conference on Artificial Intelligence, 2026. arXiv:2511.13290

    work page arXiv 2026

  23. [23]

    S. Levine. Reinforcement learning and control as probabilistic inference: T utorial and review. arXiv preprint arXiv:1805.00909, 2018. doi:10.48550/arXiv.1805.00909. URL https://arxiv.org/abs/1805.00909

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.1805.00909 2018

  24. [24]

    In: Al-Onaizan, Y., Bansal, M., Chen, Y.-N

    T. Liang, Z. He, W. Jiao, X. Wang, Y. Wang, R. Wang, Y. Yang, S. Shi, and Z. Tu. Encouraging divergent thinking in large language models through multi-agent debate. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 17889--17904. Association for Computational Linguistics, 2024. doi:10.18653/v1/2024.emnlp-main....

    work page doi:10.18653/v1/2024.emnlp-main.992 2024

  25. [25]

    P. C. Mahalanobis. Recent experiments in statistical sampling in the I ndian S tatistical I nstitute. Journal of the Royal Statistical Society, 109 0 (4): 0 325--378, 1946

    work page 1946

  26. [26]

    P. J. McCarthy. Pseudo-replication: Half samples. Review of the International Statistical Institute, 37 0 (3): 0 239--264, 1969

    work page 1969

  27. [27]

    Controlled decoding from language models

    S. Mudgal, J. Lee, H. Ganapathy, Y. Li, T. Wang, Y. Huang, Z. Chen, H.-T. Cheng, M. Collins, T. Strohman, J. Chen, A. Beutel, and A. Beirami. Controlled decoding from language models. In Proceedings of the 41st International Conference on Machine Learning, volume 235 of Proceedings of Machine Learning Research, pages 36486--36503. PMLR, 2024. URL https://...

    work page arXiv 2024

  28. [28]

    BLEnD: A Benchmark for LLMs on Ev- eryday Knowledge in Diverse Cultures and Languages,

    J. Myung, N. Lee, Y. Zhou, J. Jin, R. A. Putri, D. Antypas, H. Borkakoty, E. Kim, C. Perez-Almendros, A. A. Ayele, et al. BLEnD : A benchmark for LLMs on everyday knowledge in diverse cultures and languages. In Advances in Neural Information Processing Systems (NeurIPS), Datasets and Benchmarks Track, 2024. doi:10.48550/arXiv.2406.09948. URL https://openr...

    work page doi:10.48550/arxiv.2406.09948 2024

  29. [29]

    J. F. Nash. The bargaining problem. Econometrica, 18 0 (2): 0 155--162, 1950. doi:10.2307/1907266

    work page doi:10.2307/1907266 1950

  30. [30]

    Rudelson and R

    M. Rudelson and R. Vershynin. Hanson-- W right inequality and sub- G aussian concentration. Electronic Communications in Probability, 18 0 (82): 0 1--9, 2013. doi:10.1214/ECP.v18-2865

    work page doi:10.1214/ecp.v18-2865 2013

  31. [31]

    M. J. Ryan, W. Held, and D. Yang. Unintended impacts of LLM alignment on global representation. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 16121--16140, Bangkok, Thailand, Aug. 2024. Association for Computational Linguistics. doi:10.18653/v1/2024.acl-long.853. URL https://aclan...

    work page doi:10.18653/v1/2024.acl-long.853 2024

  32. [32]

    Rohit Saxena and Frank Keller

    S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto. Whose opinions do language models reflect? arXiv preprint arXiv:2303.17548, 2023

    work page arXiv 2023

  33. [33]

    A Roadmap to Pluralistic Alignment

    T. Sorensen, J. Moore, J. Fisher, M. Gordon, N. Mireshghallah, C. M. Rytting, A. Ye, L. Jiang, X. Lu, N. Dziri, et al. A roadmap to pluralistic alignment. arXiv preprint arXiv:2402.05070, 2024. doi:10.48550/arXiv.2402.05070. URL https://arxiv.org/abs/2402.05070

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2402.05070 2024

  34. [34]

    Takemoto

    K. Takemoto. The moral machine experiment on large language models. Royal Society Open Science, 11 0 (2): 0 231393, 2024. doi:10.1098/rsos.231393. URL https://royalsocietypublishing.org/rsos/article/11/2/231393/92489

    work page doi:10.1098/rsos.231393 2024

  35. [35]

    A. M. Turner, L. Thiergart, G. Leech, D. Udell, J. J. Vazquez, U. Mini, and M. MacDiarmid. Activation addition: Steering language models without optimization. arXiv preprint arXiv:2308.10248, 2024. doi:10.48550/arXiv.2308.10248. URL https://arxiv.org/abs/2308.10248

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.10248 2024

  36. [36]

    Tversky and D

    A. Tversky and D. Kahneman. Advances in prospect theory: Cumulative representation of uncertainty. Journal of Risk and Uncertainty, 5 0 (4): 0 297--323, 1992. doi:10.1007/BF00122574. URL https://link.springer.com/article/10.1007/BF00122574

    work page doi:10.1007/bf00122574 1992

  37. [37]

    X. Wang, J. Wei, D. Schuurmans, Q. V. Le, E. H. Chi, S. Narang, A. Chowdhery, and D. Zhou. Self-consistency improves chain of thought reasoning in language models. In International Conference on Learning Representations (ICLR), 2023. doi:10.48550/arXiv.2203.11171. URL https://openreview.net/forum?id=1PL1NIMMrw. arXiv:2203.11171

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2203.11171 2023

  38. [38]

    Williams, P

    G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou. Information-theoretic model predictive control: Theory and applications to autonomous driving. IEEE Transactions on Robotics, 34 0 (6): 0 1603--1622, 2018. doi:10.1109/TRO.2018.2865891. URL https://ieeexplore.ieee.org/abstract/document/8558663

    work page doi:10.1109/tro.2018.2865891 2018

  39. [39]

    K. M. Wolter. Introduction to Variance Estimation. Springer, 2nd edition, 2007

    work page 2007

  40. [40]

    J. Yao, X. Yi, J. Wang, Z. Dou, and X. Xie. CAReDiO : Cultural alignment of LLM via representativeness and distinctiveness guided data optimization. arXiv preprint arXiv:2504.08820, 2025. doi:10.48550/arXiv.2504.08820. URL https://arxiv.org/abs/2504.08820

    work page doi:10.48550/arxiv.2504.08820 2025

  41. [41]

    Zewail, A

    A. Zewail, A. Figueroa, J. Graham, and M. Atari. Moral stereotyping in large language models. Proceedings of the National Academy of Sciences, 123 0 (10): 0 e2519941123, 2026. doi:10.1073/pnas.2519941123. URL https://www.pnas.org/doi/10.1073/pnas.2519941123

    work page doi:10.1073/pnas.2519941123 2026

  42. [42]

    Zhang, X

    B. Zhang, X. Zhao, J. Li, H. Chen, and Z. Chen. Mind the gap in cultural alignment: Task-aware culture management for large language models. arXiv preprint arXiv:2602.22475, 2026

    work page arXiv 2026

  43. [43]

    A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, et al. Representation engineering: A top-down approach to AI transparency. arXiv preprint arXiv:2310.01405, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023