LLM

LLM-driven agents are advanced LLMs (Large Language Models) enhanced with memory, tool functionality, reflection capabilities, and other systems, enabling them to respond to environmental stimuli in a manner akin to human reactions. These agents have become increasingly popular across various domains, including coding assistance, personal assistant technology, and social simulations.

In the realm of LLM-driven social simulations, scenarios typically feature a multi-agent environment, where agents interact with both physical items and other agents. This creates a rich, dual-layered environment comprising both physical and social elements. The pioneering work in this field was conducted by researchers from Stanford and Google, who developed the first LLM-driven AI town. This groundbreaking project introduced a "perceive-execute-reflect" paradigm, marking the inception of first-generation cognitive architecture for AI communities.

However, the practical application of these AI towns in games faces two significant hurdles. Firstly, previous models often treated the physical and social environments as separate entities. While LLMs excel in navigating limitless textual scenarios, the physical world's constraints limit the scope of agent interactions, thereby hindering the simulation of complex socio-economic dynamics and diminishing the experience's immersiveness. Secondly, earlier iterations of AI towns struggled to connect meaningfully with the real world. The drive for socialization—a key motivator for game engagement—was not effectively harnessed, as simulated societies lacked tangible connections among users.

Our product addresses these challenges head-on. We've pioneered the integration of an economic system within the simulated society, effectively merging the physical and social environments. This economic system allows for the emergence of social emotions from agent interactions, influencing commodity popularity and pricing. Consequently, agents' decisions regarding purchases, shop openings, and the refinement of the physical environment are affected. Our agents possess a sophisticated understanding of the economic system, enabling autonomous decision-making concerning employment and creative endeavors.

Moreover, we bridge the gap between the virtual and real worlds by enabling our agents to produce real-world valuables, such as drawings, music, and poetry. By integrated with the strong equipments for multimodal composing, not only do our agents have the ability to create human-level artworks, but can also critique their values. These AI-built creations can be shared among users, fostering strong community ties and vibrant user communities.

Through these innovations, our work overcomes existing limitations, enhancing the integration of physical and social environments in AI towns and establishing meaningful connections between the virtual experiences and the real world. This not only enriches the gaming experience but also paves the way for more immersive and socially interconnected virtual societies.