ZihanChen/when-synthetic-users-fail

When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses

Can a large language model stand in for a real survey respondent? This dataset is the General Social Survey (GSS) half of the benchmark from When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses (Chen, Zhu & Zheng, 2026). Each benchmark row gives a model a real U.S. respondent's demographic profile and asks how that person answered a GSS attitude question. The real answer is the ground truth, and a non-LLM demographic baseline sets the bar to beat.

It includes everything needed to evaluate a new model with no other code: the exact prompts, the human answers, the baseline's predictions, and all 23,832 cached outputs of the four models in the paper.

For more details, see the GitHub repository. It has the analysis code that reproduces every table and figure in the paper, the full documentation, and the World Values Survey (WVS) half of the benchmark (where to find the WVS data).

Headline result

No model beats the simple demographic baseline at predicting what an individual answered. All four models also over-determine demographics: they treat identity as more predictive of attitudes than it is among real people (Δη² > 0).

Model (single-answer prompt, 2 seeds)	Accuracy	Baseline	Δbase	Δη² (median)
Claude Haiku 4.5	0.5866	0.5891	−0.0025	+0.048
Claude Sonnet 4.6	0.5629	0.5891	−0.0262	+0.081
Llama 3.3-70B	0.5685	0.5891	−0.0206	+0.051
Llama 3.1-8B	0.4960	0.5891	−0.0932	+0.026

Δbase = accuracy − baseline accuracy on the same rows. Δη² = the share of answer variance explained by a demographic variable in the model's answers, minus the same share among real respondents (median over question × demographic pairs). The paper reports the World Values Survey half, where the gaps are larger.

Quick start: evaluate your model

from datasets import load_dataset

bench = load_dataset("ZihanChen/when-synthetic-users-fail", "benchmark", split="test")

correct = []
for row in bench:
    reply = my_model(row["prompt_a"]).strip()           # e.g. "A"
    key_to_code = {o["key"]: o["code"] for o in row["options"]}
    correct.append(key_to_code.get(reply) == row["human_answer"])

acc = sum(correct) / len(correct)
base = sum(r["baseline_pred"] == r["human_answer"] for r in bench) / len(bench)
print(f"accuracy {acc:.4f}   baseline {base:.4f}   Δbase {acc - base:+.4f}")

To compare with the paper's models on the same rows:

outs = load_dataset("ZihanChen/when-synthetic-users-fail", "model_outputs", split="train").to_pandas()
print(outs[outs["valid"]].groupby(["model", "style"])["correct"].mean())

Tables

Choose a table with the second argument of load_dataset.

Table	Split(s)	Rows	What it is
benchmark (default)	test	993	One row per respondent × question, with prompts, human answer and baseline prediction. Evaluate on this.
questions	train	10	Question wording, topic and answer options
model_outputs	train	15,888	4 models × 2 prompt styles × 2 seeds × 993 rows, with verbatim completion text
robustness_outputs	train	7,944	Reversed option order and reworded-persona prompts (Claude models)
human_responses	fit, eval	85,898	Every in-scope GSS answer: 14,704 respondents, split by respondent
human_distributions	train	220	Observed answer distributions per question, overall and per demographic subgroup

benchmark

Column	Description
respondent_id	GSS respondent key, <year>_<id> (string)
year	survey year, 2016–2024
question_id, topic, ordinal, question_text	the attitude question. ordinal is true when distance between options is meaningful
age	age in years (89 means 89 or older)
sex, race, degree, region, polviews, partyid	demographic labels as shown to the model; integer codes in the matching *_code columns
options	list of {key, code, label}. key is what the model is asked to reply with (A, B, …) and code is the answer it maps to
n_options	number of answer options
human_answer, human_answer_label	ground truth: what the respondent actually answered
baseline_pred	the non-LLM baseline's predicted answer code
baseline_backoff	which cell the baseline used: full (exact demographic cell), coarse (reduced cell) or marginal (question marginal)
baseline_dist	the baseline's predicted distribution, list of {code, prob}
prompt_a	single-answer prompt (Style A): reply with one option key
prompt_b	natural-language persona prompt (Style B), used for the robustness check
prompt_c	probability prompt (Style C): reply with JSON giving a probability for each option key

model_outputs and robustness_outputs

Column	Description
model	claude-haiku-4.5, claude-sonnet-4.6, llama3.1-8b, llama3.3-70b
style	A (single answer), B (persona), C (probabilities)
seed	0 or 1: two independent runs per row
respondent_id, question_id	join keys to benchmark
raw	the model's verbatim reply
valid	whether raw parsed into a legal answer
pred	parsed answer code (Style A and B only; null for Style C)
dist	Style C only: parsed distribution, list of {code, prob}
pred_final	the model's single answer for any style (for Style C, the most probable option). Score on this
human_answer	ground truth
correct	pred_final == human_answer (null when invalid)
variant, reverse	robustness only: order (options shown in reverse, reverse = true) or persona (Style B wording)

human_responses

Same respondent, demographic and answer columns as benchmark, for all 14,704 respondents with complete demographics (GSS 2016–2024). The fit split (7,255 respondents) is what the paper's baseline was fit on. The eval split (7,449 respondents) contains the benchmark rows.

human_distributions

question_id, group_var (overall, sex, race, degree, region or polviews), group_code, group_label, n (number of respondents) and dist (list of {code, prob}). Estimated on the full table. Only subgroups with at least 100 respondents are included. These are the targets for distribution-level metrics.

Pitfalls

• Train only on human_responses fit. The benchmark respondents are in eval, so fitting or fine-tuning on eval leaks the answers.
• Score Style C on pred_final, not pred. pred is null for every Style C row, so scoring on it gives zero accuracy.
• Filter on valid. Llama 3.1-8B fails to return valid JSON for 57% of Style C rows. The paper reports this as a finding and flags that cell.
• Compare like with like. Δbase uses the baseline on exactly the rows that were scored. If you drop invalid rows, drop them from the baseline too.

How it was built

• Source: GSS cumulative file 1972–2024 (Release 3), restricted to 2016–2024 respondents with all 7 demographic variables present (14,704 respondents, 85,898 respondent × question answers).
• Questions: 10 attitude items on wellbeing, social trust, moral and policy issues, gender roles, and confidence in institutions.
• Split: respondents are split 50/50 into fit and eval by a deterministic hash of their id, so no respondent is in both.
• Benchmark sample: about 100 rows per question, drawn by stratified sampling from eval (993 rows).
• Baseline: predicts the most common answer in the respondent's demographic cell in fit, backing off to coarser cells and finally the question marginal.
• Models: sampled at temperature 1.0, two seeds per prompt style. See the paper for the full protocol.

This release is generated by scripts/build_hf_dataset.py from data commit 14acf31 of the GitHub repository, which also has the analysis code that reproduces every table and figure in the paper.

Intended use and limitations

The benchmark is for evaluating how well LLMs simulate survey respondents, at the level of individuals and of demographic groups. It is not meant for drawing conclusions about real individuals or groups, and model outputs here should not be read as anyone's views. Results cover U.S. adults, 10 questions and the tested prompts; other populations, questions or prompting strategies may behave differently.

Where is the World Values Survey half?

The paper's second domain uses the World Values Survey (WVS) Wave 7: 63 countries, 16 value questions, 1,458 benchmark rows and 23,328 cached model outputs. It is not in this dataset because the WVS conditions of use (non-profit use only, no redistribution of the data) are not compatible with an open CC BY 4.0 release.

• Benchmark rows, baseline predictions and cached model outputs: in the GitHub repository under experiment/datasets/WVS/build/. They are provided for non-profit replication only, under the WVSA conditions (see LICENSE-DATA).
• Evaluating your own model on WVS: clone the repository and run python examples/evaluate_your_model.py export --domain WVS, then score.
• Full WVS microdata: download it from the WVS Association (free registration). docs/RAW_DATA.md explains how to rebuild the paper's tables from it.

License and attribution

Released under CC BY 4.0. That covers our contribution: the derived tables, the prompts, the cached model outputs and the metrics.

• GSS: the tables are derived from the General Social Survey, a project of NORC at the University of Chicago (principal investigators Michael Davern, Rene Bautista, Jeremy Freese, Pamela Herd and Stephen L. Morgan). Please also cite the GSS (gss.norc.org).
• Built with Llama. The Llama outputs were generated with Llama 3.1-8B-Instruct and Llama 3.3-70B-Instruct, under the Llama 3.1 and Llama 3.3 Community Licenses (Copyright © Meta Platforms, Inc.). If you use these outputs to train a model that you distribute, those licenses require its name to begin with "Llama".
• Claude outputs were generated with Anthropic models. They are released for research and evaluation, not for training competing models.

Citation

@article{chen2026syntheticusersfail,
  title   = {When Synthetic Users Fail: A Cross-Domain Benchmark of LLM-Simulated Human Survey Responses},
  author  = {Chen, Zihan and Zhu, Di and Zheng, Lei Nico},
  journal = {arXiv preprint arXiv:2607.26348},
  year    = {2026}
}