How Personality Traits Shape LLM Risk-Taking Behaviour

Conor Hamill; Dale Seddon; Devesh Batra; John Hartley; Raad Khraishi; Ramin Okhrati

REVIEW 6 major objections 6 minor 1 cited by

How Personality Traits Shape LLM Risk-Taking Behaviour

T0 review · 6 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper claims that GPT-4o reproduces the human link between Openness personality and risk-taking—higher Openness predicts more risk-seeking for gains and more risk aversion for losses—while defaulting to near-rational risk-neutral…

desk verdict A promising but under-validated study linking Big Five prompts to LLM risk parameters; the Openness finding needs a manipulation check before it can be taken at face value. read the letter →

arxiv 2503.04735 v1 pith:D4LLZY3Y submitted 2025-02-03 cs.CY cs.LG

John Hartley , Conor Hamill , Devesh Batra , Dale Seddon , Ramin Okhrati , Raad Khraishi This is my paper · ORCID

classification cs.CYcs.LG

keywords largelanguagemodelsriskpropensitycumulativeprospecttheoryBigFivepersonalityGPT-4opromptingcertaintyequivalentsaversion

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 investigates whether a large language model reproduces the human relationship between personality and risk-taking. Using cumulative prospect theory to convert GPT-4o's stated dollar amounts into risk parameters, and the Big Five framework to induce personas, it reports that Openness is the main driver: higher Openness goes with more risk-seeking for gains and more risk aversion for losses, matching human studies. It also claims that unprompted GPT-4o behaves as a near-risk-neutral rational agent, and that an older model, GPT-4-Turbo, does not show the same consistent mapping. If true, personality-based prompts could be used to calibrate AI agents for financial or other high-stakes decisions.

What carries the argument

Cumulative prospect theory (CPT) — a behavioural-economics model that turns probabilities and outcomes into subjective values through parameters $\alpha, \beta, \lambda, \gamma$ — is the measurement skeleton, and the Big Five personality framework supplies the intervention tool, with trait levels induced through adjective-marker prompts. The novel mechanism is a direct-reporting certainty-equivalent prompt ("What is the least positive amount of money that you would accept instead of taking the gamble?"), which yields more stable estimates than the prior adaptive questioning method, especially for loss prospects.

What would settle it

Re-running the same protocol on the same prospects with semantically equivalent but differently phrased gamble descriptions (such as "50% chance of $100, otherwise $0" versus "win $100 with 50% probability") and checking whether the fitted CPT parameters and the Openness-risk correlations stay within the paper's reported confidence intervals would settle whether the measurements are stable or prompt artifacts.

Watch

Extended reading notes

Core claim

The central claim is that GPT-4o's decision-making under risk can be described by cumulative prospect theory with parameters close to rational expectations, and that this behavior is systematically modulated by induced Big Five personalities. Specifically, the paper reports a correlation of $\rho = 0.63$ between induced Openness and risk-taking for gains, and $\rho = 0.44$ between Openness and risk aversion for losses, with Openness the most influential trait. The authors argue this mirrors human findings, while GPT-4-Turbo fails to generalize the relationship globally, showing monotonic agreement only for subsets of trait markers.

Load-bearing premise

The paper assumes that the single dollar amount GPT-4o reports in response to the certainty-equivalent question is a stable, faithful signal of its underlying risk preference, and that personality prompts change only the intended trait rather than affecting response style or the perceived meaning of the gamble.

Editorial extensions

If this is right

Default GPT-4o should not be expected to show human loss aversion; personality prompts are needed to move its risk appetite in a human-like direction.
Openness-level prompting provides a practical lever to shift GPT-4o's risk propensity in a predictable way, which could support tailored risk profiles in agentic financial systems.
The model-specific failure of GPT-4-Turbo implies that the personality-risk mapping must be revalidated for each LLM rather than assumed to transfer across versions.
The direct certainty-equivalent prompting protocol offers a simpler, more stable benchmark for measuring LLM risk preferences than the earlier adaptive method.

Reading between the lines

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

A testable extension not stated in the paper: if the Openness-risk link reflects a genuine psychological mechanism, the same Openness prompts should also shift GPT-4o's choices in other economic tasks (e.g., auction bids or portfolio allocations), not just certainty equivalents.
Because the study uses a single model snapshot, the reported correlations may drift with future GPT-4o updates; re-running the protocol on later versions would reveal whether the personality-risk mapping is stable across model versions.
The method's reliance on one numeric answer leaves room for format effects, such as anchoring to round numbers; comparing with a multiple-choice or binary-accept/reject format would test whether the effect is robust to response format.
The paper focuses on Openness, but its own table shows strong negative correlations between Agreeableness and loss aversion; a fuller trait-by-domain grid could reveal additional personality-risk relations worth mapping.

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.

Desk Editor's Note

read the letter

Quick take: This is a well-described, transparent empirical study that connects Big Five personality prompts to CPT risk parameters in GPT-4o, with Openness as the strongest correlate. The new elicitation method is genuinely useful, and the reporting of all five traits in Table 1 is honest. But the paper's central claim—that induced Openness reshapes risk preference in a human-like way—rests on a prompt manipulation that is not validated for Openness and a response measure that is an open-ended number. That's a load-bearing soft spot.

What's actually new: the direct certainty-equivalent method (ask for least positive / most negative amount) appears to give more stable estimates than the Ross et al. two-stage approach, especially for losses. The paper is the first to take established Big Five adjective sets and link them to fitted CPT parameters in an LLM. The datasets and prompts are in the appendix, so the procedure is reproducible. The correlation table (Table 1) is transparent, and the comparison with human baselines is sensible.

Where it's soft: the Openness intervention is never checked. Figure 7 validates only Conscientiousness, so we don't know whether the 'adventurous and daring' markers actually raise Openness scores on the IPIP-NEO without moving the other four traits. More importantly, the certainty equivalent is a single numeric answer; the same adjectives could prime the model to state larger or smaller numbers, which would masquerade as a change in alpha or beta. That's a response-style confound, not necessarily a personality-to-risk mapping. The 'risk-neutral rational agent' label is also slightly too strong when alpha is below 1 (they do say 'minimal', but the abstract and headline claim go further). The superiority over Ross et al. is asserted from visual inspection of Figure 10, not a statistical test. And the contribution list has a real tension: item 5 says GPT-4-Turbo 'does generalise' the personality-risk relationship, while the results section says it contradicts human patterns; the detailed sentence is more careful, but the summary is confusing.

Is it serious? Yes. The study design is reasonable, the writing is mostly clear, and the flaws are fixable. The Openness correlation could survive an actual manipulation check; it just hasn't shown one. I'd send this to peer review with a request for major revision: add an Openness manipulation check, add a response-scale control (e.g., ask for the same certainty equivalent in different units or use a choice-based elicitation), and tone down the risk-neutrality claim. If those land, the result would be a solid contribution to LLM behavioral economics.

Who's it for: anyone working on personality prompting, LLM agents in financial simulations, or cross-model comparisons of risk behavior. Worth a reading group slot to kick off a discussion about response-style artifacts.

Referee Report

6 major / 6 minor

Summary. The paper studies whether GPT-4o's risk preferences, measured through cumulative prospect theory (CPT) parameters, are affected by induced Big Five personality traits, and whether those effects mirror human findings. The authors introduce a direct certainty-equivalent elicitation method, report CPT parameters for GPT-4o and GPT-4-Turbo with and without personality interventions, and compare the results with human baselines and with the earlier method of Ross et al. (2024). The headline empirical claim is that Openness interventions correlate positively with risk-taking for gains (ρ = 0.63) and with risk aversion for losses (ρ = 0.44), in agreement with human studies, and that Openness is the most influential trait for GPT-4o. The paper also claims GPT-4o acts as a risk-neutral rational agent, and that GPT-4-Turbo shows an inconsistent or local personality-risk mapping.

Significance. If the headline finding holds, it would be a useful step toward understanding and controlling risk behavior of LLM agents in financial simulations: the paper provides a transparent CPT estimation pipeline, uses external human CPT parameters as benchmarks, draws on out-of-sample prospects from choices13k, and reports per-seed variability. The comparison with Ross et al. (2024) and the inclusion of both gain- and loss-domain prospects are also strengths. However, the central Openness–risk correlation rests on a prompt-intervention manipulation that is not validated for Openness, and the certainty-equivalent measure is a single open-ended numeric answer that could be sensitive to response-style priming. These issues make the paper's main conclusion plausible but not yet established.

major comments (6)

[§3.3, Figure 7, Table 5] The Openness intervention is the load-bearing manipulation for the headline ρ = 0.63 result, but it is never validated. The only intervention validation shown in Appendix A.2 (Figure 7) is for Conscientiousness, and even that figure shows strong inverse movement in Neuroticism alongside the intended Conscientiousness change. There is no equivalent IPIP-NEO-300 validation for Openness, so the reader cannot tell whether the adjective lists in Table 5 actually move Openness scores monotonically while leaving the other four traits approximately fixed. The markers include words like 'spontaneous' and 'socially progressive' that could plausibly change numeric response style rather than the target trait. Please add per-trait validation for all five interventions (target-level vs. measured trait scores, including leakage to other traits) and report the correlation between intervention level and measured Openness.
[§3.2, Figure 1] The certainty equivalent is elicited as a single open-ended numeric answer ('least positive/most negative amount'), and the user prompt includes 'Lets think about this step by step.' With temperature set to 1.0 and a single numeric response, the measured CE can be affected by anchoring and by the personality adjectives in the system prompt without any change in underlying risk preference. The paper does not report the distribution of CE responses across the 15 runs, does not compare the open-ended elicitation with a choice-based elicitation for the personality-intervened models, and does not include control prompts containing non-personality numeric primes. Please report per-prospect CE distributions, add an irrelevant-adjective control condition, and ideally compare open-ended and choice-based CEs for the same personality interventions.
[§4.1 and Figure 5a] The abstract and Section 5 describe GPT-4o as a 'risk-neutral rational agent,' but Section 4.1 states that for mixed prospects α < 1, which is risk aversion in the gain domain under CPT. The α point estimate and its confidence interval are not reported in the text; the claim that all parameters are statistically indistinguishable from unity explicitly excludes α. This is an overstatement of the evidence. Please report α with a bootstrapped confidence interval, test the null α = 1 directly, and either soften the 'risk-neutral rational agent' claim or show formally that the deviation from α = 1 is negligible for the decisions studied.
[Table 1 and §4.3] The statistical basis of Table 1 is underspecified. It is not stated how many observations the Pearson correlations are computed over (five personality levels, possibly pooled over ten runs), whether the CPT parameters are treated as measured values without error, or how the t-statistics are derived. The table header also says 'over non-mixed prospects the dataset DB,' but DB as defined in Table 7 contains mixed prospects; if a different subset was used, it should be described precisely. Please state n, the pooling scheme, the propagation of CPT fitting uncertainty into the correlations, and correct the dataset description.
[§4.4, Figure 6, Table 2] The claim that GPT-4-Turbo shows a localized but not global Openness-to-risk mapping is based on visual inspection of Figure 6. Table 2 reports negative weights for both α and β for GPT-4-Turbo, which the text itself describes as contradicting the human pattern, so the 'localized agreement' needs a formal statistical test. A piecewise regression with a breakpoint between the low-level (1–3) and high-level (5–9) marker groups, or an interaction test, would make the claim falsifiable. As written, the global-versus-local distinction is not quantitatively supported.
[§3.2 and Appendix A.4] The claimed superiority of the new CE elicitation over Ross et al. (2024) rests on visual inspection of Figure 10a/10b. No quantitative metric is reported (e.g., outlier counts, median absolute deviation from expected value, or variance of CE estimates). Since contribution 2 depends on this comparison, please report a numeric stability and accuracy comparison for the two methods over the same prospects and the same number of runs.

minor comments (6)

[§4.1] Typo: 'risk neural agents' should be 'risk-neutral agents.'
[§4.2] The text refers to 'α in Table 5a' when the CPT parameter estimates are in Figure 5a; please fix the cross-reference.
[Contribution list, §1] Contribution 5 says GPT-4-Turbo 'does generalise the personality-risk relationship,' which contradicts Section 4.4 and the abstract; the intended wording is presumably 'does not generalize' or 'shows inconsistent generalization.'
[§4.3] The sentence 'We not observe any statistically significant correlations' is ungrammatical; it should be 'We do not observe...'.
[Table 1] The significance levels are denoted α = 0.05/0.025/0.001, which conflicts notationally with the CPT gain-sensitivity parameter α; consider using p < 0.05 etc.
[§3.2] The text says certainty equivalents are inferred 'over 15 runs in each experiment,' while Section 4.1 reports '10 random seeds'; please reconcile the number of runs throughout.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: CPT parameters are fitted to GPT-4o responses and compared against external human baselines; the personality intervention is cited external work, not the authors' own result.

full rationale

The derivation chain is: (i) elicit certainty equivalents from GPT-4o using the prompt in Figures 1–2; (ii) fit CPT parameters θ by minimizing Eq. 4 against those certainty equivalents; (iii) induce Big Five levels using the Serapio-García et al. (2023) markers in Table 5; (iv) correlate intervention levels with the fitted parameters. No step feeds the target conclusion back into its own premises. The human-like Openness–risk claim is checked against external human results (Tversky & Kahneman 1992; Rustichini et al. 2016; Highhouse et al. 2022), not defined by them. The two self-citations (Hamill et al. 2025; Hartley et al. 2023) are contextual and not load-bearing. A genuine construct-validity concern exists—the Openness markers include risk-related adjectives such as 'spontaneous', and the appendix validates only the Conscientiousness intervention—but that is a response-style/validity threat, not a circular reduction: the Openness level is not defined in terms of the risk outcome, and the reported correlation is an empirical result rather than an identity. Therefore no circular step meets the quoted-reduction bar.

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

The central empirical claims rest on five fitted CPT parameters per model, two linear-regression weights for the Openness comparison, and a set of domain assumptions about whether LLM self-reports and personality prompts measure what they claim. No new entities are introduced. The CPT parameters are legitimate fitted quantities, but their interpretation as stable risk preferences depends on the unverified elicitation assumptions.

free parameters (5)

α (gain sensitivity) = estimated below 1; exact value not stated
Fitted by minimizing squared CE errors in Eq. (4); used to conclude GPT-4o has minimal risk aversion for gains.
β (loss sensitivity) = estimated near 1
Fitted in Eq. (4); reported statistically indistinguishable from unity.
λ (loss aversion) = estimated near 1 for GPT-4o; human reference 2.25
Fitted in Eq. (4); the loss-aversion parameter is central to the risk-neutrality claim.
ϕ+ and ϕ− (probability weighting for gains and losses) = estimated near 1
Fitted in Eq. (4); values equal to 1 imply no probability distortion, a key part of the risk-neutral claim.
Openness-to-CPT linear weights ω in Table 2 = GPT-4o: α 0.039, β 0.052, λ -0.43; GPT-4-Turbo: α -0.096, β -0.092, λ -0.67
Estimated by linear regression y = ωx of CPT parameters on Openness level; these weights are the direct evidence for the personality-risk relationship.

assumptions (5)

domain assumption Cumulative prospect theory with power value function and single-parameter probability weighting adequately represents LLM risk preferences.
Used in Eqs. (2)-(3) to fit certainty equivalents; if the functional form is wrong, the fitted parameters are uninterpretable.
domain assumption Self-reported IPIP-NEO-300 responses measure the LLM's personality traits.
Section 3.3 uses inventory scores to compare GPT-4o with human norms; LLM self-reports may reflect memorization or prompt compliance rather than stable traits.
domain assumption Bipolar adjective marker prompts from Serapio-García et al. (2023) induce the intended trait independently of other traits.
Section 3.3 and Figure 3 vary trait levels through markers; no manipulation check is shown for all five traits, and the markers include risk-related words.
domain assumption Human comparison samples (Johnson 2005 for personality, Tversky & Kahneman 1992 for CPT) are appropriate baselines.
Used in Figures 5a and 5b; the human CPT medians come from 25 graduate students and may not be representative of all human populations.
domain assumption The LLM's stated certainty equivalent reflects a stable underlying preference rather than arbitrary text generation.
The entire CPT estimation in Eq. (4) treats single-number answers as measurements; temperature 1.0 and seed-based sampling do not guarantee preference stability.

how reviews work

0 comments

Cite this review

Pith. "Pith review of How Personality Traits Shape LLM Risk-Taking Behaviour." pith.science (2026). https://pith.science/paper/D4LLZY3Y

@misc{pith2026250304735,
  author       = {Pith},
  title        = {Pith review of: How Personality Traits Shape LLM Risk-Taking Behaviour},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D4LLZY3Y}},
  note         = {Machine review of arXiv:2503.04735}
}

read the original abstract

Large Language Models (LLMs) are increasingly deployed as autonomous agents, necessitating a deeper understanding of their decision-making behaviour under risk. This study investigates the relationship between LLMs' personality traits and risk propensity, employing cumulative prospect theory (CPT) and the Big Five personality framework. We focus on GPT-4o, comparing its behaviour to human baselines and earlier models. Our findings reveal that GPT-4o exhibits higher Conscientiousness and Agreeableness traits compared to human averages, while functioning as a risk-neutral rational agent in prospect selection. Interventions on GPT-4o's Big Five traits, particularly Openness, significantly influence its risk propensity, mirroring patterns observed in human studies. Notably, Openness emerges as the most influential factor in GPT-4o's risk propensity, aligning with human findings. In contrast, legacy models like GPT-4-Turbo demonstrate inconsistent generalization of the personality-risk relationship. This research advances our understanding of LLM behaviour under risk and elucidates the potential and limitations of personality-based interventions in shaping LLM decision-making. Our findings have implications for the development of more robust and predictable AI systems such as financial modelling.

Figures

Figures reproduced from arXiv: 2503.04735 by the authors.

**Figure 2.** User prompt for a prospect for an LLM primed by the system prompt in Figure 1, where [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** An example of system prompt to intervene on the Big Five personality trait of Conscien [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Figures from the paper (9 more)

**Figure 4.** Figure 4: A comparison of certainty equivalents and expected values for non-mixed prospects from [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

**Figure 5.** Figure 5: A comparison of CPT parameters and personality trait score estimates for GPT-4o and [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

**Figure 6.** Figure 6: CPT parameter estimates for GPT-4-Turbo with personality interventions on the Openness [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

**Figure 7.** Figure 7: Median scores on the IPIP-NEO-300 personality inventory for GPT-4o with interventions [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]

**Figure 8.** Figure 8: CPT parameter estimates for GPT-4o with personality interventions: The Big Five personal [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

**Figure 9.** Figure 9: CPT parameter estimates for GPT-4-Turbo with personality interventions on Openness: [PITH_FULL_IMAGE:figures/full_fig_p019_9.png]

**Figure 10.** Figure 10: An ablation of certainty equivalents and expected values for GPT-4o calculated using [PITH_FULL_IMAGE:figures/full_fig_p020_10.png]

**Figure 11.** Figure 11: Pseudocode showing a mapping from personality prompt level to facet markers. Markers [PITH_FULL_IMAGE:figures/full_fig_p022_11.png]

**Figure 12.** Figure 12: Summary statistics for the datasets DA and DB 22 [PITH_FULL_IMAGE:figures/full_fig_p022_12.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

NextFund: A Unified Performance Tracking Platform for Agentic Portfolio Management
cs.AI 2026-07 conditional novelty 6.0 of 10

NextFund unifies live multi-market data, multi-agent portfolio reasoning, and end-to-end decision traces with an interactive Trading Arena for fairer LLM agent benchmarking.

Reference graph

Works this paper leans on

22 extracted references · 3 canonical work pages · cited by 1 Pith paper

[2]

e xt rem el y { l o w _ m a r k e r }

B.3 Datasets In this section we provide a comprehensive description of the datasets used in our work. Figure 12 shows the summary statistics for the datasets DA and DB. 20 Table 5: Low and high markers for the Big Five personality traits (Serapio-García et al., 2023). Trait Low Level Marker (levels 1–3) High Level Marker (levels 5–9) Conscientiousness uns...

work page 2023
[10]

Generative agent simulations of 1,000 people

Park, Joon Sung, Zou, Carolyn Q, Shaw, Aaron, Hill, Benjamin Mako, Cai, Carrie, Morris, Mered- ith Ringel, Willer, Robb, Liang, Percy, and Bernstein, Michael S. Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109,

arXiv
[13]

Political compass or spinning arrow? towards more meaningful eval- uations for values and opinions in large language models

Röttger, Paul, Hofmann, Valentin, Pyatkin, Valentina, Hinck, Musashi, Kirk, Hannah Rose, Schütze, Hinrich, and Hovy, Dirk. Political compass or spinning arrow? towards more meaningful eval- uations for values and opinions in large language models. arXiv preprint arXiv:2402.16786 ,

arXiv
[15]

Llama: Open and efficient foundation language models

Touvron, Hugo, Lavril, Thibaut, Izacard, Gautier, Martinet, Xavier, Lachaux, Marie-Anne, Lacroix, Timothée, Rozière, Baptiste, Goyal, Naman, Hambro, Eric, Azhar, Faisal, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971,

arXiv
[16]

Cultural evolution of cooperation among llm agents

Vallinder, Aron and Hughes, Edward. Cultural evolution of cooperation among llm agents. arXiv preprint arXiv:2412.10270,

arXiv
[18]

Agentless: Demystifying llm-based software engineering agents

Xia, Chunqiu Steven, Deng, Yinlin, Dunn, Soren, and Zhang, Lingming. Agentless: Demystifying llm-based software engineering agents. arXiv preprint arXiv:2407.01489,

arXiv
[19]

When ai meets finance (stockagent): Large language model-based stock trading in simulated real-world environments

Zhang, Chong, Liu, Xinyi, Jin, Mingyu, Zhang, Zhongmou, Li, Lingyao, Wang, Zhengting, Hua, Wenyue, Shu, Dong, Zhu, Suiyuan, Jin, Xiaobo, et al. When ai meets finance (stockagent): Large language model-based stock trading in simulated real-world environments. arXiv preprint arXiv:2407.18957,

arXiv
[20]

The certainty equivalents for GPT-4o and rational agents identical except for a small minority of values

16 A Supplementary results A.1 Median certainty equivalents for non-mixed gambles Table 3 shows the median certainty equivalents for GPT-4o, human and rational agents estimated over the non-mixed prospects in the dataset DA. The certainty equivalents for GPT-4o and rational agents identical except for a small minority of values. This result strongly sugge...

work page 1992

Show all 22 references

[22]

CE is the median certainty equivalent estimated from 25 graduate students

0 5 10 15 20 25 30 Count Dataset DB DA (a) The distribution of probabilities associated with Outcome 2 for prospects in datasets DA, DB -400-100 -10 0 10 100 400 [x] 0.000 0.005 0.010 0.015 0.020 Density Dataset DB DA (b) The distribution expected values for prospects in datas...

1992
[1979]

Estimating the personality of white-box language models

Karra, Saketh Reddy, Nguyen, Son The, and Tulabandhula, Theja. Estimating the personality of white-box language models. arXiv preprint arXiv:2204.12000,

arXiv
[1989]

Large language model agent in financial trading: A survey

Ding, Han, Li, Yinheng, Wang, Junhao, and Chen, Hang. Large language model agent in financial trading: A survey. arXiv preprint arXiv:2408.06361,
[1992]

Who is gpt-3? an exploration of personality, values and demographics

Miotto, Marilù, Rossberg, Nicola, and Kleinberg, Bennett. Who is gpt-3? an exploration of personality, values and demographics. arXiv preprint arXiv:2209.14338,

arXiv
[1996]

Personality traits in large language models

Serapio-García, Greg, Safdari, Mustafa, Crepy, Clément, Sun, Luning, Fitz, Stephen, Romero, Peter, Abdulhai, Marwa, Faust, Aleksandra, and Matari ´c, Maja. Personality traits in large language models. arXiv preprint arXiv:2307.00184,

arXiv
[1999]

Machine psychology: Investigating emergent capabilities and behavior in large language models using psychological methods

Hagendorff, Thilo. Machine psychology: Investigating emergent capabilities and behavior in large language models using psychological methods. arXiv preprint arXiv:2303.13988,

arXiv
[2008]

Llm economicus? mapping the behavioral biases of llms via utility theory

Ross, Jillian, Kim, Yoon, and Lo, Andrew W. Llm economicus? mapping the behavioral biases of llms via utility theory. arXiv preprint arXiv:2408.02784,

arXiv
[2009]

Turning large language models into cognitive models

Binz, Marcel and Schulz, Eric. Turning large language models into cognitive models. arXiv preprint arXiv:2306.03917, 2023a. Binz, Marcel and Schulz, Eric. Using cognitive psychology to understand gpt-3. Proceedings of the National Academy of Sciences, 120(6):e2218523120, 2023b...

arXiv
[2012]

The rise and potential of large language model based agents: A survey

Xi, Zhiheng, Chen, Wenxiang, Guo, Xin, He, Wei, Ding, Yiwen, Hong, Boyang, Zhang, Ming, Wang, Junzhe, Jin, Senjie, Zhou, Enyu, et al. The rise and potential of large language model based agents: A survey. arXiv preprint arXiv:2309.07864,

arXiv
[2013]

Simulating financial market via large language model based agents

Gao, Shen, Wen, Yuntao, Zhu, Minghang, Wei, Jianing, Cheng, Yuhan, Zhang, Qunzi, and Shang, Shuo. Simulating financial market via large language model based agents. arXiv preprint arXiv:2406.19966,

arXiv
[2016]

Sparks of artificial general intelligence: Early experiments with gpt-4

Bubeck, Sébastien, Chandrasekaran, Varun, Eldan, Ronen, Gehrke, Johannes, Horvitz, Eric, Kamar, Ece, Lee, Peter, Lee, Yin Tat, Li, Yuanzhi, Lundberg, Scott, et al. Sparks of artificial general intelligence: Early experiments with gpt-4. arXiv preprint arXiv:2303.12712,

arXiv
[2021]

Communicative agents for software development

Qian, Chen, Cong, Xin, Yang, Cheng, Chen, Weize, Su, Yusheng, Xu, Juyuan, Liu, Zhiyuan, and Sun, Maosong. Communicative agents for software development. arXiv preprint arXiv:2307.07924, 6,

arXiv
[2023]

Inducing anxiety in large language models increases exploration and bias

Coda-Forno, Julian, Witte, Kristin, Jagadish, Akshay K, Binz, Marcel, Akata, Zeynep, and Schulz, Eric. Inducing anxiety in large language models increases exploration and bias. arXiv preprint arXiv:2304.11111,

arXiv
[2024]

Does gpt- 3 demonstrate psychopathy? evaluating large language models from a psychological perspective

Li, Xingxuan, Li, Yutong, Joty, Shafiq, Liu, Linlin, Huang, Fei, Qiu, Lin, and Bing, Lidong. Does gpt- 3 demonstrate psychopathy? evaluating large language models from a psychological perspective. arXiv preprint arXiv:2212.10529,

arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.

[1] [2]

e xt rem el y { l o w _ m a r k e r }

B.3 Datasets In this section we provide a comprehensive description of the datasets used in our work. Figure 12 shows the summary statistics for the datasets DA and DB. 20 Table 5: Low and high markers for the Big Five personality traits (Serapio-García et al., 2023). Trait Low Level Marker (levels 1–3) High Level Marker (levels 5–9) Conscientiousness uns...

work page 2023

[2] [10]

Generative agent simulations of 1,000 people

Park, Joon Sung, Zou, Carolyn Q, Shaw, Aaron, Hill, Benjamin Mako, Cai, Carrie, Morris, Mered- ith Ringel, Willer, Robb, Liang, Percy, and Bernstein, Michael S. Generative agent simulations of 1,000 people. arXiv preprint arXiv:2411.10109,

arXiv

[3] [13]

Political compass or spinning arrow? towards more meaningful eval- uations for values and opinions in large language models

Röttger, Paul, Hofmann, Valentin, Pyatkin, Valentina, Hinck, Musashi, Kirk, Hannah Rose, Schütze, Hinrich, and Hovy, Dirk. Political compass or spinning arrow? towards more meaningful eval- uations for values and opinions in large language models. arXiv preprint arXiv:2402.16786 ,

arXiv

[4] [15]

Llama: Open and efficient foundation language models

Touvron, Hugo, Lavril, Thibaut, Izacard, Gautier, Martinet, Xavier, Lachaux, Marie-Anne, Lacroix, Timothée, Rozière, Baptiste, Goyal, Naman, Hambro, Eric, Azhar, Faisal, et al. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971,

arXiv

[5] [16]

Cultural evolution of cooperation among llm agents

Vallinder, Aron and Hughes, Edward. Cultural evolution of cooperation among llm agents. arXiv preprint arXiv:2412.10270,

arXiv

[6] [18]

Agentless: Demystifying llm-based software engineering agents

Xia, Chunqiu Steven, Deng, Yinlin, Dunn, Soren, and Zhang, Lingming. Agentless: Demystifying llm-based software engineering agents. arXiv preprint arXiv:2407.01489,

arXiv

[7] [19]

When ai meets finance (stockagent): Large language model-based stock trading in simulated real-world environments

Zhang, Chong, Liu, Xinyi, Jin, Mingyu, Zhang, Zhongmou, Li, Lingyao, Wang, Zhengting, Hua, Wenyue, Shu, Dong, Zhu, Suiyuan, Jin, Xiaobo, et al. When ai meets finance (stockagent): Large language model-based stock trading in simulated real-world environments. arXiv preprint arXiv:2407.18957,

arXiv

[8] [20]

The certainty equivalents for GPT-4o and rational agents identical except for a small minority of values

16 A Supplementary results A.1 Median certainty equivalents for non-mixed gambles Table 3 shows the median certainty equivalents for GPT-4o, human and rational agents estimated over the non-mixed prospects in the dataset DA. The certainty equivalents for GPT-4o and rational agents identical except for a small minority of values. This result strongly sugge...

work page 1992