The paper introduces Lyfe Agents, autonomous generative agents capable of simulating complex social behaviors in a virtual environment. The key features of these agents include a low-cost option-action framework for decision making, asynchronous self-monitoring for consistency, and a ‘Summarize-and-Forget’ memory mechanism for cost-effective prioritization of memory items. The agents have been tested in a custom 3D virtual environment, LyfeGame, showing their ability to exhibit human-like social reasoning. The techniques used allow Lyfe Agents to operate at a computational cost 10-100 times lower than existing alternatives. The authors conclude by highlighting the potential of such agents to enrich human social experiences in virtual worlds.
Publication date: 5 Oct 2023
Project Page: https://arxiv.org/abs/2310.02172v1
Paper: https://arxiv.org/pdf/2310.02172
