Pith. sign in

REVIEW 4 major objections 5 minor 53 references

Some Large Language Models Exhibit Consistent Risk Attitudes

T0 review · 4 major / 5 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Most large language models show stable, domain-general risk attitudes, not task-specific responses.

desk verdict A well-structured empirical framework for measuring LLM risk attitudes, but the headline cross-domain rank stability is likely confounded with a shared numeric-response style and needs control conditions before the trait interpretation holds. read the letter →

arxiv 2607.16197 v1 pith:3573OJCG submitted 2026-04-24 cs.AI

Bowen Sun , Rui Min , Yuxi Wang , Brian Odegaard , Qi Wang , Jing Du This is my paper
classification cs.AI
keywords largelanguagemodelsriskattitudebelief-to-decisionmappingcross-domainstabilityorderedlogisticregressionhuman-AIalignmentsensitivitybehavioralconsistency
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 asks whether large language models have a stable 'risk attitude' — a consistent tendency to turn a perceived level of danger into a more or less cautious action. To test this, the authors strip away the factual side of perception and isolate the mapping from a model's stated risk belief to its categorical decision, fitting that mapping with ordered logistic regression and summarizing it with a single index (area under the curve, where lower means more cautious). Across three structurally different tasks — drone navigation, clinical triage, and financial allocation — five of the six tested models kept the same relative rank order on this index, while one model reversed rank in one task. The same models also produced repeatable beliefs and decisions within each task. The authors conclude that risk attitude is a stable, model-level behavioral trait, and that current LLMs cluster in a narrow band of the human risk-attitude distribution rather than spanning it.

What carries the argument

The central object is the belief-to-decision (BC→RD) mapping: the conditional relationship between a model's stated contextual risk belief (0–100, normalized to [0,1]) and its categorical risk decision (five ordered levels, with 1 = most cautious). The machinery is ordered logistic regression, which fits a monotone S-curve to that mapping and yields two summary indices: the slope β (risk sensitivity) and the area under the fitted curve, AUC (risk attitude bias, range [1,5], lower = more cautious). This decomposition is what lets the authors separate how much risk a model perceives from how it prefers to act on that perception. Cross-domain stability is then tested by ranking models by AUC wi

What would settle it

A direct check would be to re-estimate the AUC ranks after replacing each task's decision discretization with a different but equally valid one (for example, using a raw continuous decision variable for navigation and allocation instead of K-means categories, or using a different number of response categories). If the cross-task rank ordering of models changes substantially under that re-discretization, the apparent trait is an artifact of the shared numeric-response format rather than a stable risk attitude. A second check: run one task with the belief scale changed from 0–100 to, say, 0–10 o

Watch

Extended reading notes

Core claim

The central claim is that, for most LLMs tested, the belief-to-decision mapping — how a reported risk belief on a 0–100 scale becomes a five-level ordinal risk decision — is stable within a task and preserves each model's relative risk posture across tasks. The authors quantify this by fitting an ordered logistic curve to each model's belief–decision pairs, using the curve's slope as risk sensitivity and the area under the curve as risk attitude bias (lower area = more cautious). They report that five of six models keep the same rank ordering of risk attitude bias across drone navigation, clinical triage, and financial allocation, while one model is most aggressive in navigation and most cau

Load-bearing premise

The load-bearing assumption is that the three tasks' ordinal decision scales are commensurable enough that a stable rank ordering of the area-under-the-curve index reflects a domain-general risk trait, rather than a shared tendency to map numeric 0–100 ratings onto five ordered response options in a consistent way.

Editorial extensions

If this is right

  • If a model is characterized as cautious in one decision domain, that characterization can be expected to transfer to other domains, so risk posture can serve as a predictor before deployment in new high-stakes tasks.
  • Risk attitude can be measured independently of factual accuracy or belief calibration, giving evaluators a distinct axis — how the model acts under uncertainty — alongside standard capability benchmarks.
  • Because risk sensitivity is task-dependent while attitude bias is stable, safety assessments should treat the decision threshold (where along the belief scale a model switches) and the responsiveness (how sharply it switches) as separate quantities.
  • Current alignment procedures appear to compress LLM risk postures into a narrow band near the human mean, so aligning behavior with the full range of human risk preferences is an open problem, not a solved one.

Reading between the lines

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

  • The paper leaves implicit that its rank-stability test could serve as a behavioral fingerprint for model identity: two differently fine-tuned versions of the same base model might be distinguished by their AUC ranks even when their benchmark scores match.
  • A natural extension would test whether the stable ordering also appears under purely descriptive, text-only risk scenarios with no dynamic environment, which would indicate whether the trait arises from instruction-following style rather than from genuine decision processes.
  • If the compression toward the human mean is caused by reinforcement-learning-from-human-feedback averaging, one testable prediction is that models with more extreme reward-model sampling would show wider AUC spread; the paper does not test this.
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

4 major / 5 minor

Summary. The paper proposes a cross-domain framework for measuring risk attitudes in LLMs by decomposing decisions into a contextual belief (BC, rated 0–100) and a categorical risk decision (RD, coded 1–5 from cautious to aggressive). Three tasks—drone navigation (DNC), clinical triage (CTD), and financial allocation (FIP)—are administered to six LLMs and 100 human participants. Ordered logistic regression is fit per model–task cell, producing a risk-sensitivity slope and an AUC-based risk-attitude bias. The paper reports (i) intra-task consistency of belief and decision, (ii) cross-domain rank-order stability of AUC, and (iii) a compressed LLM risk-attitude distribution relative to humans, interpreting these as evidence of a stable, model-level risk trait.

Significance. If the claims were fully established, the paper would make a useful contribution to machine psychology and AI alignment: it offers a concrete protocol for separating risk perception from risk choice, compares six contemporary LLMs directly with human participants, and makes trial-level data publicly available. The OLR fits and rank computations are straightforward and reproducible in principle. However, the headline conclusions currently rest on a small-sample rank-stability test that becomes significant only after excluding a discrepant model, on a design that shares response-format features across tasks, and on a human-divergence claim without a formal statistical comparison. The contribution is potentially significant, but the evidence as presented is not yet convincing.

major comments (4)
  1. [Results, Inter-task Universality; Appendix D.7; Table 9] The central evidence for claim (ii) is the sentence 'Kendall's W = 1.00, p = 0.017, excluding Grok 4.' This is a post-hoc exclusion of the only model that reverses rank ordering between DNC and the other two tasks. With all six models, DNC–CTD and DNC–FIP correlations are τ_b = 0.33 (p = 0.469), which are not significant; only CTD–FIP is perfect. Thus the cross-domain rank-stability claim is supported by one task pair and by deleting the most informative observation. Please report full-sample statistics, pre-specify any exclusion rule, and temper the 'most tested LLMs' wording accordingly.
  2. [Results, Inter-task Universality; Materials and Methods D.2; Tables 1–2] The paper asserts that cross-domain invariance 'cannot be explained by shared surface features or task-specific heuristics,' but the three tasks share the same response architecture: a 0–100 belief scale and a five-category decision scale coded in the same cautious-to-aggressive direction. A model with a consistent tendency to choose low (or high) category numbers on any ordered scale, or with a consistent numeric-anchoring style, would produce exactly the observed rank stability without any latent risk disposition. DNC differs in that RD is derived from action logs, but CTD and FIP are direct ordinal labels, and all are pooled into the same 1–5 coding. A non-risk control task with the same response format, a reversed decision-scale condition, or a continuous/binary decision mode is needed to rule out this confound. As written, the central interpretation of cross-domain stability is not
  3. [LLMs vs. Human Risk Baselines; Figure 6] The paper claims that LLM risk attitudes are 'compressed' or 'restricted' relative to the human distribution, but no statistical test of variance or distribution is reported. Figure 6 shows a visual comparison; there is no Levene/Brown–Forsythe test, no permutation or bootstrap interval for the variance ratio, and no numeric summary of the human vs. LLM AUC distributions. Without this, claim (iii) is unsupported. Please add formal comparisons (e.g., variance-ratio test, bootstrapped confidence intervals) and report the underlying distributions.
  4. [Materials and Methods D.4–D.7; Figures 4–5] AUC values are point estimates from fitted OLR curves, and all cross-model ranks are computed from these point estimates. No confidence intervals or bootstrap distributions are provided, so it is unclear whether the observed rank orders—especially the 5/6 perfect concordance—are robust to sampling variability. With six models and 100 trials per cell, the rank-stability analysis should include uncertainty intervals on AUC and on the Kendall coefficients.
minor comments (5)
  1. [Appendix B.7, D.2] The coding of CTD direction is confusing: B.7 says lower ESI values indicate a conservative strategy, while D.2 says 'larger values indicate more aggressive decisions.' Please state explicitly which ESI level maps to category 1 (most cautious) in the analysis.
  2. [Eq. (1) and Eq. (9)] The expected-decision function is written E_i(x) in Eq. (1) but E[RD | B_C] in Section D.4. Define one notation and use it consistently.
  3. [Figure 6] Figure 6 lacks axis labels, units, and error bars. Please add numeric scales and a legend so the claimed compression can be evaluated visually.
  4. [Appendix D.2] The K-means discretization uses k=5 and seed 42. This is a free parameter; please report sensitivity of the AUC values and model ranks to the number of clusters and random seed.
  5. [Data Availability] The repository contains data, but analysis code is only 'available from the corresponding author upon reasonable request.' Please deposit the analysis code alongside the data for reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: cross-task AUC rank stability is an empirical finding, not a construction from the model specification or a self-citation chain.

full rationale

The paper's derivation chain is: per entity and task, fit an ordered logistic regression of the ordinal risk decision on normalized contextual belief (Eqs. 5-6); define risk sensitivity as the slope beta and risk attitude bias as the AUC of the resulting expected-decision curve (Eq. 9); then compare AUC ranks across DNC, CTD, and FIP. Each entity-task cell is fit independently: D.4 states 'Each cell is estimated independently, with no information shared across entities within a task,' and no parameters or fitted values are shared among the three tasks. The cross-domain rank stability (Table 9, Kendall's W and tau) is therefore an empirical outcome of separate fits, not something forced by the equations, by shared fitted parameters, or by a prior citation. The paper does not relabel a fitted value as a held-out prediction: its cross-domain claim is a rank-order comparison of independently estimated AUCs, not an out-of-sample forecast. The references are standard external methods (McCullagh's OLR, NIST RSD) and there is no load-bearing self-citation, imported uniqueness theorem, or ansatz smuggled in by citation. The main validity concern, that the three tasks share a 0-100 belief scale and five ordered response categories so stable ordering could partly reflect a response-style confound, is a substantive interpretive threat but not circularity: it offers an alternative explanation of the empirical pattern rather than showing that the pattern is equivalent to the inputs by construction. The paper itself also flags generalization as open: 'whether the risk attitudes characterized here generalize across this space, and whether a finite, standardized set of paradigms could serve as a comprehensive risk-attitude assessment instrument ... remains an open and important question.' Thus no circular step is exhibited.

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

The analysis introduces no new physical or formal entities. The main extra statistical choices are the K-means discretization parameters and the coding of ESI; the inferential assumptions of OLR and the comparability of task scales/human data are the key domain assumptions. The 'risk attitude' construct is imported from psychology, not newly invented.

free parameters (3)
  • K-means discretization for DNC SI = k=5, seed 42
    The continuous Strategy Index is discretized into five ordinal risk-decision categories using pooled K-means; the choice of k and seed determines category boundaries and therefore AUC for every model in the DNC task.
  • K-means discretization for FIP allocations = k=5, seed 42
    Portfolio allocation vectors are discretized into five ordinal categories using pooled K-means over the simplex; this modeling choice affects the ordinal response and hence the fitted OLR and AUC.
  • CTD ESI-to-risk-category mapping = ESI 1=most cautious, ESI 5=most aggressive
    The paper assigns ESI 1 (most urgent) to the most cautious category, reasoning that early intervention is conservative; an alternative coding would alter AUC values and potentially rank order.
assumptions (5)
  • standard math Proportional odds / parallel slopes assumption in ordered logistic regression
    Invoked in Eq. 2 and Methods D.3/D.4; if the effect of contextual belief is not constant across decision categories, the fitted OLR curves and AUC estimates are misspecified.
  • domain assumption Repeated trials for each model are independent and identically distributed realizations of the same entity-task mapping
    Intra-task consistency and OLR fitting assume that zero-shot trials with memory reset are exchangeable samples of a stable underlying distribution.
  • domain assumption The three tasks share no meaningful surface features, so cross-domain rank stability reflects a general disposition
    Stated in Results (Inter-task Universality); this is required to interpret rank stability as evidence of a latent trait.
  • domain assumption Human and LLM measurements are commensurable in the same analytical framework
    Human data come from graphical interfaces with different (unreported) trial counts per participant; the human-vs-LLM compression result assumes these differences do not bias the comparison.
  • domain assumption The elicited contextual belief (BC) is measured independently enough from the decision to reveal a true mapping
    In FIP, BC and allocation are output in the same JSON; in DNC, BC is retrospective; measurement coupling could inflate the BC→RD association.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Some Large Language Models Exhibit Consistent Risk Attitudes." pith.science (2026). https://pith.science/paper/3573OJCG

@misc{pith2026260716197,
  author       = {Pith},
  title        = {Pith review of: Some Large Language Models Exhibit Consistent Risk Attitudes},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3573OJCG}},
  note         = {Machine review of arXiv:2607.16197}
}
read the original abstract

As artificial intelligence systems are deployed in open-ended, high-stakes settings, a critical dimension remains unmeasured: how perceived risk is translated into action. We test whether large language models (LLMs) exhibit systematic and consistent risk attitudes under uncertainty. We introduce a cross-domain framework that decouples contextual risk belief from categorical decision, and apply it to six representative LLMs and 100 human participants across spatial navigation, clinical triage, and financial allocation tasks. Using regression models, we extract each agents belief-to-decision mapping and quantify risk sensitivity and risk attitude bias. We find that most tested LLMs exhibit (i) robust intra-task consistency, indicating stable mappings from contextual belief to risk decision within a fixed task domain; (ii) cross-domain rank-order stability, preserving relative risk posture across tasks; and (iii) a convergence toward a restricted risk-attitude distribution relative to the broader human baseline. These results reveal risk attitude as a stable and previously uncharacterized dimension of LLM behavior, establishing a foundation for evaluating and aligning AI systems in open-ended decision-making and motivating further investigation into the origins of these intrinsic behavioral dispositions.

Figures

Figures reproduced from arXiv: 2607.16197 by the authors.

Figure 1
Figure 1. Framework for isolating and measuring risk attitude. Agent behavior under uncertainty is decomposed into a sequence of transformations from observations (Ot) to factual belief (BF ), contextual belief (BC), and categorical risk decision (RD). By isolating the mapping from contextual belief to decision (BC → RD), the framework separates risk perception from action, enabling direct measurement of risk attitude indepen… view at source ↗
Figure 2
Figure 2. Cross-domain experimental design for measuring risk attitude. Three struc￾turally distinct decision-making tasks are constructed to share a common analytical structure while differing in domain semantics: drone navigation control (DNC), clinical triage decision (CTD), and financial investment portfolio (FIP). In each task, agents receive sequential observations under un￾certainty, report a contextual risk belief (BC… view at source ↗
Figure 3
Figure 3. Intra-task convergence of contextual belief and risk decision across models and tasks. Each panel shows repeated responses (N = 30 per model) under an identical task condition, with the horizontal axis denoting contextual risk belief (BC) and the vertical axis denoting risk decision (RD). Columns correspond to three representative conditions with increasing task risk (low, moderate, and high), and rows correspond to… view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: Belief-to-decision mappings across tasks and models. For each model, fitted curves show the relationship between contextual belief and ordinal risk decision, with results from the three tasks overlaid (color-coded). Estimated slope parameters are negative (β <ˆ 0) acro…
Figure 5
Figure 5. Figure 5: Cross-domain comparison of risk attitude bias and risk sensitivity. (a) Risk attitude bias for each model, quantified as the area under the fitted contextual belief-to- risk decision curve, with theoretical range [1, 5]. Lower values indicate more risk-averse behavior …
Figure 6
Figure 6. Figure 6: LLM risk attitudes relative to the human behavioral distribution. Human participants span a broad range of risk attitudes across identical task environments, whereas LLMs cluster within a comparatively narrow region of that distribution. This compression indicates that…
Figure 7
Figure 7. Figure 7: Human-subject interface for the Drone Navigation Control (DNC) task. The interface allows participants to observe the navigation environment, control drone movement, and provide a post-trial environmental danger rating. E.5 Human Experiment for Financial Investment Por…
Figure 8
Figure 8. Figure 8: Human-subject interface for the Clinical Triage Decision (CTD) task. The interface presents patient information, evolving vital signs, and controls for entering risk judgments and final ESI decisions. E.6 Alignment Between Human and LLM Experiments The human-subject ex…
Figure 9
Figure 9. Figure 9: Human-subject interface for the Financial Investment Portfolio (FIP) task. The interface presents recent market trajectories and allows participants to report perceived market risk and choose portfolio allocations. Rather than treating human performance as a normative …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 3 canonical work pages

  1. [1]

    The role of ai in emergency department triage: An integrative systematic review.Intensive and Critical Care Nursing, 89:104058,

    Rabie Adel El Arab and Omayma Abdulaziz Al Moosa. The role of ai in emergency department triage: An integrative systematic review.Intensive and Critical Care Nursing, 89:104058,

  2. [2]

    Andrew Taylor, Chris Chmura, Jeremiah Hinson, Benjamin Steinhart, Rohit Sangal, Ar- jun K

    R. Andrew Taylor, Chris Chmura, Jeremiah Hinson, Benjamin Steinhart, Rohit Sangal, Ar- jun K. Venkatesh, Haipeng Xu, Inessa Cohen, Isaac V. Faustino, and Scott Levin. Impact of artificial intelligence–based triage decision support on emergency department care.NEJM AI, 2(3):AIoa2400296, 2025. doi: 10.1056/AIoa2400296. URLhttps://ai.nejm.org/doi/full/ 10.10...

    work page doi:10.1056/aioa2400296 2025

  3. [3]

    Origbo, Ayokunle Osonuga, Eghosasere Egbon, and David B

    Adebayo Da’Costa, Jennifer Teke, Joseph E. Origbo, Ayokunle Osonuga, Eghosasere Egbon, and David B. Olawade. Ai-driven triage in emergency departments: A review of benefits, challenges, and future directions.International Journal of Medical Informatics, 197:105838,

  4. [4]

    High-performance medicine: the convergence of human and artificial intelligence

    Eric J Topol. High-performance medicine: the convergence of human and artificial intelligence. Nature Medicine, 25(1):44–56, 2019

    2019

  5. [5]

    doi: https://doi.org/10.1016/j.ijmedinf.2025.105838

    ISSN 1386-5056. doi: https://doi.org/10.1016/j.ijmedinf.2025.105838. URLhttps: //www.sciencedirect.com/science/article/pii/S1386505625000553

    arXiv 2025

  6. [6]

    Explainable artificial intelligence for crypto asset allocation.Finance Research Letters, 47:102941, 2022

    Golnoosh Babaei, Paolo Giudici, and Emanuela Raffinetti. Explainable artificial intelligence for crypto asset allocation.Finance Research Letters, 47:102941, 2022. ISSN 1544-6123. doi: https://doi.org/10.1016/j.frl.2022.102941. URLhttps://www.sciencedirect.com/science/ article/pii/S1544612322002021

    arXiv 2022

  7. [7]

    Artificial intelligence, dynamic capabilities, and cor- porate financial asset allocation.International Review of Financial Analysis, 96:103773,

    Yu Li, Huiyi Zhong, and Qiye Tong. Artificial intelligence, dynamic capabilities, and cor- porate financial asset allocation.International Review of Financial Analysis, 96:103773,

  8. [8]

    J. B. Heaton, N. G. Polson, and J. H. Witte. Deep learning for finance: Deep portfolios.Applied Stochastic Models in Business and Industry, 33(1):3–12, 2017

    2017

Show all 53 references
  1. [9]

    On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

    Rishi Bommasani, Drew A Hudson, Ehsan Adeli, Russ Altman, Simran Arora, Sydney von Arx, Michael S Bernstein, Jeannette Bohg, Antoine Bosselut, Emma Brunskill, et al. On the opportunities and risks of foundation models.arXiv preprint arXiv:2108.07258, 2021

    2021 arXiv

  2. [10]

    Causal artificial intel- ligence for high-stakes decisions: The design and development of a causal machine learning model.IEEE Access, 10:24327–24339, 2022

    Bukhoree Sahoh, Kanjana Haruehansapong, and Mallika Kliangkhlao. Causal artificial intel- ligence for high-stakes decisions: The design and development of a causal machine learning model.IEEE Access, 10:24327–24339, 2022. doi: 10.1109/ACCESS.2022.3155118

    2022

  3. [11]

    A domain-specific risk-attitude scale: Measuring risk perceptions and risk behaviors.Journal of behavioral decision making, 15(4): 263–290, 2002

    Elke U Weber, Ann-Renee Blais, and Nancy E Betz. A domain-specific risk-attitude scale: Measuring risk perceptions and risk behaviors.Journal of behavioral decision making, 15(4): 263–290, 2002. 32

    2002

  4. [12]

    Risk attitude and preference.Wiley Interdisciplinary Reviews: Cognitive Sci- ence, 1(1):79–88, 2010

    Elke U Weber. Risk attitude and preference.Wiley Interdisciplinary Reviews: Cognitive Sci- ence, 1(1):79–88, 2010

    2010

  5. [13]

    Individual risk attitudes: Measurement, determinants, and behavioral consequences.Jour- nal of the European Economic Association, 9(3):522–550, 2011

    Thomas Dohmen, Armin Falk, David Huffman, Uwe Sunde, Jürgen Schupp, and Gert G Wag- ner. Individual risk attitudes: Measurement, determinants, and behavioral consequences.Jour- nal of the European Economic Association, 9(3):522–550, 2011

    2011

  6. [14]

    Rout- ledge, 2017

    David Hillson and Ruth Murray-Webster.Understanding and managing risk attitude. Rout- ledge, 2017

    2017

  7. [15]

    Genetic variation in preferences for giving and risk taking.Quarterly Journal of Economics, 124(2):809–842, 2009

    David Cesarini, Christopher T Dawes, Magnus Johannesson, Paul Lichtenstein, and Björn Wallace. Genetic variation in preferences for giving and risk taking.Quarterly Journal of Economics, 124(2):809–842, 2009

    2009

  8. [16]

    Assessing the construct validity of risk attitude.Manage- ment Science, 46(10):1337–1348, 2000

    Joost ME Pennings and Ale Smidts. Assessing the construct validity of risk attitude.Manage- ment Science, 46(10):1337–1348, 2000

    2000

  9. [17]

    Decisions from experience and the effect of rare events in risky choice.Psychological Science, 15(8):534–539, 2004

    Ralph Hertwig, Greg Barron, Elke U Weber, and Ido Erev. Decisions from experience and the effect of rare events in risky choice.Psychological Science, 15(8):534–539, 2004

    2004

  10. [18]

    Reasoningthefastandfrugalway: Modelsofbounded rationality.Psychological Review, 103(4):650–669, 1996

    GerdGigerenzerandDanielGGoldstein. Reasoningthefastandfrugalway: Modelsofbounded rationality.Psychological Review, 103(4):650–669, 1996. doi: 10.1037/0033-295X.103.4.650

    1996 doi

  11. [19]

    Who takes risks when and why? Determinants of risk taking

    Bernd Figner and Elke U Weber. Who takes risks when and why? Determinants of risk taking. Current Directions in Psychological Science, 20(4):211–216, 2011

    2011

  12. [20]

    Emergent abilities of large language models.Transactions on Machine Learning Research, 2022

    Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, et al. Emergent abilities of large language models.Transactions on Machine Learning Research, 2022

    2022

  13. [21]

    Probing the psychology of ai models.Proceedings of the National Academy of Sciences, 120(10):e2300963120, 2023

    Richard Shiffrin and Melanie Mitchell. Probing the psychology of ai models.Proceedings of the National Academy of Sciences, 120(10):e2300963120, 2023

    2023

  14. [22]

    Risk as feelings

    George F Loewenstein, Elke U Weber, Christopher K Hsee, and Ned Welch. Risk as feelings. Psychological bulletin, 127(2):267, 2001

    2001

  15. [23]

    Kernels of selfhood: Gpt-4o shows humanlike patterns of cognitive dissonance moderated by free choice.Proceedings of the National Academy of Sciences, 122(20):e2501823122, 2025

    Steven A Lehr, Ketan S Saichandran, Eddie Harmon-Jones, Nykko Vitali, and Mahzarin R Ba- naji. Kernels of selfhood: Gpt-4o shows humanlike patterns of cognitive dissonance moderated by free choice.Proceedings of the National Academy of Sciences, 122(20):e2501823122, 2025

    2025

  16. [24]

    Machine psychology.arXiv preprint arXiv:2303.13988, 2023

    Thilo Hagendorff, Ishita Dasgupta, Marcel Binz, Stephanie CY Chan, Andrew Lampinen, Jane X Wang, Zeynep Akata, and Eric Schulz. Machine psychology.arXiv preprint arXiv:2303.13988, 2023

    2023 arXiv

  17. [25]

    Using cognitive psychology to understand GPT-3.Proceedings of the National Academy of Sciences, 120(6):e2218523120, 2023

    Marcel Binz and Eric Schulz. Using cognitive psychology to understand GPT-3.Proceedings of the National Academy of Sciences, 120(6):e2218523120, 2023

    2023

  18. [26]

    Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Pondé de Oliveira Pinto, Jared Kaplan, Harrison Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. Evaluating large language models trained on code.arXiv preprint arXiv:2107.03374, 2021

    2021 arXiv

  19. [27]

    Artificial intelligence, values, and alignment.Minds and machines, 30(3):411– 437, 2020

    Iason Gabriel. Artificial intelligence, values, and alignment.Minds and machines, 30(3):411– 437, 2020. 33

    2020

  20. [28]

    Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300, 2020

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring massive multitask language understanding.arXiv preprint arXiv:2009.03300, 2020

    2009 arXiv

  21. [29]

    Alan Heckert, James J

    N. Alan Heckert, James J. Filliben, Carroll Croarkin, Barry Hembree, William F. Guthrie, Paul Tobias, and Julia Prinz. Handbook 151: Nist/sematech e-handbook of statistical methods, 2002

    2002

  22. [30]

    Regression models for ordinal data.Journal of the Royal Statistical Society: Series B (Methodological), 42(2):109–127, 1980

    Peter McCullagh. Regression models for ordinal data.Journal of the Royal Statistical Society: Series B (Methodological), 42(2):109–127, 1980

    1980

  23. [31]

    Concrete problems in AI safety.arXiv preprint arXiv:1606.06565, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete problems in AI safety.arXiv preprint arXiv:1606.06565, 2016

    2016 arXiv

  24. [32]

    Influence of safety climate on risk tolerance and risk-taking behavior: A cross-cultural examination.Safety Science, 146:105559, 2022

    S Bhandari and M R Hallowell. Influence of safety climate on risk tolerance and risk-taking behavior: A cross-cultural examination.Safety Science, 146:105559, 2022. doi: 10.1016/j.ssci. 2021.105559

    2022

  25. [33]

    Safety risk tolerance in the construction industry: Cross-cultural analysis.Journal of Construction Engineering and Management, 146 (4):04020022, 2020

    R Salas, M Hallowell, R Balaji, and S Bhandari. Safety risk tolerance in the construction industry: Cross-cultural analysis.Journal of Construction Engineering and Management, 146 (4):04020022, 2020. doi: 10.1061/(ASCE)CO.1943-7862.0001789

    2020

  26. [34]

    Personal characteristics and risk tolerance in a natural experiment

    Peter Brous and Bo Han. Personal characteristics and risk tolerance in a natural experiment. The Journal of Risk Finance, 23(2):155–168, 2022. doi: 10.1108/JRF-11-2021-0176

    2022 doi

  27. [35]

    Deep reinforcement learning from human preferences

    Paul F Christiano, Jan Leike, Tom B Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep reinforcement learning from human preferences. InAdvances in Neural Information Processing Systems, volume 30. Curran Associates, 2017

    2017

  28. [36]

    Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730–27744, 2022

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Advances in neural information processing systems, 35:27730...

    2022

  29. [37]

    Risks and rewards in adolescent decision-making.Current opinion in psychology, 48:101457, 2022

    Anna CK Van Duijvenvoorde, Jorien van Hoorn, and Neeltje E Blankenstein. Risks and rewards in adolescent decision-making.Current opinion in psychology, 48:101457, 2022

    2022

  30. [38]

    Prospect theory: An analysis of decision under risk

    Daniel Kahneman and Amos Tversky. Prospect theory: An analysis of decision under risk. InHandbook of the fundamentals of financial decision making: Part I, pages 99–127. World Scientific, 2013

    2013

  31. [39]

    Advances in prospect theory: Cumulative representation of uncertainty.Journal of Risk and Uncertainty, 5(4):297–323, 1992

    Amos Tversky and Daniel Kahneman. Advances in prospect theory: Cumulative representation of uncertainty.Journal of Risk and Uncertainty, 5(4):297–323, 1992

    1992

  32. [40]

    Perceptionofrisk.Science, 236(4799):280–285, 1987

    PaulSlovic. Perceptionofrisk.Science, 236(4799):280–285, 1987. doi: 10.1126/science.3563507

    1987 doi

  33. [41]

    Take caution in using llms as human surrogates.Proceedings of the National Academy of Sciences, 122(24):e2501660122, 2025

    Yuan Gao, Dokyun Lee, Gordon Burtch, and Sina Fazelpour. Take caution in using llms as human surrogates.Proceedings of the National Academy of Sciences, 122(24):e2501660122, 2025

    2025

  34. [42]

    Whose opinions do language models reflect? InProceedings of the International Conference on Machine Learning, pages 29971–30004

    Shibani Santurkar, Esin Durmus, Faisal Ladd, Cinoo Lee, Percy Liang, and Tatsunori Hashimoto. Whose opinions do language models reflect? InProceedings of the International Conference on Machine Learning, pages 29971–30004. PMLR, 2023. 34

    2023

  35. [43]

    Large language models show amplified cognitive biases in moral decision-making.Proceedings of the National Academy of Sciences, 122(25):e2412015122, 2025

    Vanessa Cheung, Maximilian Maier, and Falk Lieder. Large language models show amplified cognitive biases in moral decision-making.Proceedings of the National Academy of Sciences, 122(25):e2412015122, 2025

    2025

  36. [44]

    Grok 4 model card.https://data.x.ai/2025-08-20-grok-4-model-card.pdf, 2025

    xAI. Grok 4 model card.https://data.x.ai/2025-08-20-grok-4-model-card.pdf, 2025. Accessed: August 2025

    2025

  37. [45]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    2025 arXiv

  38. [46]

    GPT-5.2 system card.https://cdn.openai.com/pdf/ 3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf, 2025

    OpenAI. GPT-5.2 system card.https://cdn.openai.com/pdf/ 3a4153c8-c748-4b71-8e31-aecbde944f8d/oai_5_2_system-card.pdf, 2025. Accessed: December 2025

    2025

  39. [47]

    Gemini 3 pro frontier safety framework report.https://storage

    Google DeepMind. Gemini 3 pro frontier safety framework report.https://storage. googleapis.com/deepmind-media/gemini/gemini_3_pro_fsf_report.pdf, 2025. Accessed: November 2025

    2025

  40. [48]

    Deepseek-v3 technical report, 2024

    DeepSeek-AI. Deepseek-v3 technical report, 2024. URLhttps://arxiv.org/abs/2412.19437

    2024 arXiv

  41. [49]

    System card: Claude Sonnet 4.5.https://www.anthropic.com/ claude-sonnet-4-5-system-card, 2025

    Anthropic. System card: Claude Sonnet 4.5.https://www.anthropic.com/ claude-sonnet-4-5-system-card, 2025. Accessed: September 2025

    2025

  42. [50]

    Ordinal regression models made easy: A tutorial on parameter interpretation, data simulation and power analysis.International Journal of Psychology, 59(6):1263–1292, 2024

    Filippo Gambarota and Gianmarco Altoè. Ordinal regression models made easy: A tutorial on parameter interpretation, data simulation and power analysis.International Journal of Psychology, 59(6):1263–1292, 2024. 35

    2024

  43. [52]

    Wang, and Robert Krovetz

    Yixin Chen, James Z. Wang, and Robert Krovetz. Clue: Cluster-based retrieval of images by unsupervised learning.IEEE Transactions on Image Processing, 14(8):1187–1201, 2005

    2005

  44. [2024]

    doi: https://doi.org/10.1016/j.irfa.2024.103773

    ISSN 1057-5219. doi: https://doi.org/10.1016/j.irfa.2024.103773. URLhttps://www. sciencedirect.com/science/article/pii/S1057521924007051

    2024

  45. [2025]

    doi: https://doi.org/10.1016/j.iccn.2025.104058

    ISSN 0964-3397. doi: https://doi.org/10.1016/j.iccn.2025.104058. URLhttps://www. sciencedirect.com/science/article/pii/S0964339725001193

    2025

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.