An open-source polling protocol that surveys social media users using multimodal large language models accurately predicted election outcomes in 50 of 51 states during the 2024 US Presidential Election, according to research published in EPJ Data Science. Developed by researcher Roberto Cerina, the framework, named PoSSUM, is designed to conduct unobtrusive polling of social media participants. In its final pre-election assessment, the system assigned the Republican candidate an estimated win probability of 0.65 while exhibiting lower state-level bias than most established polling organizations. The study authors report that these predictive results indicate the potential of automated language model systems to serve as unobtrusive alternatives to conventional survey methods.
The architecture incorporates multimodal large language models capable of analyzing multiple forms of online content, specifically leveraging live user posts, shared images, and other digital traces. By extracting and evaluating these active digital records, the protocol builds synthetic respondent groups, termed silicon samples. According to the research, these silicon samples capture information not present in the LLM's training data.
To counteract the observable selection biases of social media platforms, PoSSUM incorporates Multilevel Regression and Post-stratification, a statistical modeling technique known as MrP. The researchers configured this post-stratification framework with structured priors specifically designed to adjust for platform-specific imbalances.
The predictive performance of the protocol was evaluated across the 2024 electoral cycle through five distinct PoSSUM polls, which were conducted and published openly on GitHub and the social media platform X. The research culminated in a final poll fielded between October 17 and October 26, 2024, analyzing a synthetic sample of 1,054 users drawn from X. Across that final fielding window, the protocol achieved its benchmark accuracy by correctly forecasting the winner in 50 of the 51 contested states.
These results highlight the need for further validation across contexts and outcomes.
Subject of Research: Multimodal language model polling protocol using social media data
Article Title: PoSSUM: a protocol for surveying social-media users with multimodal LLMs
Article References: Cerina, R. (2026). PoSSUM: a protocol for surveying social-media users with multimodal LLMs. EPJ Data Science. https://doi.org/10.1140/epjds/s13688-026-00705-7
Image Credits: AI Generated
DOI: 10.1140/epjds/s13688-026-00705-7
Keywords: artificial intelligence, large language models, social media polling, election forecasting, multilevel regression, post-stratification, PoSSUM, computational social science, protocol, surveying, social-media, users
News Source: Denise Maddox. (November 7, 2026). Multimodal AI Polling Protocol Forecasts US Election Outcomes via Social Media. Scienmag.



