Recent breakthroughs in Large Language Models (LLMs) revolutionize the field of Natural Language Processing (NLP), driving advancements in comprehension, reasoning, and content generation. These models achieve remarkable performance in specific tasks but encounter limitations in complex social interactions, motivating the development of multi-agent systems that enable dynamic collaboration and competition. LLM-based multi-agent systems, using natural language as the primary medium of interaction, emerge as key enablers of digital social relationships and collective intelligence. At the same time, these systems introduce challenges in modeling agent social behaviors, optimizing collaboration efficiency, and adapting game-theoretic strategies. This survey addresses gaps in the existing literature by systematically exploring the theoretical foundations of LLM-powered social agents, drawing insights from cognitive science and sociology. We first introduce relevant sociological and cognitive theories and propose framework to model social intelligence in multi-agent systems. Based on this framework, we comprehensively review key techniques for social agent construction, collaboration, competition, as well as their application scenarios and evaluation methodologies. Finally, we identify future research directions, with particular emphasis on constructing large-scale cognitive world models and modeling the growth and evolution of multi-agent systems. This work establishes a solid foundation for advancing the understanding of social dynamics in LLM-based multi-agent environments.
Traditional source separation methods rely on coarse-grained categorical labels through labeling all vocals collectively without distinguishing individual voices in an audio mixture, which inherently limits the ability to isolate single tracks. While fine-grained annotations could partially address this issue, they demand substantial resources and face challenges in extracting tracks from raw signals. To overcome these limitations, we propose to extract each track through decomposing the patterns of data generation. Specifically, we refine Variational Stochastic Dirichlet Process-VAE, a variational autoencoder framework through replacing the standard variational distribution by a variational stochastic Dirichlet process (VSDP). Among our proposed framework, the encoder, leveraging stick-breaking constructions, adaptively partitions the latent space into clusters, while the decoder designed to recover each component achieves implicit signal separation. Its advantage is that the reconstruction target can be shifted from the raw input to its individual components. Experiments demonstrate our method’s efficacy in two scenarios: (1) Under coarse-grained source definitions, it reaches near-state-of-the-art performance (SDR=10.3); (2) For fine-grained track separation, what’s more, it identifies 83% of individual vocal tracks with an average SDR of 7.8, which cannot be obtained by other SOTA methods without the help of annotations.
