Synthetic survey respondents are AI models that fill in questionnaires as if they were, say, a 40-year-old rural teacher in Chile. They are cheap and fast, but they tend to sound like one person in different hats. A team from Tsinghua University now says it can build each respondent from plug-in blocks instead, and it reports cutting distribution error by up to 52.8% on the World Values Survey.
Key takeaways
- The method is called PARADE. Each demographic trait (one country, one age band, one religion) gets its own small add-on to the model.
- On the World Values Survey, the paper builds 106 demographic blocks plus 1 shared block that teaches the model how to answer surveys.
- Blocks are combined when the model runs, so it can simulate trait combinations it never saw in training.
- KL divergence, a measure of how far the AI’s answer spread sits from real people’s, fell by up to 52.8% against the strongest baseline.
- The authors say the method assumes traits add up independently, and that it should support real surveys, not replace them.
What is PARADE, in plain English?
The paper, “Composable Parametric Demographic Injection for Values Survey Simulation”, was submitted to arXiv on 13 September 2026 and appeared in the 20 October 2026 listings. Its ten authors list Tsinghua University’s computer science department and Quancheng Laboratory, and Qingyao Ai is the corresponding author.
The core tool is LoRA (low-rank adaptation). A LoRA is a small set of extra weights clipped onto a large model, so you can teach it something new without retraining the whole thing. PARADE uses one shared “Task LoRA” for general survey-answering skill. On top of that sit separate “Demographic LoRAs”, one for each value of each trait.
Think of a spice rack. The Task LoRA is the base recipe. Each demographic block is a jar: “Brazil”, “aged 30-49”, “university degree”, “rural”. To simulate a person, you take down the jars that match their profile and stir them in.
Why do synthetic survey respondents need a fix?
Researchers want simulated respondents because real surveys are slow and expensive, and some groups are hard to reach. If a model could stand in for a hard-to-reach group, you could test question wording or explore that group’s likely answers before paying for fieldwork.
So far the results have been patchy. The most common method, demographic prompting, means writing “You are a 55-year-old Catholic woman from Poland” into the prompt. The PARADE authors point out that this breaks easily when the wording changes, and it adds extra text to every single question you ask.
Other teams have found deeper problems. A cross-domain benchmark from researchers at Stevens Institute of Technology and UMass Boston, “When Synthetic Users Fail”, found that no LLM it tested beat the strongest simple baseline at predicting individual World Values Survey answers. It also found that models treat identity as far more predictive than it really is. In one US example, political views explained about 1.5% of real variation in confidence in banks, but up to about 67% in the model.
Earlier work from part of the same Tsinghua group, Parametric Social Identity Injection (accepted to KDD 2026), called this “Diversity Collapse”: inside the model, different identities blur together, and the answers flatten into one bland middle.
How do the plug-in blocks work?
The World Values Survey setup uses eight traits: gender (2 values), age group (3), country (66), religion (10), education (9), marital status (6), employment (8) and urban or rural residence (2). That gives 106 demographic blocks, or 107 counting the shared Task LoRA.
Stirring the blocks together naively causes them to interfere with each other, so the authors added “composition calibration” in two parts:
- Learned weights per trait. Each trait, such as country, gets a dial that can turn its influence down to 0.5x or up to 1.5x.
- An orthogonality penalty. This pushes the demographic blocks to use different directions from the Task LoRA, so they don’t overwrite general survey skill.
Both parts pull their weight. In the paper’s ablation test on one model, removing the orthogonality penalty raised KL divergence by 79%, and removing the weights raised it by 51.1%. Removing the demographic blocks entirely dropped accuracy from 47.71% to 26.99%.
How well did it score?
The team tested three open models of about 8-9 billion parameters: Llama-3.1-8B-Instruct, Qwen3.5-9B and Ministral-3-8B-Instruct. They used two datasets, the World Values Survey and SocioBench.
The headline is about answer spreads, not single guesses. Against the strongest baseline, PARADE cut KL divergence by up to 52.8% and Jensen-Shannon divergence (a similar distance measure) by up to 43.6%, both with Ministral on WVS. Accuracy on individual answers stayed roughly level with the best baselines, at 47.7% to 48.9% on WVS.
The more interesting test covers unseen combinations. The authors picked WVS respondents whose full eight-trait profile never appeared in training, though each single trait did. There were 915 eligible people, sampled 100 at a time. Across these settings, PARADE ranked first 24 times and second four times.
How does it compare with other approaches?
| Approach | How identity gets in | Main weakness |
|---|---|---|
| Persona prompting | Text description in every prompt | Breaks when wording changes; extra text on every question |
| Single shared LoRA | One fine-tuned adapter for everyone | No explicit demographic conditioning |
| Mixture of adapters (e.g. CuMA) | Several experts routed by demographics | Reuses trait information less efficiently |
| PARADE | One reusable block per trait value, combined when the model runs | Assumes traits add up independently |
Where does it fall short?
The authors list the limits plainly, which is worth crediting. The big one is additivity. PARADE treats “young” plus “religious” plus “rural” as a sum, but real people’s traits interact. Being young and religious in one country can mean something different from the same pair elsewhere.
Coverage is fixed too. The method can recombine trait values it has seen, but it cannot invent a block for a country missing from training. The evaluation also scores each question’s answer spread separately, because full profiles are too rare to measure. And the ablation ran on only one model.
The paper’s ethics statement warns that demographic conditioning can reproduce stereotypes, and that synthetic respondents should complement representative surveys, not replace them. My opinion: that warning matters more than the 52.8%. A tool that is better at sounding like a group is also better at making a stereotype look like data.
What happens next
The authors name two directions: blocks that model how traits interact, and calibration that changes depending on the question. Neither has been built yet. The paper is a v1 preprint, it has not been peer reviewed, and the arXiv page links no code repository.
The bigger test is outside the lab. Cutting divergence on benchmarks is one thing. Showing that a simulated group leads a researcher to the same conclusion as a real sample is harder, and the Stevens benchmark suggests current models often fail it. If you follow how small models get evaluated, our pieces on an LLM calibration agent and cheap models misreading the night sky cover the same gap between benchmark scores and reliable answers.
FAQ
What are synthetic survey respondents?
They are AI language models set up to answer survey questions as a person with a given profile would, such as a specific age, country and education level. Researchers use them to explore likely answers before running a real survey.
What does the 52.8% figure measure?
It is the largest drop in KL divergence against the strongest baseline. KL divergence measures how far the model’s spread of answers sits from real respondents’ spread. The figure comes from the Ministral model on the World Values Survey, so it is not an accuracy score for individual answers.
Can PARADE replace real surveys?
No. The authors say synthetic respondents should complement representative surveys. The method assumes traits act independently, it cannot handle trait values missing from training, and it may reproduce stereotypes.
If you work with survey data, read the full paper’s limitations section before you trust any synthetic sample, then test it against one real group you already know well.
Leave a comment