AI Machine Learning & Data Science Research

Stanford U & Google’s Generative Agents Produce Believable Proxies of Human Behaviours

In the new paper Generative Agents: Interactive Simulacra of Human Behavior, a team from Stanford University and Google Research presents agents that draw on generative models to simulate both individual and emergent group behaviours that are humanlike and based on their changing experiences and environment.

The quality and fluency of AI bots’ natural language generation are unquestionable, but how well can such agents mimic other human behaviours? Researchers and practitioners have long considered the creation of a sort of sandbox society populated by agents with human behaviours as an approach for gaining insights into various interactions, interpersonal relationships, social theories, and more — and we may be getting there.

In the new paper Generative Agents: Interactive Simulacra of Human Behavior, a team from Stanford University and Google Research presents agents that draw on generative models to simulate both individual and emergent group behaviours that are humanlike and based on their identities, changing experiences, and environment.

The team summarizes their main contributions as follows:

  1. Generative agents, believable simulacra of human behaviour that are dynamically conditioned on agents’ changing experiences and environment.
  2. A novel architecture that makes it possible for generative agents to remember, retrieve, reflect, interact with other agents, and plan through dynamically evolving circumstances.
  3. Two evaluations (a controlled evaluation and end-to-end evaluation) that establish causal effects of the importance of components of the architecture, as well as identify breakdowns arising from, e.g., improper memory retrieval.
  4. Discussion of the opportunities and ethical and societal risks of generative agents in interactive systems.

The team sought to build a virtual open-world framework where intelligent agents go about their business and engage with each other in natural language — planning their days, sharing news, forming relationships, and coordinating group activities — all while conditioning their behaviours on both the changing environment and their past experiences. The team’s novel agent architecture combines a large language model (LLM) with mechanisms that synthesize and extract information based on the LLM outputs such that the agents learn from their past experiences how to make more accurate real-time inferences while maintaining long-term character coherence.

A core agent component is their memory stream, a database that stores a comprehensive record of past experiences. The agent can retrieve relevant information from its memory stream, reflect on this, and plan actions with regard to its changing environment. Agents can also recursively synthesize records to higher-level observations to guide more complex behaviours.

In their empirical study, the team enlisted human evaluators and had 25 of their proposed generative agents interact with each other in natural language over two full game days as non-player characters (NPC) in a Smallville sandbox environment built with the Phaser web game development framework. In the experiment, the agents maintained character coherence and demonstrated believable proxies of humanlike behaviours in remembering, planning, reacting, and reflecting.

This work advances the development of LLM-based simulacra populated by agents with dynamic and interactive humanlike behaviours, with potential applications in role-playing, social prototyping, immersive environments and gaming.

The paper Generative Agents: Interactive Simulacra of Human Behavior is on arXiv.


Author: Hecate He | Editor: Michael Sarazen


We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.

3 comments on “Stanford U & Google’s Generative Agents Produce Believable Proxies of Human Behaviours

  1. It would be fascinating to play out scenarios in the SIM City environment applying different heuristics into multiple games running side by side and seeing how they play out differently. I guess people are doing that all the time now.

  2. Pingback: Reading list of GenAI 101 for whoever needs it – Unsatisfied

Leave a Reply

Your email address will not be published. Required fields are marked *

%d bloggers like this: