LLMs Simulate Survey Respondents with High Accuracy

Chan-Tung Ku, Chan Hsu, Pei-Cing Huang, Frank Cheng-shan Liu, I-Ling Cheng, Yihuang Kang· July 7, 2026 View original

Summary

This research introduces "cross-survey transfer" to rigorously evaluate LLMs as simulated human survey respondents, finding that zero-shot LLMs can predict answers to unseen questions with 52% accuracy, closing the gap with supervised models. It also clarifies nuances around variance collapse and safety alignment effects in LLMs.

A new evaluation framework called "cross-survey transfer" has been proposed to assess the effectiveness of large language models (LLMs) in simulating human survey respondents. Unlike previous methods that focused on distributional comparisons, this framework tests an LLM's ability to predict answers to entirely new questions within the same survey, given its responses to an initial set. Using data from the Taiwan Election and Democratization Study (TEDS) 2024, the study found that zero-shot LLMs achieved 52% accuracy on genuinely unseen items, approaching within 6 percentage points of a supervised random forest model trained on the same population. The research also provided clearer insights into common LLM limitations, showing that "variance collapse" can also affect supervised models and "safety alignment" effects vary significantly across different model families.

Why it matters

Silicon sampling offers a promising, cost-effective method to augment traditional survey research, enabling faster insights and potentially reducing reliance on human respondents for certain types of data collection.

How to implement this in your domain

  1. 1Experiment with LLMs to generate synthetic survey data for preliminary market research.
  2. 2Develop internal benchmarks using cross-survey transfer to validate LLM-generated responses.
  3. 3Integrate LLM-based respondent simulation into survey design workflows to test question efficacy.
  4. 4Analyze the cost-benefit of using silicon sampling versus traditional human surveys for specific research needs.

Who benefits

Market ResearchSocial SciencesPublic Opinion PollingProduct Development

Key takeaways

  • LLMs can simulate human survey respondents with notable accuracy.
  • Cross-survey transfer is a rigorous method for evaluating LLM-based sampling.
  • LLMs can predict answers to unseen questions, nearing supervised model performance.
  • Limitations like variance collapse and safety alignment are more nuanced than previously thought.

Original post by Chan-Tung Ku, Chan Hsu, Pei-Cing Huang, Frank Cheng-shan Liu, I-Ling Cheng, Yihuang Kang

"arXiv:2607.03091v1 Announce Type: new Abstract: Silicon sampling-using large language models (LLMs) to simulate human survey respondents-has emerged as a promising approach for augmenting traditional survey research. However, most evaluations rely on distributional comparisons ra…"

View on X

Originally posted by Chan-Tung Ku, Chan Hsu, Pei-Cing Huang, Frank Cheng-shan Liu, I-Ling Cheng, Yihuang Kang on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI in Marketing

AI Engineering & DevToolsAI ResearchAI in Marketing

New Bandit Algorithm Handles Drifting Rewards and Actions

This research introduces a new approach for non-stationary linear bandits with round-specific feasible decision sets, addressing limitations of existing methods that assume orthogonal structures. It achieves optimal dynamic regret by viewing the problem through a misspecification-reduction lens, partitioning the horizon into blocks and relating regret to fixed-parameter benchmarks.

Zihao Hu, Yuan Yao, Jiheng Zhang, Zhengyuan ZhouJul 7, 2026
AI Engineering & DevToolsAI in MarketingAI Research

Concept Graphs Mitigate Bias in Text-to-Image Diffusion Models.

CO-ALIGN, a novel bias mitigation approach, reduces harmful bias in Text-to-Image (T2I) diffusion models by aligning concepts within the model's internal concept ontology using concept graphs. This method significantly improves fairness and image quality while reducing semantically incoherent outputs.

Mansi, Avinash Kori, Francesco LeofanteJul 7, 2026
AI Engineering & DevToolsAI ResearchAI in Marketing

APeB Benchmarks LLM Agent Personalization for Raw Queries

This research introduces APeB, a new benchmark for evaluating the personalization capabilities of LLM-powered agents, particularly when handling underspecified user queries and noisy interaction histories. It reveals that current models struggle with intent and preference discovery, highlighting the need for dedicated history-utilization modules.

Garry Yang, Zizhe Chen, Xinru Chen, Yongqiang Chen, Jianxiang Wang, Deyu Zou, Linyi Ding, Jialiang Wu, Yunzhong He, Yu Gong, James Cheng, Huaixiao TouJul 7, 2026