Reproduce “Plurals: A System for Guiding LLMs Via Simulated Social Ensembles” 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
Joshua Ashkinaze, Emily Fry, Narendra Edara, Eric Gilbert, Ceren Budak
The University of Michigan developed Plurals, a system that guides large language models (LLMs) through simulated social ensembles grounded in deliberative democracy theory. It generates outputs that better resonate with specific human audiences, with preferences for Plurals outputs ranging from 66% to 88% over standard methods, and allows for the creation of value-based ethical guardrails.
Key insights
- Interactional pluralism, a novel theoretical contribution, suggests that AI pluralism exists in both agent characteristics and their interaction protocols.
- Integrating intersectional demographic data from sources like ANES for persona generation leads to significantly more diverse LLM outputs.
- LLM agents can effectively follow complex deliberative instructions (e.g., focusing on rational arguments or emotional communication) during multi-turn interactions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.