Reproduce “Synthetic Social Media Influence Experimentation via an Agentic Reinforcement Learning Large Language Model Bot” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bailu Jin, Weisi Guo
Researchers from Cranfield University created a synthetic social media environment that combines agent-based modeling with a reinforcement learning agent and large language models for ethical experimentation on online influence. The system demonstrated that an RL opinion leader could consistently acquire followers, achieving a 10% gain, with performance optimized by a constrained action space and the agent's self-observation.
Key insights
- Constraining the Reinforcement Learning agent's action space and incorporating self-observation are crucial for stable and consistent generation of opinion leaders for topic-specific influence.
- The 'creativity' setting of the LLM significantly impacts opinion dynamics; 'Narrow' settings promote opinion convergence, while 'Creative' settings result in more dispersed and chaotic opinion distributions.
- Increased environmental complexity (follow-unfollow dynamics) and reduced observability (part-observable scenarios) introduce greater variance and learning challenges for the RL agent.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.