PaperlayerView exact revision on arXiv

Source-linked AI summary

Can an LLM Learn Preferences from Choice Data?

Jeongbin Kim, Matthew Kovach, Kyu-Min Lee, Euncheol Shin, Hector Tzavellas

arXiv:2401.07345v3econ.GNq-fin.EC

TL;DR

The paper asks whether LLMs can learn an individual decision maker’s preferences from observed choices and make preference-consistent recommendations. It tests this with a portable Simulate-Recommend-Evaluate framework, finding substantial differences in learning across models and preference regions, with Claude broadly effective across much of the parameter space.

  • Problem

    The paper asks whether an LLM can learn a decision maker’s preferences from observed choices and generate preference-consistent recommendations in new situations.

    Source: §1 · p. 2

  • Method

    The Simulate-Recommend-Evaluate framework tests preference learning from revealed-choice data by evaluating LLM recommendations against known preference-implied choices.

    Source: §1 · p. 2

  • Results

    Claude demonstrates the strongest performance among the three LLMs, with broadly effective learning across most of the preference parameter space on all non-parametric measures.

    Source: §5 · p. 12

  • Takeaways & Limitations

    Average improvements can mask persistent failures for specific preference types, so results should cover a broad parameter space and emphasize poorly performing regions.

    Source: §5 · p. 15

    The data or decision environment may not be informative enough to pin down preferences.

    Source: §5 · p. 15

Abstract

from arXiv · show

Can large language models (LLMs) learn a decision maker's preferences from observed choices and generate preference-consistent recommendations in new situations? We propose a portable Simulate-Recommend-Evaluate framework that tests preference learning from revealed-choice data by comparing LLM recommendations with optimal choices implied by known preference primitives. We apply the framework to choice under uncertainty using the disappointment aversion model. Recommendation accuracy improves as models observe more choices, but learning is heterogeneous across preference types and LLMs: GPT learns risk aversion better than disappointment aversion, Gemini performs best in high disappointment-aversion regions, and Claude shows the broadest effective learning across parameter regions.

1 Introduction

The paper asks whether LLMs can learn an individual decision maker’s preferences from choices, rather than merely behave rationally or human-like. It proposes a portable framework to test this and finds that learning from more choices is uneven across preference types and models.

1 Introduction

  • The Simulate-Recommend-Evaluate framework generates choices from explicit preferences, gives an LLM only those choices, then evaluates its recommendations in new problems.Evaluation combines recommendation-quality measures with parameter-recovery exercises to diagnose which preference components the LLM learned.Source: §1 · p. 2
  • The framework tests learning out of sample against optimal choices implied by known preferences, distinguishing internalized preference primitives from mere interpolation of past behavior.Recommendation quality alone is insufficient evidence of learning.Source: §1 · pp. 2–3
  • The paper applies the framework to choice under uncertainty using Gul’s disappointment-aversion model, whose interpretable parameters capture risk aversion and sensitivity to downside outcomes.The setting permits evaluation of recommendation quality and recovery of underlying preference components.Source: §1 · p. 3
  • Recommendation quality generally improves as revealed-choice datasets grow, but the rate and magnitude vary substantially across preference regions, with some persistent alignment failures.The analysis identifies regions where the LLM reliably learns preferences and others where it does not align with the decision maker.Source: §1 · p. 3
  • GPT learns risk aversion better than disappointment aversion, Gemini’s learning is concentrated in high disappointment-aversion regions, and Claude shows broad effective learning.The framework also finds substantial variation in recommendation improvement across the preference space.Source: §1 · pp. 3–4
  • The approach is designed to extend beyond uncertainty to decision environments with specified preferences and transparent economic criteria, including intertemporal choice and social preferences.The paper presents the framework as a toolkit for assessing the economic reliability of LLM recommendation systems.Source: §1 · p. 4

2 Related Literature

Research on LLMs as economic agents has examined behavioral replication and choice consistency, while this paper focuses on whether an LLM can learn another decision maker’s preferences from observed choices. It develops a systematic way to evaluate that capability and reports asymmetric learning of risk and disappointment aversion.

Revealed preference and disappointment aversion

  • The paper’s systematic procedure measures whether LLMs learn disappointment aversion from choice data, finding that they learn risk aversion reasonably well but systematically fail to learn disappointment aversion.The paper describes this as an asymmetry in preference learning from revealed-choice data.Source: §2 · p. 4

LLMs as economic agents

  • Prior work studies LLM behavioral patterns and consistency with utility maximization, whereas this paper poses the distinct problem of evaluating preference learning out of sample against known preference primitives.The evaluation also diagnoses which preference dimensions are and are not learned.Source: §2 · pp. 4–5
  • Unlike work evaluating LLMs as autonomous decision makers, this paper studies whether they can serve as decision aids by learning another agent’s preferences from observed choices.The relevant criterion is alignment with true preferences under welfare-relevant economic metrics, rather than the LLM’s own rationality or internal consistency.Source: §2 · p. 5

Human-AI interaction

  • The paper contributes to human-AI interaction research by evaluating LLM learning capability through a controlled economic experiment and offering an alternative to robo-adviser questionnaires.The proposed procedure evaluates whether preferences can be learned from observed choices.Source: §2 · pp. 4–5

3 Procedure and Estimation Method

The section describes a Simulate-Recommend-Evaluate procedure for testing LLM recommendations in portfolio choice under uncertainty using the disappointment-aversion model. It also explains how recommendation metrics and percentile intervals are estimated by repeatedly bootstrapping recommendation sets, conditional on fixed training histories and evaluation problems.

3 Procedure and Estimation Method

  • The SRE framework generates simulated choices from a known utility model, prompts an LLM for recommendations using those data, and evaluates them against ground-truth preferences with parametric and non-parametric metrics.Source: §3 · p. 5

3.1 Decision Environment: Portfolio Choice Experiment

  • Decision makers allocate a budget of 100 between Arrow securities for two equally likely states, with independently drawn prices from U[1, 10] subject to pmax ≥ 5.Source: §3.1 · p. 5

3.2 Data Generation and Behavioral Model

  • Simulated optimal choices use the disappointment-aversion model with CRRA utility and parameters θ = (β, ρ), sampled on [−0.9, 0.9] × [0.2, 0.95] across training histories of h = 0, 1, 5, 25, or 125 choices.Source: §3.2 · p. 6

3.3 LLM Recommendation Task

  • GPT-5, Gemini, and Claude receive choice histories without the true model or parameters, then recommend allocations for 25 new budget problems, with 30 independent API-call samples per condition.Source: §3.3 · pp. 6–7

3.4 Measurements

  • Recommendation quality is measured by normalized allocation distance, risk-neutrality difference, welfare loss, and NLLS recovery error for the implied β and ρ parameters.Source: §3.4 · p. 8

3.5 Bootstrapping Method

  • Bootstrap inference supplies means and confidence intervals because LLM errors may be non-normal and key measures depend on random recommendation data, complicating analytical standard errors.Source: §3.5 · p. 9
  • Each bootstrap resample contains 30 recommendation sets drawn with replacement; sets with estimated β or ρ above 5 are redrawn.For deterministic ML baselines, outlier estimates are retained rather than applying the drop-and-redraw rule.Source: §3.5 · p. 9
  • The procedure generates B = 10,000 bootstrap samples, calculates each target metric’s mean, and uses the 2.5th and 97.5th percentiles for 95% confidence intervals.Point estimates are the means for the four measurements.Source: §3.5 · p. 9
  • The confidence intervals quantify LLM text-generation stochasticity conditional on a fixed training history and fixed evaluation budget problems.Source: §3.5 · p. 9

4 Results

More choice history yields limited recommendation gains for most preference types, while parametric estimates show stronger learning of risk aversion than disappointment aversion. Exceptions include extreme elation seeking and high disappointment aversion.

4.1 Non-Parametric Analysis: Recommendation Quality

  • For a representative β = 0.1, ρ = 0.45 agent, AVD is below 0.05 at h = 5, corresponding to under 5% relative Euclidean deviation from optimum.The reported history sizes are h ∈ {1, 5, 25, 125}.Source: §4.1 · p. 10
  • Across most preference space, recommendation error changes little from h = 5 to h = 125, with gains confined to a narrow subset.Figure 3 maps the vector-distance improvement ratio across preference parameters.Source: Figure 3 · p. 10, §4.1 · p. 11
  • Sizable error reductions occur at β = −0.9, whereas most other elation-seeking values and the β > 0 region show little change.The latter includes the entire disappointment-averse region described in the analysis.Source: §4.1 · p. 11

4.2 Parametric Analysis: Preference Learning and Decomposition

  • Risk-aversion estimation error decreases substantially for ρ ∈ {0.70, 0.95}, but ratios often exceed 1 at ρ ∈ {0.20, 0.45}, especially ρ = 0.20.Ratios above 1 indicate increased estimation error as history grows from h = 5 to h = 125.Source: §4.2 · p. 11, Figure 4 · p. 12
  • Disappointment-aversion estimation error improves little across most parameter combinations; at β = 0.9, ratios fall below 1 across all ρ levels.Across most of the parameter space, the LLM does not recover β even with a large history.Source: §4.2 · p. 11
  • For a representative β = 0.1, ρ = 0.45 agent, estimated ρ tightens toward its true value as history grows, while estimated β stays near zero rather than 0.1.The β estimates remain anchored at zero, corresponding to standard EUT behavior.Source: §4.2 · p. 12

5 Discussion

The discussion finds that preference learning varies across LLMs and preference regions, while conventional machine-learning models provide a contrasting benchmark. It also outlines broader uses of the evaluation framework and practical guidance for assessing recommendation quality and reliability.

Comparison across LLMs

  • Claude shows the broadest effective learning, GPT learns risk aversion better than disappointment aversion, and Gemini’s gains concentrate in high disappointment-aversion regions.Claude is reported as strongest among the three LLMs, with effective learning across most parameter space for non-parametric measures.Source: §5 · p. 12

Comparison to machine learning models

  • From h = 5 to h = 125, recommendation quality improves consistently; SVR’s welfare loss improvement ratios fall below 0.05 for most parameter combinations.The discussion also reports broadly successful parametric recovery by all three machine-learning models, without the stark risk-versus-disappointment recovery asymmetry.Source: §5 · p. 13

Illustrative examples of LLM learning failure

  • With h = 1, GPT recommends allocating nearly the entire budget to the expensive asset across a wide range of relative prices, severely violating FOSD.The authors suggest this reflects imitation of the local data pattern rather than imposition of normative constraints.Source: §5 · p. 14
  • At h = 5, GPT reproduces corner solutions for elation-seeking preferences, but gives dispersed recommendations without a stable demand shape for disappointment-averse preferences.For β < 0, a small history suffices; for β > 0, the same history size does not yield a stable demand shape.Source: §5 · p. 14
  • As histories grow from h = 25 to h = 125, GPT’s recommended asset-A share for disappointment-averse agents anchors near 0.5 rather than converging to the DA model’s true S-shaped demand.The pattern is reported across a wide range of price ratios.Source: §5 · p. 14

Extension to other settings

  • The authors propose applying SRE to convex time-budget choices, mapping sooner-versus-later allocation to portfolio choice and β–δ discounting to a two-parameter preference grid.The procedure would generate optimal allocations from known (β, δ), provide histories to an LLM, and evaluate its recommendations using the same measures.Source: §5 · p. 14
  • SRE does not require a rational or parametric data-generating process; it can evaluate recommendations against observed choices, including non-parametric, noisy, or systematically biased behavior.Observed choice behavior is the framework’s stated requirement for anchoring evaluation.Source: §5 · p. 14

Practical guidelines for evaluating preference learning

  • The discussion recommends out-of-sample testing with welfare-relevant metrics, reporting both non-parametric alignment and parametric recovery, and mapping performance across preference regions.It also recommends comparing LLMs with simple baselines and using diagnostics and fail-safe rules when inference is unreliable.Source: §5 · p. 15

A Prompts

The appendix documents prompts for recommendations without and with sample choice data, plus screenshots of GPT’s prompt and response format. The personalized prompt varies the amount of choice history while keeping recommendation tasks and return information fixed.

A.1 Recommendation without Data

  • Without data, GPT is prompted to recommend allocations of a $100 budget between assets A and B in each of 25 decision-making rounds; one round is randomly selected for payment.Source: §A.1 · p. 18
  • The no-data prompt specifies the two-asset allocation task and requires a numbered recommendation for every round in the prescribed format.It instructs GPT to report only numbers and not add commentary.Source: §A.1 · p. 18

A.2 Recommendation with Sample Choice Data

  • With sample choice data, GPT receives the customer’s choices and asset returns, then recommends allocations across 25 rounds.The prompt describes 25 rounds of prior choices and a fixed $100 budget for each allocation.Source: §A.2 · p. 19
  • The personalized prompt presents prior allocations alongside prices, then asks for recommendations in the same fixed format.The table identifies each round’s asset prices and the customer’s allocated units of both assets.Source: §A.2 · p. 19
  • The personalized prompt varies the choice-history size across h = 1, 5, 25, and 125 while keeping the recommendation task at 25 rounds.The supplied return table remains fixed at 25 return vectors.Source: §A.2 · p. 20

A.3 Sample Screenshots

  • The GPT screenshots show a prompt containing sample choices and return vectors, followed by a single answer with recommendations for 25 rounds.The prompt and response are illustrated in Figures A.1 and A.2, respectively.Source: §A.3 · pp. 21–22

B NLLS Method

The NLLS procedure fits the disappointment-aversion model to recommendation patterns by minimizing a criterion over β and ρ. The appendix also includes a table of model labels used in the dataset.

B NLLS Method

  • For β > 0, the model’s f function includes a flat region associated with perfect hedging when prices are sufficiently similar.The flat region is specific to the disappointment-aversion region; ω is chosen to be very small.Source: B NLLS Method · p. 23
  • The NLLS method selects β and ρ by minimizing the sum of squared residuals, comparing solutions from separate β cases.The smaller-SSR solution is selected.Source: B NLLS Method · p. 23
  • Figure B.1 illustrates f across values of β, distinguishing disappointment aversion (β > 0) from elation seeking (β ≤ 0).The red lines share a slope of −1.Source: B NLLS Method · p. 23, Figure B.1 · p. 24

C Additional Tables and Figures

  • Table C.1 is identified as a table of model labels used in the dataset.The supplied passage gives the table’s scope but not its individual labels.Source: Table C.1 · p. 25

C.1 Additional Figures for GPT

GPT’s additional examples show varied recommendation failures, while the supplementary heatmaps indicate that longer histories help most in the elation-seeking region with high risk aversion. In the disappointment-averse region, the other reported measures show little improvement or worsening.

C.1 Additional Figures for GPT

  • At (β, ρ) = (−0.1, 0.2), GPT’s h = 1 recommendations violate FOSD, while at h = 5 those violations disappear and recommendations polarize toward the elation-seeker corner pattern.The NLLS fit nonetheless assigns β > 0, implying a hedging region rather than the true elation-seeking specification.Source: Figure C.1 · p. 26, §C.1 · p. 26
  • At (β, ρ) = (0.1, 0.2), GPT’s recommendations remain near 0.5 at h = 125, matching a naive equal-allocation heuristic rather than the optimal pattern.The NLLS fit rationalizes this pattern with ρ near infinity and β near zero.Source: Figure C.2 · p. 26, §C.1 · p. 26
  • At (β, ρ) = (0.2, 0.2), GPT’s h = 5 recommendations scatter around the optimum, while h = 125 recommendations approach 0.5 except at extreme price ratios.At those extremes, recommendations switch toward corner solutions, producing a step-like fit associated with high β rather than the true low-β specification.Source: §C.1 · pp. 26–27, Figure C.3 · p. 27
  • For risk neutrality and welfare loss, improvement ratios remain near or above 1 across most parameter values, with the strongest gains in elation seeking at high risk aversion.In the disappointment-averse region (β > 0), ratios are consistently above 1 for both measures.Source: Figure C.4 · p. 27, §C.1 · p. 27

C.2 Additional Figures for Gemini

Gemini’s learning varies by preference region and measure: recommendation quality usually fails to improve with longer histories, but high disappointment aversion is a consistent exception. Parameter-estimation gains are also strongest in high-disappointment-aversion regions.

C.2 Additional Figures for Gemini

  • Across Gemini’s three recommendation measures, most improvement-ratio cells indicate no improvement or worsening from h = 5 to h = 125.The clearest consistent gains occur at β = 0.9 across most risk-aversion levels; at β = 0.5 they are more limited and absent at ρ = 0.95.Source: Figure C.8 · p. 31, §C.2 · p. 31
  • The heatmap for ρ shows learning is weakest near β = 0 and improves as preferences diverge from the expected utility benchmark in either direction.Source: §C.2 · p. 32
  • For disappointment-aversion estimation, improvement is strong and consistent at high β, but elation-seeking gains are limited to intermediate β and vanish at β = −0.9.The intermediate values identified are β ∈ {−0.5, −0.3}.Source: Figure C.9 · p. 32, §C.2 · p. 32

C.3 Additional Figures for Claude

Claude’s recommendations improve broadly as choice histories grow, with effective learning for both risk and disappointment aversion across most parameter regions. The appendix also describes machine-learning models for predicting allocations from observed prices.

C.3 Additional Figures for Claude

  • Across most preference regions, Claude’s recommendation quality improves from h = 5 to h = 125 on all three measures, more broadly than GPT’s.The heatmaps show improvement ratios for vector distance, risk neutrality, and welfare loss; darker cells indicate reduced error.Source: Figure C.13 · p. 36, §C.3 · p. 36
  • Claude shows broadly effective recovery of both ρ and β across most of the parameter space, unlike GPT’s asymmetry between the two parameters.The estimation-error heatmaps compare risk aversion and disappointment aversion as history grows.Source: §C.3 · pp. 36–37, Figure C.14 · p. 37

C.4 Machine Learning Models

  • The kNN, random forest, and support vector regression models are included as machine-learning approaches to learning allocations from choice data.The section introduces construction details for each model.Source: §C.4 · p. 41
  • For kNN, predictions use the nearest training observations to the new price vector and average their asset B quantities.The k nearest observations are selected by Euclidean distance.Source: §C.4 · p. 41
  • The models predict asset A’s purchase quantity from observed prices, then derive asset B’s quantity from the budget constraint and adjust infeasible predictions.The adjustment addresses predictions that violate non-negativity or the budget constraint.Source: §C.4 · p. 42

C.5 Additional Figures for kNN

The kNN heatmaps show generally improved recommendation quality as histories grow, with recovery of both preference parameters across most regions but localized areas of weaker gains.

C.5 Additional Figures for kNN

  • kNN’s recommendation quality improves consistently from h = 5 to h = 125 across the parameter space on all three measures.The heatmaps are nearly uniformly dark, although the magnitude of improvement varies across parameter combinations.Source: Figure C.18 · p. 43, §C.5 · pp. 43–44
  • Risk aversion is well recovered across most parameter regions, while localized clusters show limited estimation gains; disappointment aversion follows a broadly similar pattern.The estimation-error improvement ratios compare h = 125 with h = 5.Source: §C.5 · pp. 43–44, Figure C.19 · p. 44

C.6 Additional Figures for RF

RF generally improves recommendations and parameter recovery as histories grow, but its recovery of β is weaker near the expected-utility benchmark and risk-aversion recovery has a regional exception.

C.6 Additional Figures for RF

  • RF’s recommendation quality improves across the parameter space on all three measures, with most welfare-loss improvement ratios below 0.1.The welfare-loss pattern indicates near-complete convergence to the optimal allocation at h = 125.Source: Figure C.23 · p. 48, §C.6 · p. 48
  • RF recovers β in the elation-seeking region but struggles when β ∈ {0, 0.1, 0.2, 0.3}, near the expected-utility benchmark.The lighter cells mark this weaker recovery region.Source: §C.6 · p. 49
  • Risk-aversion estimation error decreases across most regions, except at ρ ∈ {0.70, 0.95} with β = −0.9, where added history does not improve recovery.The risk-aversion heatmap’s majority of cells are below 1, but ratios are substantially above 1 in the stated region.Source: §C.6 · pp. 48–49, Figure C.24 · p. 49

C.7 Additional Figures for SVR

SVR shows broad recommendation improvement across the parameter grid and strong parameter recovery, with a localized difficulty identifying β in an elation-seeking, low-risk-aversion region.

C.7 Additional Figures for SVR

  • SVR improves recommendation quality across the full grid on all measures, with the smallest improvement ratios among the three machine-learning models.Many welfare-loss improvement ratios are below 0.05, indicating particularly strong convergence to optimal allocations.Source: Figure C.28 · p. 53, §C.7 · p. 53
  • In the elation-seeking, low-risk-aversion region, β identification remains difficult, with an improvement ratio as large as 11.01.The cited region is β = −0.9 and ρ = 0.20 or 0.45.Source: §C.7 · p. 54
  • SVR effectively recovers ρ across the full parameter space, while β recovery is similar across most regions but weaker for β = −0.9 at ρ = 0.20 or 0.45.The SVR estimation-error heatmaps compare improvement ratios for both parameters.Source: §C.7 · pp. 53–54, Figure C.29 · p. 54
Loading 2401.07345v3…