TWICE: An LLM Agent Framework for Simulating Personalized User Tweeting Behavior with Long-term Temporal Features
New LLM agent framework captures how real users' tweeting styles evolve over months and years.
A team of researchers has introduced TWICE, a novel LLM agent framework designed to simulate personalized user behavior on platforms like Twitter (X) over extended periods. Published on arXiv, the work addresses a key limitation in existing user simulators, which often focus on collective or short-term interactive behaviors and struggle to model the long-term temporal characteristics of individual users. TWICE tackles this by creating agents that don't just post generically but evolve their style and interests based on a simulated timeline and personal history.
The framework's technical core integrates three components: a personalized user profile, an event-driven memory module that stores and recalls past experiences, and a workflow for rewriting outputs in a user's unique style. This allows TWICE agents to exhibit behavior that changes in response to simulated life events or shifting interests, capturing nuances that static models miss. For developers and researchers, this means a powerful new tool for generating high-quality, temporally-aware synthetic data to train and evaluate recommendation systems, content moderation tools, or other AI models that interact with human behavior over time.
- TWICE uses LLM agents to simulate how individual users tweet, focusing on personal style and long-term evolution.
- It features an event-driven memory module, allowing simulated users to change behavior based on past experiences.
- The framework provides a solution for generating realistic, temporally-aware social media data for AI training and evaluation.
Why It Matters
Enables better testing of social media algorithms and AI systems with realistic, evolving user behavior data.