pith. sign in

arxiv: 2605.30036 · v2 · pith:UMGHOMTTnew · submitted 2026-05-28 · 💻 cs.AI · cs.CL

Teaching Values to Machines: Simulating Human-Like Behavior in LLMs

Asaf Yehudai , Naama Rozen , Ariel Gera This is my paper

Pith reviewed 2026-06-29 07:09 UTC · model grok-4.3

classification 💻 cs.AI cs.CL
keywords large language modelshuman valuespsychological questionnairesbehavior simulationvalue alignmentpopulation modelingAI personas
0
0 comments X

The pith

Value-prompted LLMs produce questionnaire responses that align with human value structures and improve population simulations.

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

The paper sets out to test whether LLMs can be given coherent human-like values drawn from established psychological theory and whether those values produce behavior patterns that match real humans. Large-scale experiments administer over five million validated questionnaire items to leading models under value prompts and compare the resulting structures and value-behavior links directly to human data. Strong agreement appears on both the organization of values and their connection to reported behaviors. Using actual distributions of values from human populations further improves how well the models simulate group-level outcomes. A sympathetic reader would care because this suggests LLMs could function as scalable, psychologically grounded stand-ins for studying human responses in value-laden settings.

Core claim

Prompting LLMs with human value profiles from psychological theory leads to questionnaire responses whose value structures and value-behavior relationships show strong agreement with patterns documented in human studies; incorporating empirical human value distributions additionally raises the fidelity of population-level behavioral simulations.

What carries the argument

Value induction through prompts derived from psychological value theory, evaluated by administering validated questionnaires to measure alignment in value structures and value-behavior links.

If this is right

  • Value-induced LLMs can act as proxies for human value-based decisions in controlled experiments.
  • Population simulations gain accuracy when they draw on real human value distributions rather than uniform or synthetic ones.
  • LLMs maintain coherent value structures that align with humans across multiple measured dimensions.
  • Value-behavior relationships observed in humans transfer to the prompted models at both individual and aggregate levels.

Where Pith is reading between the lines

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

  • This method could support testing how different value groups respond to policies or scenarios at scale without recruiting human participants.
  • Value alignment might be checked in open-ended tasks by comparing LLM outputs to human benchmarks on the same value scales.
  • The approach offers a route to quantify how well an AI system's defaults match or diverge from measured human value distributions.

Load-bearing premise

That LLM answers to psychological questionnaires reflect the same underlying value constructs and causal relationships found in humans rather than statistical patterns picked up during training.

What would settle it

Finding that value-prompted LLMs produce inconsistent or non-matching patterns on a fresh set of validated behavioral measures never used in the original prompting or evaluation would indicate the claimed alignment does not hold.

Figures

Figures reproduced from arXiv: 2605.30036 by Ariel Gera, Asaf Yehudai, Naama Rozen.

Figure 1
Figure 1. Figure 1: Overview of our work. We apply value-prompting to induce value-aligned LLMs, leading to coherent [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: The Human Value Theory Continuum: A circular model showing 10 core human values. Adjacent values align, while opposing values conflict. cognitive consistency (Bem, 1972; Sagiv and Roc￾cas, 2021). Understanding this dynamic interplay is central to modeling human action. 3 Experimental Setup Value-prompting To steer LLMs toward a single dominant value, we use a prompting method, value￾prompting, based on Sch… view at source ↗
Figure 3
Figure 3. Figure 3: Behavioral agreement of Llama-3-70B un￾der four high-order values across domains like politics, ethics, and personality. Value-prompting produces dis￾tinct, interpretable behavior patterns, highlighting co￾herent value-behavior relationships in the model. various aspects of an LLM’s “persona”, i.e., be￾havioral characteristics. These behaviors include personality, views on religion, politics, and ethics. E… view at source ↗
Figure 4
Figure 4. Figure 4: (a) Correlation matrix of high-order value vectors for Qwen3-235B-A22B-Instruct, showing human-like [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: (a) MDS map for GPT-OSS-120B, showing a human-like circular structure. (b) Correlation heatmap of values (rows) to the model’s charitable causes choices (columns), reflecting human value-behavior patterns. We compare three settings: (1) Priming Only: reg￾ular value-prompting, (2) Test Only: presenting the filled-out PVQ questionnaire, and (3) Priming & Test: a combination of value-prompting with the filled… view at source ↗
Figure 6
Figure 6. Figure 6: (Part 1/3) [PITH_FULL_IMAGE:figures/full_fig_p018_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: (Part 2/3) [PITH_FULL_IMAGE:figures/full_fig_p019_6.png] view at source ↗
Figure 6
Figure 6. Figure 6: Behavioral agreement of (a) Flan-XXL, (b) LLaMA-3-8B, (c) GPT-oss-20b, (d) GPT-oss-120b, and (e) [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Correlation heatmaps for value vectors for (a) LLaMA-3-8B, (b) LLaMA-3-70B, (c) GPT-OSS-20B, (d) [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Correlation heatmaps for value vectors with value-name only prompts. Correlation heatmaps show only [PITH_FULL_IMAGE:figures/full_fig_p022_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: MDS maps with four different models and population distributions. We can see that all of them exhibit a [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
read the original abstract

Large Language Models (LLMs) demonstrate a remarkable capacity to adopt different personas and roles; however, it remains unclear whether they can manifest behavior that adheres to a coherent, human-like value structure. In this work, we draw on established psychological value theory to induce human-like values in LLMs and assess their alignment with patterns observed in human studies. Using validated psychological questionnaires, we conduct large-scale experiments -- over 5 million questions -- to evaluate value structures and value-behavior relationships in leading LLMs and compare them to humans. Our findings reveal strong agreement between value-prompted LLMs and humans across both dimensions. Moreover, incorporating human value distributions enhances population-level simulations with value-induced LLMs. These findings highlight the potential of value-induced LLMs as effective, psychologically grounded tools for simulating human behavior.

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

3 major / 2 minor

Summary. The paper claims that prompting LLMs with human value distributions drawn from psychological theory produces questionnaire responses whose correlational structure and value-behavior links match those observed in human populations. Large-scale experiments (over 5 million questions) are reported to show strong agreement on both dimensions, and the authors further claim that seeding population-level simulations with human value distributions improves fidelity when using value-induced LLMs.

Significance. If the empirical agreement is robust and not reducible to surface-level mimicry of training data, the work would supply a concrete, questionnaire-grounded method for creating psychologically plausible LLM agents. The scale of the evaluation and the direct comparison to external human datasets are positive features; reproducible code or parameter-free derivations are not mentioned.

major comments (3)
  1. [§3.2] §3.2 (Prompt Construction): The manuscript provides no explicit templates, few-shot examples, or validation procedure for the value-induction prompts. Without this information it is impossible to determine whether the reported agreement arises from induced latent value constructs or from the model simply retrieving statistical patterns already present in its training distribution (which includes many human survey responses).
  2. [§4.1–4.2] §4.1–4.2 (Statistical Analysis): No mention is made of multiple-testing correction, pre-registered analysis plans, or robustness checks across prompt paraphrases and model families. The central claim of “strong agreement” therefore cannot be evaluated for statistical reliability or sensitivity to analysis choices.
  3. [§5] §5 (Population Simulations): The improvement attributed to human value distributions is presented without a clear baseline that holds prompt style and sampling procedure constant while varying only the value distribution. It is therefore unclear whether the gain is due to psychologically grounded value induction or simply better prompt calibration.
minor comments (2)
  1. [Table 1, Figure 2] Table 1 and Figure 2 captions should explicitly state the exact LLMs, temperature settings, and number of samples per condition.
  2. [§3] The abstract states “over 5 million questions” but the methods section does not break this number down by questionnaire, model, or condition; a supplementary table would improve transparency.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for their constructive comments, which highlight important aspects of reproducibility and methodological rigor. We address each major comment below and indicate planned revisions to strengthen the manuscript.

read point-by-point responses

  1. Referee: [§3.2] §3.2 (Prompt Construction): The manuscript provides no explicit templates, few-shot examples, or validation procedure for the value-induction prompts. Without this information it is impossible to determine whether the reported agreement arises from induced latent value constructs or from the model simply retrieving statistical patterns already present in its training distribution (which includes many human survey responses).

    Authors: We agree that explicit documentation of the prompts is essential for evaluating whether the observed alignments reflect induced value constructs. The value-induction method draws directly from validated psychological instruments (e.g., Schwartz Value Survey), and the large-scale reproduction of human correlational structures across independent dimensions makes simple surface retrieval less plausible. In the revised manuscript we will add the complete prompt templates, any few-shot examples, and the validation procedure to an appendix, enabling readers to replicate and assess the induction process. revision: yes

  2. Referee: [§4.1–4.2] §4.1–4.2 (Statistical Analysis): No mention is made of multiple-testing correction, pre-registered analysis plans, or robustness checks across prompt paraphrases and model families. The central claim of “strong agreement” therefore cannot be evaluated for statistical reliability or sensitivity to analysis choices.

    Authors: We will incorporate multiple-testing corrections (FDR) and additional robustness analyses across prompt paraphrases and model families in the revision. These checks will be reported alongside the original results. However, because the study was exploratory rather than confirmatory, no pre-registration was performed; we will state this limitation explicitly. revision: partial

  3. Referee: [§5] §5 (Population Simulations): The improvement attributed to human value distributions is presented without a clear baseline that holds prompt style and sampling procedure constant while varying only the value distribution. It is therefore unclear whether the gain is due to psychologically grounded value induction or simply better prompt calibration.

    Authors: We accept that an explicit control isolating the value distribution is required. In the revised version we will add a baseline condition that keeps prompt phrasing, sampling procedure, and model identical while substituting a non-human (e.g., uniform or model-default) value distribution. This will directly test whether the reported fidelity gains stem from the human value seeding. revision: yes

standing simulated objections not resolved
  • Pre-registration of the analysis plan (the study was exploratory and conducted without prior registration).

Circularity Check

0 steps flagged

No significant circularity; direct empirical comparison to external human datasets

full rationale

The paper's core claims rest on large-scale empirical measurement: value-prompted LLMs are administered validated psychological questionnaires (over 5 million questions), and their response patterns and value-behavior correlations are compared directly to independent human survey data. No equations or derivations reduce reported agreement metrics to quantities fitted from the LLM responses themselves; no self-citations supply load-bearing uniqueness theorems or ansatzes; and the human reference distributions are external benchmarks rather than outputs of the same experimental pipeline. This structure is self-contained against external data and exhibits none of the enumerated circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The central claim rests on the transfer of human psychological constructs to LLMs via prompting and on the assumption that questionnaire responses capture the same constructs in both populations.

axioms (2)
  • domain assumption Established psychological value theory can be induced in LLMs via prompting to produce coherent value structures comparable to humans.
    Invoked when the paper states it draws on psychological value theory to induce values and compares resulting structures to human studies.
  • domain assumption Validated psychological questionnaires measure equivalent constructs when administered to LLMs and to humans.
    Required for the claim of strong agreement across value structures and value-behavior relationships.

pith-pipeline@v0.9.1-grok · 5665 in / 1279 out tokens · 31162 ms · 2026-06-29T07:09:08.652762+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

10 extracted references · 8 canonical work pages · 3 internal anchors

  1. [1]

    Preprint, arXiv:2208.10264

    Using large language models to simulate mul- tiple humans and replicate human subject studies. Preprint, arXiv:2208.10264. Lisa P. Argyle, Ethan C. Busby, Nancy Fulda, Joshua R. Gubler, Christopher Rytting, and David Wingate

    work page arXiv

  2. [2]

    Anat Bardi and Shalom H Schwartz

    Out of one, many: Using language mod- els to simulate human samples.Political Analysis, 31(3):337–351. Anat Bardi and Shalom H Schwartz. 2003. Values and behavior: Strength and structure of relations.Per- sonality and social psychology bulletin, 29(10):1207– 1220. Daryl J Bem. 1972. Self-perception theory.Advances in experimental social psychology, 6. Mar...

    work page arXiv 2003

  3. [3]

    Gian Vittorio Caprara, Guido Alessandri, and Nancy Eisenberg

    Applied multidimensional scaling and unfold- ing. Gian Vittorio Caprara, Guido Alessandri, and Nancy Eisenberg. 2012. Prosociality: the contribution of traits, values, and self-efficacy beliefs.Journal of personality and social psychology, 102(6):1289. Gian Vittorio Caprara, Patrizia Steca, Arnaldo Zelli, and Cristina Capanna. 2005. A new scale for mea- s...

    2012

  4. [4]

    Hyung Won Chung, Le Hou, Shayne Longpre, et al

    The contribution of personality traits and self- efficacy beliefs to academic achievement: A longitu- dinal study.British journal of educational psychol- ogy, 81(1):78–96. Hyung Won Chung, Le Hou, Shayne Longpre, et al

  5. [5]

    Scaling Instruction-Finetuned Language Models

    Scaling instruction-finetuned language models. arXiv preprint arXiv:2210.11416. Ella Daniel and Maya Benish-Weisman. 2019. Value development during adolescence: Dimensions of change and stability.Journal of personality, 87(3):620–632. Francesca Danioni, Daniela Barni, Claudia Russo, Ioana Zagrean, and Camillo Regalia. 2022. Perceived sig- nificant others’...

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  6. [6]

    Towards Measuring the Representation of Subjective Global Opinions in Language Models

    Bringing values back in: The adequacy of the european social survey to measure values in 20 countries.Public opinion quarterly, 72(3):420–445. Esin Durmus, Karina Nguyen, Thomas I Liao, Nicholas Schiefer, Amanda Askell, Anton Bakhtin, Carol Chen, Zac Hatfield-Dodds, Danny Hernandez, Nicholas Joseph, et al. 2023. Towards measuring the representation of sub...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  7. [7]

    Technical report, National Bureau of Economic Research

    Automated social science: Language models as scientist and subjects. Technical report, National Bureau of Economic Research. Jared Moore, Tanvi Deshpande, and Diyi Yang. 2024. Are large language models consistent over value- laden questions?arXiv preprint arXiv:2407.02996. OpenAI, :, Sandhini Agarwal, Lama Ahmad, Jason Ai, Sam Altman, Andy Applebaum, Edwi...

    work page arXiv 2024

  8. [8]

    Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kada- vath, et al

    Training language models to follow instruc- tions with human feedback.Advances in neural in- formation processing systems, 35:27730–27744. Ethan Perez, Sam Ringer, Kamile Lukosiute, Karina Nguyen, Edwin Chen, Scott Heiner, Craig Pettit, Catherine Olsson, Sandipan Kundu, Saurav Kada- vath, et al. 2023. Discovering language model behav- iors with model-writ...

    work page arXiv 2023

  9. [9]

    SocialIQA: Commonsense Reasoning about Social Interactions

    To compete or to cooperate? values’ impact on perception and action in social dilemma games.Eu- ropean Journal of Social Psychology, 41(1):64–77. Leonard Salewski, Stephan Alaniz, Isabel Rio-Torto, Eric Schulz, and Zeynep Akata. 2024. In-context im- personation reveals large language models’ strengths and biases.Advances in Neural Information Process- ing...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    It is important to him/her to take care of people he/she is close to

    A new empirical approach to intercultural com- parisons of value preferences based on schwartz’s theory.Frontiers in Psychology, 11:1723. Asaf Yehudai, Taelin Karidi, Gabriel Stanovsky, Ariel Goldstein, and Omri Abend. 2024. A nurse is blue and elephant is rugby: Cross domain alignment in large language models reveal human-like patterns. Preprint, arXiv:2...

    work page arXiv 2024