MIDSim: LLM agents simulate social media info diffusion across channels
New system models both social links and algorithmic exposure for realistic virality.
Researchers from the Chinese Academy of Sciences have developed MIDSim, an LLM-powered multi-agent system that simulates how information spreads across social media. Previous models largely focused on diffusion through social links (e.g., reposts, mentions), ignoring the powerful influence of platform algorithms like recommender systems that curate content feeds. MIDSim bridges this gap by modeling both social and algorithmic exposure streams simultaneously. Each user agent is driven by a large language model with personalized profiles (interests, demographics, posting history), allowing for nuanced behavioral responses.
The team constructed real-world diffusion datasets from Sina Weibo, RedNote (Xiaohongshu), and Twitter, containing actual diffusion records, user profiles, historical posts, and social relationships. In experiments replicating real diffusion events, MIDSim significantly outperformed epidemic-based, cascade-based, and point process baselines in matching macro-level diffusion curves (e.g., total shares over time). It also generated more diverse and contextually relevant comment content than prior approaches. This work opens the door to more accurate simulations of how news, memes, and misinformation propagate through modern multi-channel social platforms.
- MIDSim models both social links and algorithmic exposure (e.g., recommender systems) for multi-channel diffusion.
- Uses LLM-powered user agents with personalized profiles to simulate realistic behavior and comment generation.
- Outperforms baselines on real-world datasets from Sina Weibo, RedNote, and Twitter.
Why It Matters
Could help predict viral trends and model misinformation spread more accurately across platforms.