Three Prompt Traits Match Economic Game Behavior

Economic experiments often treat preferences, beliefs, and strategic reasoning as separate explanations for behavior, yet it remains difficult to represent one person consistently across games with different rules and action spaces. The authors propose using an LLM as a behavioral simulator whose prompt contains a small, interpretable type vector. Rather than train a model on individual decisions, they search over verbal trait profiles and ask which profiles reproduce observed choice distributions. Their central result is that three prompted characteristics—Risk Aversion, Strategic Sophistication, and Trust—closely match aggregate behavior in a large cross-game dataset.

Core Contribution

The paper turns prompt wording into a model-selection device. A type vector assigns Likert-scale values to behavioral characteristics, such as “Risk Aversion: 4 out of 5,” then conditions an LLM to play a game role. The authors vary both the candidate characteristics and their levels, comparing generated choices with human choices after normalizing each game’s action range.

This differs from fitting a conventional predictive model to observed features. The vector is intended to be portable and legible: the same profile can be used across distinct games, and its dimensions retain behavioral labels rather than becoming latent numerical factors. The paper therefore asks a structural question, not merely whether an ML model can minimize in-sample error: how many named dimensions are needed before an AI-generated distribution resembles human play?

Technical Approach

The evaluation covers 119,147 decisions by 78,657 subjects from more than 35 countries, spanning ten classic economic game roles. For each candidate keyword set, the authors generate choices from type vectors at five Likert levels per dimension. They then fit mixtures over the resulting prompt-conditioned behavior to minimize normalized Wasserstein distance from the human action distribution. A default system prompt without an assigned type provides a direct reference point.

The search begins with one-dimensional types and adds dimensions up to five. Figure 2 isolates how each behavioral keyword shifts choices in each role, which matters because a useful prompt dimension must induce variation rather than merely supply an intuitive label. For each dimensionality, the authors identify the keyword combination with the lowest mean distance across all ten roles for the aggregate comparison.

Figure 3 makes the modeling target concrete. It places the human distributions above LLM distributions under the default prompt and under fitted mixtures with one through five dimensions; Nash and total-payoff-maximizing actions are marked where applicable. The comparison is distributional, so a match is not just agreement on an average action. It tests whether the prompted mixture captures dispersion and multimodality across games.

The paper also fits individuals who played at least five roles, seeking a best-matching vector for each person. It examines whether those vectors predict choices in a held-out game and compares their mean absolute error, scaled by the game’s action range, with a HistGBT model. Separate checks compare the five economic keywords with Big Five OCEAN traits and placebo keywords, while a cross-model analysis repeats keyword effects for five LLMs.

Results and Analysis

Adding prompt dimensions improves aggregate distributional fit, and the three-dimensional Risk Aversion–Strategic Sophistication–Trust combination closely matches aggregate behavior across the ten roles. The result is stronger than showing that an LLM can generate plausible game play: the default prompt is explicitly evaluated, and the fitted mixtures are judged against the full human distributions using repeated 1,000-draw comparisons. Figure 4 reports distances by dimensionality and game, alongside a human-sampling benchmark, making clear that the gains are evaluated role by role rather than inferred from a single pooled average.

The individual analysis supplies the more consequential test. The authors report that the type vectors required to fit cross-game individuals cluster into fewer than a dozen groups, and that type-based profiles can predict behavior in games held out from fitting despite changed rules and action sets. Figure 7 benchmarks those held-out errors against HistGBT, a more conventional supervised predictor. That comparison is the right stress test for the paper’s interpretability claim: a compact verbal representation is useful only if its portability does not come at a prohibitive predictive cost.

The evidence supports a narrower conclusion than a general theory of human behavior. Within these ten economic roles, a three-trait prompt family appears sufficient to approximate aggregate distributions and offers a compact basis for individual matching. The study does not establish that the traits are stable psychological measurements, nor that they will carry to non-economic decisions. Its contribution is a practical empirical representation: prompt vectors can expose low-dimensional regularities that would be difficult to inspect in raw decision data alone.

Limits in Practice

The method depends on the behavior induced by a particular LLM, system prompt, trait vocabulary, and sampling procedure. Figure 9 checks directional keyword effects across five models, but cross-model agreement in correlations is not the same as identical fitted distributions or identical individual-level predictions. The search also evaluates a bounded menu of five economic characteristics and five Likert levels; low dimensionality within that menu need not imply that other behavioral domains are intrinsically three-dimensional. Finally, the held-out validation remains within a family of classic economic games, where choices are scalar and incentives are stylized.