https://arxiv.org/abs/2304.03442
This document is an academic paper describing Generative Agents. The main points are as follows
- Generative Agents are computational agents that simulate human behavior and can be applied to a variety of interactive applications such as immersive environments, interpersonal communication exercises, and prototyping tools.
- The system consists of 25 agents in a The Sims-like environment, where users can observe and intervene in the agents’ daily planning, sharing news, forming relationships, and coordinating group activities.
- The architecture that enables Generative Agents extends the language model to fully record the agent’s experience in natural language, compile memories into high-dimensional reflections over time, and dynamically retrieve memories for use in action planning.
- Evaluation results showed that Generative Agents can generate believable behavior both at the individual agent level and at the society-wide level.
- The architecture validation revealed that observation, planning, and reflection contribute significantly to the credibility of an agent’s actions.
- We proposed an architecture and interaction method that enables believable simulation of human behavior by integrating a language model and a computational interactive agent.
https://github.com/joonspk-research/generative_agents
This page is auto-translated from [/nishio/Generative Agents: Interactive Simulacra of Human Behavior](https://scrapbox.io/nishio/Generative Agents: Interactive Simulacra of Human Behavior) using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.