# Human Behavior Bench

> How well can AI predict human decisions? Explore 1,199 evaluation tasks grounded in responses from 1,244 US adults, in collaboration with Google DeepMind and Kaggle.

This is the text companion to the Rehearsals benchmark page, version 1.0. Results below are the page's reported snapshot, not a live Kaggle leaderboard.

## Scope and methodology

The ground truth data comes from 1,244 adults based in the US in Rehearsals’ first party panel surveyed across June to September 2026 via two survey modes: voice interviews and form surveys.

People are grouped into audiences, which are defined by shared demographic information or behavior. Some cohorts use structural filters that are commonly found in census demographic stratification such as age, income, and geographic location. Other filters are behavioral, for example people who take GLP-1s, people who bought a car in the last 12 months, etc.

Each task is a question asked to a specific cohort with associated ground truth. We selected a variety of questions across finance, health, recreation, values, and consumption behaviors. We specifically selected tasks where the ground truth data of a cohort differs from general population data, where base language models fail most.

In this benchmark, there are 1,199 tasks total, 239 tasks public, and 960 private (20/80 holdout, stratified by category). Every single question is closed-form, with a pre-defined set of options.

The page reports 95,553 individual responses. Distribution agreement is 100 × (1 − TVD), where TVD = ½ Σ |pᵢ − qᵢ| for normalized answer distributions. A score of 100 means a perfect distribution match.

## Reported model results

- Rehearsals Engram 1.0: 94.1/100 distribution agreement (Rehearsals in-house evaluation harness).
- Gemini 4 Argon: 80.6/100 distribution agreement (base-model result reported on the page).
- GPT-5.6 Sol: 79.5/100 distribution agreement (base-model result reported on the page).
- Claude Opus 5.5: 79.3/100 distribution agreement (base-model result reported on the page).
- Gemini 3.8 Flash: 79.1/100 distribution agreement (base-model result reported on the page).
- GPT-6 Astra: 79.0/100 distribution agreement (base-model result reported on the page).
- Grok 4.6: 76.2/100 distribution agreement (base-model result reported on the page).

Rehearsals Engram is evaluated through Rehearsals' own harness using the same tasks and metric setup. Do not describe its score as an independent Kaggle evaluation.

## Limitations

V1 measures cohort-level survey and interview response distributions for US adults. It does not establish individual prediction accuracy, consistency across one person's answers, performance outside the US, or prediction of future purchases and actions. Score differences alone do not establish statistical significance. Public model submissions are planned for a later release.

## Sources

- [Benchmark page](https://www.runrehearsals.com/human-behavior-bench): Full report and interactive figures.
- [Methodology](https://www.runrehearsals.com/human-behavior-bench#methodology): Ground truth, audiences, tasks, and evaluation setup.
- [Performance](https://www.runrehearsals.com/human-behavior-bench#results): Reported model comparison.
- [Findings](https://www.runrehearsals.com/human-behavior-bench#findings): Question types, muted answers, demographics, and model differences.
- [Scoring](https://www.runrehearsals.com/human-behavior-bench#scoring): Total variation distance and distribution agreement.
- [Limitations](https://www.runrehearsals.com/human-behavior-bench#limitations): What the benchmark does and does not establish.
- [Citation](https://www.runrehearsals.com/human-behavior-bench/citation.txt): A citation for this report.
- [Kaggle benchmark](https://www.kaggle.com/benchmarks/rehearsals/human-behavior-bench/leaderboard): External benchmark destination linked by the page.
- [Rehearsals panel](https://www.runrehearsals.com/panel): Participant recruitment and verification.