Small behavioral profiles improve AI simulations—but the gains depend on the game

Choosing information, not writing richer biographies

Language models are increasingly used to simulate responses to rules or incentives before researchers test them with people. The difficult question is whether those simulations remain useful when the conditions change. A model that matches behavior under familiar rules may fail under a different payment schedule or bargaining structure.

In their October 2026 preprint, Robin Na and colleagues treat the description supplied to a simulator as something to test, rather than assume useful. Their central question is which features of earlier behavior improve predictions for different participants facing new experimental designs.

From records to profiles

The researchers first defined candidate features using behavioral literature and author-reviewed scoring rules. One language model examined experimental records, including actions, messages and, where available, surveys. It assigned scores from zero to ten only when the record offered qualifying evidence. Unsupported features were omitted rather than given a neutral score. Each included feature carried supporting evidence and the model’s stated certainty.

A separate simulation model received selected portions of these profiles alongside the rules of a game. Across comparisons, the researchers held the sampled profiles and simulation settings fixed while changing which features were visible.

They tested features individually, then added features one at a time when doing so improved the specified learning comparisons. This search selected two-feature profiles in each primary application. Validation then used different participants and experimental designs absent from learning, without revising the selected contents.

The experiments covered three settings: repeated public-goods games involving contributions to a shared account; three-person bargaining over chips with private values; and generalized money-request games in which numerical choices earned different payoffs and bonuses. Validation included 250 public-goods games, 48 bargaining games and 884 money-request designs.

What helped the simulations

The leading feature differed by setting. In public goods, it described withdrawal after adverse peer behavior, such as reducing contributions after others failed to contribute. In bargaining, it described package breadth: a tendency to negotiate larger bundles of chips. In money-request games, it described sensitivity to payoffs and incentives.

The authors compared simulated and human outcome distributions, not simply their averages. Their normalized alignment scale assigns zero to an uninformed reference and 100 to the estimated agreement between two resampled human datasets of the same size. These scores are not percentages of correct predictions.

The leading one-feature profiles scored 65.5 for public goods, 77.5 for bargaining and 73.5 for money requests. Without profiles, the corresponding scores were −1.0, 40.5 and 12.2. Paired bootstrap comparisons favored the leading profiles over no profile and each of three general-purpose persona libraries.

More detail did not reliably help. Adding the second selected feature produced no clear validation improvement in any family. The one-feature profile clearly outperformed the complete profile in bargaining; differences between those conditions remained inconclusive in the other two settings. An inconclusive comparison does not establish equal performance.

Better distributions are not better individual forecasts

The money-request results carry an important qualification. The main evaluation pooled choices across different payoff designs sharing a bonus rule. Profiles improved the match between those pooled distributions.

However, secondary evaluations comparing each simulated request with human choices under its particular payoff design favored simulations without profiles. The reported aggregate gain therefore did not extend to those design-specific action comparisons.

Nor did the study test whether a profile predicts the same person across situations. Learning and validation involved different participants. In money-request games, almost everyone made only one choice, so profiles represented weighted response types rather than established individual personalities.

Transfer has boundaries

Profiles also traveled unevenly between game families. Public-goods and bargaining profiles improved the pooled money-request simulations. In the reverse direction, the leading money-request profile substantially worsened public-goods predictions, scoring −68.0—below both the uninformed reference and the no-profile condition.

The authors consequently argue for evaluating behavioral descriptions in the setting where they will be used. Their findings do not establish stable psychological traits or causal mechanisms. Most features lacked independent benchmark measurements or human-coded ratings, and changing the generation or simulation model changed which features were selected.

Other limits remain: the search did not examine every feature combination; the study did not evaluate demographic-group differences; and general-purpose personas described different populations, preventing a clean separation of information effects from population differences. The practical contribution is a testable way to choose simulation inputs, not evidence that synthetic participants can broadly replace human experiments.