New Framework Simulates Human Daily Activities, Enhancing Agent Proactivity

Researchers have developed a desire-driven intelligent agent framework capable of simulating human daily activities and interactions. This framework generates action sequences that reflect natural human behavior more accurately compared to previous methods based on explicit goal and role scenarios or rule-based decision-making systems like LLMob, ReAct, and BabyAGI. The team found that the D2A-generated action sequences effectively reduce the dissatisfaction level of various desires (more rational), showing a high degree of consistency with human driving models. They also expanded the D2A framework to multiple intelligent agent environments, discovering that it can produce behavior sequences that are richer and more lifelike. According to peer reviews, this research introduces a novel desire-driven proactive framework for modeling human-like activities and interactive systems. The team achieved this by converting psychological theories into computable models, where an internal value system replaces traditional high-level task management systems in agents. This allows the agents to generate more intelligent and natural behavior sequences through self-initiated actions rather than pre-programmed instructions. The researchers conducted experiments in various settings, both indoor and outdoor, using different objects and scenarios. They used clear performance metrics, including heat maps and dissatisfaction curves, to demonstrate the relationship between desire value and agent behavior, as well as the efficiency of their method compared to baseline models. These analyses showed significant improvements in behavior simulation quality. Initially, the team's work received some skepticism, with early peer ratings averaging around 5553, which is below the standard line for acceptance. However, they supplemented their experiment data with a robust theoretical foundation and extensive interdisciplinary applications, ultimately convincing four reviewers to raise their scores to 6666 and above. This validated the novelty and practicality of their framework concept. However, the model is currently relatively simple and does not account for the layered structure of various value dimensions or incorporate complex psychological mechanisms such as emotional regulation and cognitive dissonance, which will be areas of future exploration. In addition, the team views the fulfillment of desires or values as a deeper motivational drive for action. They plan to conduct more work on multi-agent social simulations to integrate these intelligent agents into human-like social settings. This could help in studying user interactions and providing personalized services, or even in enhancing the interaction and engagement in non-player characters (NPCs) in multiplayer games. Overall, this research offers a new direction in the development of intelligent agents, emphasizing a more natural, proactive approach based on self-determined value systems. The team believes this framework has significant potential for applications in large-scale social simulators and interpersonal interaction systems, as well as in enriching the behavior of NPCs in games, thereby increasing the game's immersion and playability.