Presented by: Aman Singhal

Venue: Large Language and Vision Models Symposium, NYU Center for Data Science

Year: 2024

Abstract

AI-powered characters that simulate believable human behavior using large language models. The agents have an architecture that stores memories, reflects on experiences, and plans future actions. Demonstrated in a Sims-like sandbox with 25 agents, they autonomously perform daily activities, form relationships, and coordinate complex social behaviors (like organizing a Valentine’s Day party from a single prompt).

Presentation Slides

View slides in new window

Key Topics

• Agent architecture: memory, reflection, and planning
• Large language models for behavioral simulation
• Memory stream and retrieval mechanisms
• Reflection and higher-level reasoning
• Planning and action generation
• Sandbox environment with 25 autonomous agents
• Emergent social behaviors and relationships
• Complex coordinated activities from simple prompts
• Applications in gaming, simulation, and AI research