← News

ResearchSources: arXiv · 16 Nov 2026

arXiv paper turns long survey histories into compact personas for LLM digital twins

Researchers describe an iterative pipeline that compresses a respondent's survey transcript into a task-specific persona, reporting accuracy gains for one model on a pricing task.

In plain words

A research team says it can replace a person's long record of survey answers with a short, learned persona that a large language model uses to predict that person's choices. On one pricing task, this compressed representation predicted unseen respondents' answers more accurately than feeding the model the full transcript. The results come from a single task and small sample, so they are not yet a general result.

What was announced

The arXiv paper (v2, announced as a replacement) reports an iterative auto-discovery pipeline that builds persona representations and learns how persona evidence influences simulated responses. The authors evaluate it on the 40-item Product Preferences–Pricing task of the Twin-2K-500 dataset, learning from 20 respondents and 20 products, then testing the frozen representation on 30 unseen respondents and 20 unseen products. Holdout answers come from two survey waves, and predictions are scored against each wave separately.

For gpt-5.4-nano, swapping the full survey transcript for the learned representation and mechanism raises exact accuracy from 0.573 to 0.623 in one wave and from 0.558 to 0.626 in the other. The authors add that larger and cross-provider simulators (gpt-5.4, gpt-5-mini, qwen3-8b) using the full transcript do not beat the baseline significantly in both waves, and none significantly outperforms the learned representation in either wave. The abstract frames this as consistent gains while keeping an explicit mapping from respondent evidence to predicted decisions.

What it does not show

  • Evidence is limited to the paper's abstract; the full article was not reviewed, so methods and caveats beyond the abstract are unverified.
  • Results cover one 40-item pricing task with a small learning set (20 respondents, 20 products); no other tasks or datasets are shown.
  • The reported gains are the authors' own measurements on one model (gpt-5.4-nano); no independent confirmation is available.

Sources

Source-support review completed 17 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 17 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.