The Innate Economic Preferences of Language Models

Joshua Foster; Joy Buchanan

REVIEW 4 major objections 5 minor 74 references

The Innate Economic Preferences of Language Models

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A language model's single-token choice rule is a random utility model, so its raw logits reveal a measurable economic preference, and fine-tuning can rewrite that preference.

desk verdict The softmax-as-RUM framing is right and the logit-as-utility-index toolkit is genuinely useful; but the structural β headline outruns the evidence — the paper's own IIA and position-bias results reject the identifying assumption, leaving risk-aversion as a robust qualitative feature, not a stable structural parameter. read the letter →

arxiv 2607.26288 v1 pith:VCXKF76P submitted 2026-07-28 econ.EM

Joy Buchanan , Joshua Foster This is my paper

classification econ.EM MSC 91B0691B1662P20

keywords largelanguagemodelsrevealedpreferencerandomutilitymodelriskaversionmultinomiallogitindependenceofirrelevantalternativesstructuraldiscretechoicefine-tuning

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 tries to establish that a language model's single-token choice rule is not just text generation but exactly the random utility model economists use for human choice. Because the model's raw logits are observed, they serve as a systematic utility index, so the model's default economic preferences can be read directly rather than inferred from behavior. Estimating risk attitudes on portfolio menus across twelve models, the paper finds every model risk-averse, with estimates varying by a factor of nearly four across developers. It also finds the models honor dominance and mostly pass transitivity, but violate independence of irrelevant alternatives and label-invariance, so preference strength is unstable near indifference. Finally, fine-tuning an open-weight model toward two target risk coefficients shows a principal can install and audit a chosen risk attitude.

What carries the argument

The load-bearing object is the softmax-to-logit identity: choice probability equals exp(u_i/τ) over the sum of exp(u_j/τ), which is precisely the conditional logit representation with i.i.d. Gumbel shocks, so a logit gap between two menu labels is a scaled utility difference. The paper couples this with the affine utility specification u_i = κV_i + ζ and the quadratic benchmark V_i = μ_i − βR_i, so identification reduces to regressing the A-minus-B logit gap on Δμ and ΔR; the risk coefficient is recovered as the ratio β = −θ2/θ1. To control position bias, each menu is mirrored, and half the difference between canonical and mirrored gaps removes the additive position premium before estimation

What would settle it

Estimate β twice for the same open-weight model, once on the D-optimal menu sample and once on a random sample from the same admissible grid, and compare the two ratios −θ2/θ1 with their joint confidence interval. If they differ beyond sampling error, or if the quadratic surface's fit collapses on the random sample, the structural quadratic-utility interpretation is menu-dependent and the 'innate preference' claim fails.

Watch

Extended reading notes

Core claim

The central claim is an equivalence: under single-token forced choice, the model's softmax decoding rule is exactly the multinomial logit random utility model, with the logit vector as the observed systematic utility index and temperature as the scale of the Gumbel noise. The paper then imposes an affine mapping from utility to logits and a quadratic mean-variance benchmark V_i = μ_i − βR_i, reducing the model's risk preference to the single parameter β. Across a subject pool of twelve open-weight and proprietary models, every estimate of β is positive and statistically significant, so all models are risk-averse; the estimates range from about 0.0031 to 0.0114, enough to change portfolio ran

Load-bearing premise

The collapse point is the maintained assumption that one quadratic utility V_i = μ_i − βR_i with a constant scale κ governs logits across all menus and prompts; the paper's own diagnostics (IIA as low as 0.12, state-dependent position bias, completeness as low as 0.758 for one 70B model) show this is only approximate, so the single β is a design-weighted average rather than a fully stable structural parameter.

Editorial extensions

If this is right

A principal can audit what preference a model brings to an underspecified instruction by reading logits directly, without the inference assumptions needed for human data.
Because all twelve models are risk-averse with β spanning roughly 0.003 to 0.011, the same portfolio menu can produce different allocations depending on which developer's model is asked.
Monotonicity and continuity at ceiling imply models respect mean-risk dominance, so the economic failure mode is concentrated in label- and menu-composition effects near indifference.
The fine-tuning result implies a mismatched default risk attitude is not fixed: a principal can install a target β and verify it out of sample with the same structural estimator.
The six diagnostics form a deployable rationality test battery for AI agents before they are entrusted with real resource-allocation decisions.

Reading between the lines

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

If the logit-as-utility identity holds beyond portfolio menus, the same read-out could certify the preferences of any single-token agent before market deployment—a step the paper does not itself take.
The paper's robustness checks show position bias is state-dependent and strongest near indifference, so the single recovered β is best read as a design-weighted average; re-estimating β on near-indifference menus only would test how structural it really is.
The isomorphism is deliberately single-token; extending it to multi-token deliberation would require showing that a chain of reasoning composes as a random utility process, which remains open.
Quadratic mean-variance utility is one maintained specification; the same logit-gap machinery could compare nested candidates such as constant-relative-risk-aversion or probability-weighted utility using non-nested tests.

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

The real contribution here is the move to read raw LLM logits as an observed systematic utility index. Proposition 2.1 is just McFadden, but the application is clean and practically useful: it gives a way to audit the default preferences an LLM agent brings to a delegated decision. The two-regime estimation (logit regression for open-weight, ML for frontier) is careful, the six-axiom battery is a sensible reusable protocol, and the empirical sweep is wide — twelve models, prompt variants, D-optimal vs random menus, CoT checks. The paper also reports its weaknesses honestly: Table 14 explicitly calls the position effect a design-weighted average, and Appendix C.4 discloses the low-β κ problem instead of burying it.

What is oversold is the word "structural" applied to the headline β. The identifying assumption — Eq. (2), constant κ, additive position bias — is rejected by the paper's own diagnostics. IIA indices run as low as 0.122 (Gemma 4), and Figure 5 shows the position bias is state-dependent, not constant. The paper treats these as findings about the models, but they equally undercut the estimator that produces β. Prompt dependence is severe: Llama 3.1 8B β swings from 0.0091 to 0.00015 across prompts, a 60-fold change; Qwen 3 8B's position bias flips sign between D-optimal and random menus. The sign of β is stable everywhere, so "all models are risk-averse in this single-token task" survives. Magnitude and structural interpretation do not. The fine-tuning exercise, meanwhile, puts the target β* in the loss and recovers it with the same estimator — a legitimate engineering demonstration that logit surfaces can be shaped, but not independent confirmation that the base-model βs are structural. Minor items: Llama 3.1 70B completeness is 0.758, so a quarter of its mass is off-menu; footnote 15 admits an fp32 robustness run was abandoned, leaving an open question; and the replication code is promised but not in the manuscript. For a methods paper, shipping that code should gate publication.

Who this is for: anyone working on LLM agents in markets, alignment, or machine-subject behavioral economics. It deserves a serious referee, but the referee should push to reword the structural claims as descriptive features of logit surfaces and to release the code. The measurement toolkit will be cited either way.

Referee Report

4 major / 5 minor

Summary. The paper proposes that a language model's single-token generation rule—softmax over logits with temperature—is exactly the multinomial logit random utility model (Proposition 2.1), so raw logits can be read as an observed systematic utility index. Under a maintained quadratic mean-variance utility V_i = μ_i − β R_i and an affine logit mapping u_i = κ V_i + ζ with constant κ across menus, the authors estimate β and κ for twelve models, using observed logits for open-weight models and sampled choices (MLE) for frontier models. They also construct six revealed-preference diagnostics (completeness, reflexivity, monotonicity, transitivity, continuity, IIA). The headline results are: monotonicity and continuity are at or near ceiling, transitivity is high, but reflexivity and IIA are frequently violated; every model shows a positive, statistically significant β, i.e., risk aversion; and fine-tuning Llama 3.1 8B with a loss on the logit gap recovers two preset β targets on held-out menus.

Significance. If the structural interpretation holds, this is a valuable measurement program: it connects LLM internals to a well-established discrete-choice framework, gives a direct reading of preference parameters from logits rather than inferring them from choices, provides a battery of rationality diagnostics, and demonstrates an explicit fine-tuning route to implant a risk attitude. The proof of Proposition 2.1 is correct, the estimators are standard, and the paper is unusually transparent: label-token maps, replication schedules, design details, and code are documented. The sign robustness of β across prompts and menus is a potentially useful finding. However, the paper's own diagnostics undercut the stronger claim that logits 'structurally identify' a stable preference parameter, so the central contribution needs substantial reframing.

major comments (4)

[§3.2, Eq. (6); §5.1, Table 3] The identifying equation (6) assumes a constant utility-to-logit scale κ across menus and a purely additive position bias that mirroring removes. The paper's own diagnostics reject this: IIA indices range from 0.122 to 0.920, reflexivity is 0.000 for some models, and completeness is as low as 0.758 (Llama 3.1 70B). Table 14 and Figure 5 further show that the position bias is state-dependent—for Qwen 3 (8B) the estimated α flips sign between D-optimal and random designs (+0.305 vs −0.987)—and Table 15 shows the position-determined choice rate rises from 0.000 on dominance menus to 0.420 on iso-utility menus. Under these conditions, the estimator in Eq. (6) is fitting a misspecified surface, and the recovered β is a design-weighted average over that surface, not a stable structural parameter. The paper should either weaken the 'structural identification' claim to a conditional approximatio
[§5.2, Table 12; Appendix D.4] The prompt-robustness results in Table 12 show that β is highly design-sensitive. For Llama 3.1 (8B), β moves from 0.0091 under the baseline prompt to 0.00015 under the Compact Table prompt—a roughly 60-fold change—and Ministral 3 (8B) moves from 0.0114 to 0.0043 across prompts. Chain-of-thought variants in Table 17 also change β materially for several models. Finding 2 ('all models exhibit risk aversion') is supported only at the level of the sign of β, not as a quantitatively stable preference parameter. Since the abstract and introduction emphasize structural identification and heterogeneity in risk attitudes, the main text should prominently qualify that the reported magnitudes are conditional on the exact prompt and menu design, and the cross-model ranking may not be robust.
[§5.3, Eq. (10); §6] The fine-tuning exercise is described as a 'structural validation' of the preference interpretation, but Eq. (10) constructs the target logit gap directly from the same maintained quadratic utility and a chosen κ, and the same estimator (Eq. (6)) then recovers β* from the fine-tuned model. The recovered β matching the target therefore confirms internal consistency of the training objective, not that the base-model β estimates correspond to an underlying structural preference. This circularity is acknowledged indirectly in Appendix C, but the main text's conclusion overstates the evidential value. I recommend re-framing Section 5.3 as a demonstration that a logit-surface target can be installed and audited, not as independent validation of the structural model.
[§5.1, Table 3; §5.2, Table 4] The maintained quadratic utility is presented as 'a useful approximation,' but the fit is poor for several frontier models: pseudo-R² is 0.634 for Claude 4.5 Haiku and 0.425 for Claude 4.6 Sonnet. Given the paper's goal of structural identification, the choice of V_i = μ_i − β R_i should be tested against alternatives (e.g., allowing a separate coefficient on σ_i or a more flexible mean-variance tradeoff), or at least the limited fit should be discussed as evidence that the quadratic benchmark is only a local approximation for some models. As it stands, the structural estimates for low-R² models are difficult to interpret as 'innate economic preferences.'

minor comments (5)

[§2.2] Typo: 'transitivty' should be 'transitivity'.
[References] Reference to 'V on Neumann and Morgenstern' should read 'von Neumann and Morgenstern.' Also check spacing in the reference list (e.g., 'Harrison, Glenn W, John A List').
[§4.3] The greedy forward D-optimal algorithm is described; Appendix B discusses the benchmark but not how the greedy implementation was validated. A brief note on the number of candidate menus and the achieved determinant would help.
[Table 3] Column headers 'Complete' and 'Reflexive' are abbreviations; spell out 'Completeness' and 'Reflexivity' for clarity, and define the temperature at which indices are evaluated.
[§5.2, Figure 3] The heatmap color scale and the yellow dashed line are described, but the figure caption could state the base bundle explicitly in the caption for each panel.

Circularity Check

1 steps flagged · score 4.0 of 10

Base-model β is a self-contained regression summary of logits; circularity appears in the fine-tuning 'validation,' which recovers its own training target with the same estimator.

fitted input called prediction [Section 5.3, Eq. (10); Appendix C.2 Eq. (19); results in Table 6]
"For each training menu, we compute the utility difference implied by the principal’s target preference and translate it into the log odds that the delegated model should assign to the two options. ∆z∗m ≡ log P∗(A|m) P∗(B|m) = κ(V∗A − V∗B) = κ(∆µm − β∗∆Rm). (10) ... The fine-tuning exercise provides a structural validation of this interpretation."

The training target in Eq. (10) is exactly the same linear logit-gap model κ(Δμ − β*ΔR) that the estimator inverts in Eq. (6) and Eq. (14) via β̂ = −θ̂2/θ̂1. The loss L(θ) = (1/M)Σ(Δz_m(θ) − Δz*_m)^2 trains the model to match that target, so recovering β* on held-out menus primarily verifies that the model learned its own training objective and that the estimator is the inverse of the loss. It is a manipulation check with a generalization component, not an independent validation of the base-model 'innate' preferences or of quadratic utility; calling it 'structural validation' is therefore by construction conditional on the inserted target.

full rationale

The central measurement of β is not circular: Proposition 2.1 is the textbook softmax/multinomial-logit equivalence (McFadden 1972), and Eq. (6) is a direct least-squares projection of observed logit gaps onto (Δμ, ΔR). Estimating β that way does not presuppose the answer; the paper's own diagnostics (IIA 0.122–0.920, state-dependent position bias in Table 14/Figure 5, prompt sensitivity in Table 12) show the maintained constant-κ quadratic specification is unstable, but instability is a misspecification/validity concern, not circularity. No load-bearing self-citation or imported uniqueness theorem is present. The one genuinely by-construction element is the fine-tuning demonstration: the target β* is inserted into the training loss and recovered by the same estimator, so calling that an independent 'structural validation' overstates what is shown. That is a secondary result, however, so the overall circularity score is moderate rather than high.

Assumptions & free parameters 5 free parameters · 7 assumptions · 0 invented entities

No new physical entities are introduced; the Gumbel shocks are the standard RUM error structure. The paper's load-bearing content is carried by the maintained quadratic utility, the affine logit scale, and the additive position-bias assumption, together with the hand-set fine-tuning hyperparameters.

free parameters (5)

β (risk aversion) per model = 0.0031 (Qwen3-8B) to 0.0114 (Ministral-3-8B)
Recovered as -θ2/θ1 from the observed-logit or sampled-choice logit regressions (Table 4); the paper's central claim of universal risk aversion is a summary of these fitted values.
κ (utility-to-logit scale) per model = 0.039 (Llama-3.1-8B) to 1.482 (Gemma-4-31B)
Estimated as the return slope θ1; assumed constant across menus within a model (Eqs. 2, 5).
α (position bias) per model = e.g., +1.383 (Qwen3-14B D-optimal), -0.969 (Qwen3-8B random grid)
Intercept in the logit-gap regression; shown to be design-dependent (Tables 13/14), so the 'position correction' in Eq. (9) is itself estimated and imperfect.
Fine-tuning loss weights (λ_mass, λ_sob, λ_ref, λ_mir, λ_iia, λ_disp, λ_exact, λ_3m, λ_cyc, λ_path, λ_path,smooth) = (0.10,0.10,0.10,0.20,0.05,0.05,0.15,0.05,0.10,0.15,0.03)
Hand-set in Appendix C.2; the fine-tuning result depends on these tuning constants, plus δ=1, ω_sgn=0.25, m_sgn=0.25, m_br=0.05, T_aux, epochs, and corpus sizes.
κ_ft (logit scale for fine-tuning targets) = κ=1 for low-β; base-model gap std for high-β
Selected by a κ∈{1,2,4} sweep on diagnostics (Table 9); β estimates are invariant to κ in the ratio, but diagnostic scores are not.

assumptions (7)

domain assumption The model's one-token generation follows softmax over the full vocabulary at fixed temperature τ (Eq. 1).
The paper restricts to single-token forced choices with τ=1, top-p=1; this is the empirical protocol, not a law of LLMs.
standard math Gumbel i.i.d. errors give the exact RUM representation of softmax (Prop. 2.1).
McFadden 1972; proof in Appendix A.
domain assumption Logits are an affine transform of a cardinal utility index u_i = κV_i + ζ, with κ constant across prompt contexts (Eq. 2).
Needed for cross-menu comparability of utility gaps; a maintained normalization.
domain assumption The benchmark utility is quadratic mean-variance, V_i = μ_i - βR_i with R_i=σ_i²+μ_i² (Eq. 4).
This functional form is imposed, not derived; β is only identified relative to it. The paper cites Markowitz (2014) in support.
domain assumption Position bias enters additively and is removed by mirroring (Eq. 9).
Mirroring assumes the position premium is independent of the utility difference; Tables 14/Figure 5 show it is state-dependent, so the correction is approximate.
domain assumption For frontier models, sampled choices are conditionally i.i.d. and the Bernoulli likelihood (Eq. 8) is correctly specified.
Standard MLE regularity; requires valid-menu responses only.
ad hoc to paper The fine-tuning target is defined by the same quadratic utility and a prespecified κ (Eq. 10).
The training objective is constructed from the maintained model, so the fine-tuning result is by construction.

how reviews work

0 comments

Cite this review

Pith. "Pith review of The Innate Economic Preferences of Language Models." pith.science (2026). https://pith.science/paper/VCXKF76P

@misc{pith2026260726288,
  author       = {Pith},
  title        = {Pith review of: The Innate Economic Preferences of Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VCXKF76P}},
  note         = {Machine review of arXiv:2607.26288}
}

read the original abstract

Language models increasingly settle real resource tradeoffs on behalf of principals yet their economic preferences remain unobserved. We demonstrate their generation rule is isomorphic to the random utility model of discrete choice. This allows internal logit scores to structurally identify preferences. Estimating risk attitudes across twelve models in a portfolio task reveals universal but heterogeneous risk aversion. Although models reject strictly dominated options, their elicited preferences fail invariance tests and violate the independence of irrelevant alternatives across varying experimental prompts. Finally, fine tuning establishes that a principal can explicitly engineer a target risk attitude.

Figures

Figures reproduced from arXiv: 2607.26288 by the authors.

**Figure 1.** Stylized visual representations of the random utility restrictions evaluated in the single-token forced-choice [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗

**Figure 2.** Decision prompt template The two observation regimes differ not only in what is recorded but in how choice probabilities are formed. In the open-weight regime, one forward pass per menu yields the exact label logits, and no replication is required. In the frontier regime, choice probabilities are formed either from repeated sampled choices or, where the provider exposes them, from first-token logprobs. Across both r… view at source ↗

**Figure 3.** Empirical indifference curves for open-weight models mapped over the portfolio space. [PITH_FULL_IMAGE:figures/full_fig_p018_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Empirical indifference curves for fine-tuned models mapped over the portfolio space. [PITH_FULL_IMAGE:figures/full_fig_p020_4.png]

**Figure 5.** Figure 5: State-dependent position bias over the feasible mean–variance grid. Each cell reports the mean paired bias [PITH_FULL_IMAGE:figures/full_fig_p034_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

74 extracted references · 5 linked inside Pith

[1]

Afriat, Sydney N. 1967. The construction of utility functions from expenditure data. International Economic Review 8 (1):67--77

1967
[2]

Akata, Elif, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. 2025. Playing Repeated Games with Large Language Models. Nature Human Behaviour 9 (7):1380--1390

2025
[3]

Andersen, Steffen, Glenn W Harrison, Morten I Lau, and E Elisabet Rutstr \"o m. 2008. Eliciting risk and time preferences. Econometrica 76 (3):583--618

2008
[4]

Apesteguia, Jose and Miguel A Ballester. 2018. Monotone stochastic choice models: The case of risk and time preferences. Journal of Political Economy 126 (1):74--106

2018
[5]

and Harald Uhlig

Araujo, Douglas K.G. and Harald Uhlig. 2026. How does AI Distribute the pie? Large Language Models and the Ultimatum Game. Working Paper 34919, National Bureau of Economic Research

2026
[6]

Berry, Steven T. 1994. Estimating discrete-choice models of product differentiation. The RAND Journal of Economics :242--262

1994
[7]

Bini, Pietro, Lin William Cong, Xing Huang, and Lawrence J. Jin. 2026. Behavioral Economics of AI : LLM Biases and Corrections. Working Paper 34745, National Bureau of Economic Research. ://www.nber.org/papers/w34745

2026
[8]

Brand, James, Ayelet Israeli, and Donald Ngwe. 2023. Using LLMs for market research. Harvard Business School Marketing Unit Working Paper (23-062)

2023

Show all 74 references

[9]

Bridle, John. 1989. Training stochastic model recognition algorithms as networks can lead to maximum mutual information estimation of parameters. Advances in Neural Information Processing Systems 2

1989
[10]

Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell et al. 2020. Language models are few-shot learners. Advances in Neural Information Processing Systems 33:1877--1901

2020
[11]

Bruhin, Adrian, Helga Fehr-Duda, and Thomas Epper. 2010. Risk and rationality: Uncovering heterogeneity in probability distortion. Econometrica 78 (4):1375--1412

2010
[12]

Charness, Gary, Uri Gneezy, and Alex Imas. 2013. Experimental methods: Eliciting risk preferences. Journal of Economic Behavior & Organization 87:43--51

2013
[13]

Choi, Syngjoo, Raymond Fisman, Douglas Gale, and Shachar Kariv. 2007. Consistency and heterogeneity of individual behavior under uncertainty. American Economic Review 97 (5):1921--1938

2007
[14]

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

2017
[15]

Chugunova, Marina and Daniela Sele. 2022. We and it: an interdisciplinary review of the experimental evidence on how humans interact with machines. Journal of Behavioral and Experimental Economics 99:101897

2022
[16]

Cohen, Alma and Liran Einav. 2007. Estimating risk preferences from deductible choice. American Economic Review 97 (3):745--788

2007
[17]

de Silva, Tim, Taha Choukhmane, Weidong Lin, and Matthew Akuzawa. 2026. AI Financial Advice: Supply, Demand, and Life Cycle Implications. Working paper, Working Paper

2026
[18]

Demuynck, Thomas and John Rehbeck. 2023. Computing revealed preference goodness-of-fit measures with integer programming. Economic Theory 76 (4):1175--1195

2023
[19]

Dietvorst, Berkeley J, Joseph P Simmons, and Cade Massey. 2015. Algorithm aversion: people erroneously avoid algorithms after seeing them err. Journal of Experimental Psychology: General 144 (1):114

2015
[20]

Einav, Liran, Amy Finkelstein, Iuliana Pascu, and Mark R Cullen. 2012. How general are risk preferences? Choices under uncertainty in different domains. American Economic Review 102 (6):2606--2638

2012
[21]

Eisfeldt, Andrea L and Gregor Schubert. 2024. AI and Finance. Tech. rep., National Bureau of Economic Research

2024
[22]

Ellis, Keaton and Wanying Huang. 2026. Should I State or Should I Show? Aligning AI with Human Preferences. arXiv preprint arXiv:2603.29317 ://arxiv.org/abs/2603.29317

2026 arXiv
[23]

Greene, William H. 2003. Econometric analysis. Pretence Hall

2003
[24]

Guan, Bryan, Tanya Roosta, Peyman Passban, and Mehdi Rezagholizadeh. 2025. The order effect: investigating prompt sensitivity to input order in LLMs. arXiv preprint arXiv:2502.04134

2025 arXiv
[25]

Gul, Faruk. 1991. A theory of disappointment aversion. Econometrica :667--686

1991
[26]

Harrison, Glenn W, John A List, and Charles Towe. 2007. Naturally occurring preferences and exogenous laboratory experiments: A case study of risk aversion. Econometrica 75 (2):433--458

2007
[27]

Harrison, Glenn W. and E. Elisabet Rutstr \"o m. 2008. Risk Aversion in the Laboratory. In Research in Experimental Economics, vol. 12. Emerald Group Publishing Limited, 41--196

2008
[28]

Hausman, Jerry and Daniel McFadden. 1984. Specification tests for the multinomial logit model. Econometrica :1219--1240

1984
[29]

Holt, Charles A and Susan K Laury. 2002. Risk aversion and incentive effects. American Economic Review 92 (5):1644--1655

2002
[30]

Holtzman, Ari, Jan Buys, Li Du, Maxwell Forbes, and Yejin Choi. 2019. The curious case of neural text degeneration. arXiv preprint arXiv:1904.09751

2019 arXiv
[31]

Horton, John J, Apostolos Filippas, and Benjamin S Manning. 2023. Large language models as simulated economic agents: What can we learn from homo silicus? Tech. rep., National Bureau of Economic Research

2023
[32]

Hu, Edward J, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Liang Wang, Weizhu Chen et al. 2022. Lora: Low-rank adaptation of large language models. Iclr 1 (2):3

2022
[33]

Imas, Alex, Kevin Lee, and Sanjog Misra. 2025. Agentic Interactions. Working paper. ://ssrn.com/abstract=5875162. Available at SSRN 5875162

2025
[34]

Kahneman, Daniel. 2003. Maps of Bounded Rationality: Psychology for Behavioral Economics. American Economic Review 93 (5):1449--1475

2003
[35]

Kahneman, Daniel and Amos Tversky. 1979. Prospect Theory: An Analysis of Decision under Risk. Econometrica 47 (2):263--291

1979
[36]

Kiefer, Jack and Jacob Wolfowitz. 1959. Optimum designs in regression problems. The Annals of Mathematical Statistics 30 (2):271--294

1959
[37]

Kim, Jeongbin, Matthew Kovach, Kyu-Min Lee, Euncheol Shin, and Hector Tzavellas. 2024. Can an LLM Learn Preferences from Choice Data? arXiv preprint arXiv:2401.07345

2024 arXiv
[38]

and Charles F

List, John A. and Charles F. Mason. 2011. Are CEO s Expected Utility Maximizers? Journal of Econometrics 162 (1):114--123

2011
[39]

Liu, Jiaxin, Yixuan Tang, Yi Yang, and Kar Yan Tam. 2025. Evaluating and aligning human economic risk preferences in llms. In Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 18185--18199

2025
[40]

Logg, Jennifer M, Julia A Minson, and Don A Moore. 2019. Algorithm appreciation: People prefer algorithmic to human judgment. Organizational Behavior and Human Decision Processes 151:90--103

2019
[41]

Louviere, Jordan J, Terry N Flynn, and Richard T Carson. 2010. Discrete choice experiments are not conjoint analysis. Journal of Choice Modelling 3 (3):57--72

2010
[42]

Louviere, Jordan J and George Woodworth. 1983. Design and analysis of simulated consumer choice or allocation experiments: an approach based on aggregate data. Journal of Marketing Research 20 (4):350--367

1983
[43]

Lu, Wei, Daniel L Chen, and Christian B Hansen. 2025. Aligning Large Language Model Agents with Rational and Moral Preferences: A Supervised Fine-Tuning Approach. arXiv preprint arXiv:2507.20796

2025
[44]

Luce, R Duncan et al. 1959. Individual choice behavior, vol. 4. Wiley New York

1959
[45]

Ludwig, Jens, Sendhil Mullainathan, and Ashesh Rambachan. 2025. Large Language Models: An Applied Econometric Framework. Working Paper 33344, National Bureau of Economic Research

2025
[46]

Manski, Charles F. 1975. Maximum score estimation of the stochastic utility model of choice. Journal of Econometrics 3 (3):205--228

1975
[47]

--- --- ---. 1977. The structure of random utility models. Theory and Decision 8 (3):229

1977
[48]

Markowitz, Harry. 2014. Mean--variance approximations to expected utility. European Journal of Operational Research 234 (2):346--355

2014
[49]

Marschak, Jacob. 1960. Binary choice constraints and random utility indicators. In Mathematical Methods in the Social Sciences, edited by Kenneth J Arrow, Samuel Karlin, and Patrick Suppes. Stanford, CA: Stanford University Press, 312--329

1960
[50]

Mat e jka, Filip and Alisdair McKay. 2015. Rational inattention to discrete choices: A new foundation for the multinomial logit model. American Economic Review 105 (1):272--298

2015
[51]

McFadden, Daniel. 1972. Conditional logit analysis of qualitative choice behavior

1972
[52]

Mei, Qiaozhu, Yutong Xie, Walter Yuan, and Matthew O. Jackson. 2024. A Turing test of whether AI chatbots are behaviorally similar to humans. Proceedings of the National Academy of Sciences 121 (9):e2313925121

2024
[53]

Mester, Loretta J. 1996. A study of bank efficiency taking into account risk-preferences. Journal of Banking & Finance 20 (6):1025--1045

1996
[54]

D-optimal

Mitchell, Toby J. 2000. An algorithm for the construction of “D-optimal” experimental designs. Technometrics 42 (1):48--54

2000
[55]

Newey, Whitney K and Daniel McFadden. 1994. Large sample estimation and hypothesis testing. Handbook of Econometrics 4:2111--2245

1994
[56]

Ouyang, Long, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray et al. 2022. Training language models to follow instructions with human feedback. Advances in Neural Information Processing Systems 35...

2022
[57]

Ouyang, Shumiao, Hayong Yun, and Xingjian Zheng. 2024. How ethical should ai be? how ai alignment shapes the risk preferences of llms. arXiv preprint arXiv:2406.01168

2024 arXiv
[58]

Park, Joon Sung, Joseph O'Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. 2023. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th Annual ACM Symposium on User Interface Software and Technology. 1--22

2023
[59]

Pezeshkpour, Pouya and Estevam Hruschka. 2024. Large language models sensitivity to the order of options in multiple-choice questions. In Findings of the Association for Computational Linguistics: NAACL 2024. 2006--2017

2024
[60]

Prelec, Dra z en. 1998. The Probability Weighting Function. Econometrica 66 (3):497--527

1998
[61]

Rafailov, Rafael, Archit Sharma, Eric Mitchell, Christopher D Manning, Stefano Ermon, and Chelsea Finn. 2023. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems 36:53728--53741

2023
[62]

Rose, John M and Michiel CJ Bliemer. 2009. Constructing efficient stated choice experimental designs. Transport Reviews 29 (5):587--617

2009
[63]

Salant, Yuval and Ariel Rubinstein. 2008. (A, f): choice with frames. The Review of Economic Studies 75 (4):1287--1296

2008
[64]

Schildberg-H \"o risch, Hannah. 2018. Are risk preferences stable? Journal of Economic Perspectives 32 (2):135--154

2018
[65]

Serapio-Garc \' a, Gregory, Mustafa Safdari, Cl \'e ment Crepy, Luning Sun, Stephen Fitz, Peter Romero, Marwa Abdulhai, Aleksandra Faust, and Maja Matari \'c . 2025. A psychometric framework for evaluating and shaping personality traits in large language models. Nature Machine...

2025
[66]

Manning, Andrey Fradkin, and John J

Shahidi, Peyman, Gili Rusak, Benjamin S. Manning, Andrey Fradkin, and John J. Horton. 2025. The Coasean Singularity? Demand, Supply, and Market Design with AI Agents. In The Economics of Transformative AI, chap. 6. University of Chicago Press. ://www.nber.org/papers/w34468. NB...

2025
[67]

and Richard F

Stanovich, Keith E. and Richard F. West. 2000. Individual differences in reasoning: Implications for the rationality debate? Behavioral and Brain Sciences 23 (5):645–665

2000
[68]

Thurstone, Louis L. 1927. A law of comparative judgment. Psychological Review 34 (4):273--286

1927
[69]

Train, Kenneth E. 2009. Discrete choice methods with simulation. Cambridge university press

2009
[70]

Varian, Hal R. 1992. Microeconomic analysis, vol. 3. Norton New York

1992
[71]

Vaswani, Ashish, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, ukasz Kaiser, and Illia Polosukhin. 2017. Attention is all you need. Advances in Neural Information Processing Systems 30

2017
[72]

Von Neumann, John and Oskar Morgenstern. 1947. Theory of games and economic behavior, 2nd rev

1947
[73]

Wei, Jason, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou et al. 2022. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35:24824--24837

2022
[74]

Zheng, Chujie, Hao Zhou, Fandong Meng, Jie Zhou, and Minlie Huang. 2024. Large language models are not robust multiple choice selectors. In International Conference on Learning Representations, vol. 2024. 19426--19454

2024

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.

[1] [1]

Afriat, Sydney N. 1967. The construction of utility functions from expenditure data. International Economic Review 8 (1):67--77

1967

[2] [2]

Akata, Elif, Lion Schulz, Julian Coda-Forno, Seong Joon Oh, Matthias Bethge, and Eric Schulz. 2025. Playing Repeated Games with Large Language Models. Nature Human Behaviour 9 (7):1380--1390

2025

[3] [3]

Andersen, Steffen, Glenn W Harrison, Morten I Lau, and E Elisabet Rutstr \"o m. 2008. Eliciting risk and time preferences. Econometrica 76 (3):583--618

2008

[4] [4]

Apesteguia, Jose and Miguel A Ballester. 2018. Monotone stochastic choice models: The case of risk and time preferences. Journal of Political Economy 126 (1):74--106

2018

[5] [5]

and Harald Uhlig

Araujo, Douglas K.G. and Harald Uhlig. 2026. How does AI Distribute the pie? Large Language Models and the Ultimatum Game. Working Paper 34919, National Bureau of Economic Research

2026

[6] [6]

Berry, Steven T. 1994. Estimating discrete-choice models of product differentiation. The RAND Journal of Economics :242--262

1994

[7] [7]

Bini, Pietro, Lin William Cong, Xing Huang, and Lawrence J. Jin. 2026. Behavioral Economics of AI : LLM Biases and Corrections. Working Paper 34745, National Bureau of Economic Research. ://www.nber.org/papers/w34745

2026

[8] [8]

Brand, James, Ayelet Israeli, and Donald Ngwe. 2023. Using LLMs for market research. Harvard Business School Marketing Unit Working Paper (23-062)

2023