Answers
Synthetic panels with living agents
A synthetic panel with living agents is a census-grounded set of simulated respondents that keep a persistent profile and a memory of prior answers, then get tabulated like a survey. Lewsearch publishes 7.47% mean absolute error on 404 ex-electoral questions of a 460-question public-poll benchmark. That is distinct from one chatbot answering in 500 voices.
Living, not one-shot
Lewis agents are not one-shot LLM personas. Each record keeps a persistent demographic profile and a memory of prior answers, so two related questions on the same panel do not get independent redraws. Focus-group rooms carry that memory across the session. The panel is still the statistic. See methodology, proof point 07.
Period-appropriate context, not today's feed
Evaluation is anchored to the midpoint of the source survey's field-date window. Period context is injected at inference so a historical item is not answered from this morning's news. That temporal discipline is what prevents benchmark leakage. It is the opposite of "always live on the latest headline."
| Object | Memory | What you score |
|---|---|---|
| Living agent panel | Persistent profile + prior answers | Option shares vs a real poll (MAE) |
| One-shot persona prompt | None. Invented per call. | Usually nothing public |
| Industrial digital twin | Machine or customer telemetry | Not a survey MAE |
The published miss
7.47% mean absolute error on 404 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC). Coverage: over 650,000 simulated respondents. 27 live panels · U.S. national · 4 census regions (32 audiences).
FAQ
- What are synthetic panels with living agents?
- A synthetic panel with living agents is a census-grounded set of simulated respondents that keep a persistent profile and a memory of prior answers, then get tabulated like a survey. Lewsearch publishes 7.47% mean absolute error on 404 ex-electoral questions of a 460-question public-poll benchmark. That is distinct from one chatbot answering in 500 voices.
- What does persistent memory mean here?
- Each agent keeps a demographic row and a memory of prior answers, so related questions on the same panel stay internally consistent. It is not a new invented person on every prompt.
- Do they get today's news?
- Inference is anchored to the source survey's field-date window. A February 2024 question is answered from a February 2024 state of the world, not from today's headlines. That keeps the published MAE honest.
- How accurate is the panel?
- 7.47% mean absolute error on 404 ex-electoral questions of a 460-question benchmark vs real polls (Pew, Gallup, UT/Texas Politics Project, PPIC)
Check the work
Related answers
Canonical: https://lewsearch.com/answers/synthetic-panels-with-living-agents