Synthetic Users Benchmark - GSS Individual Accuracy: leaderboard
LLMs as synthetic survey respondents (2026): given the demographics of a real General Social Survey respondent, the model answers 10 U.S. social-attitude questions as that person would, with a single answer. Share of actual answers predicted, in percent; higher is better.
Metric: Individual accuracy (%; General Social Survey 2016-2024 waves, 10 U.S. social-attitude questions, 993 sampled respondent-question rows, profile of age, sex, race, degree, region, political views and party; single-answer prompt (Style A): given the demographic profile of a real respondent, return one answer option; exact-match share of sampled respondent-question pairs whose actual answer is predicted; two sampling runs per cell at temperature 1.0 with a 100-token cap, unreadable answers flagged invalid). Source: arxiv.org. Saturation forecast: Around 2029. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude Haiku 4.5 | 58.7 | #568 |
| 2 | Llama 3.3 70B Instruct | 56.9 | #1357 |
| 3 | Claude Sonnet 4.6 | 56.3 | #161 |
| 4 | Llama 3.1 8B Instruct | 49.6 | #4246 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=synthetic-users-benchmark-gss-individual-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-13.