Efficient Neural-Symbolic Data System via Multi-Agent Collaboration

The paradigm of scaling Large Language and Multimodal Models (LLMs/MLMs) is encountering a significant bottleneck, as the 'Scaling Law' of amassing more data and compute becomes unsustainable. To power the next generation of AI, a new paradigm of data systems is required that efficiently fuses the powerful perceptual capabilities of neural networks with the formal, verifiable logic of symbolic systems in order to better manage and analyze data. Recent efforts have demonstrated the potential of this paradigm, yet existing systems remain limited by black-box neural architectures, expensive inference costs, and inadequate support for complex, iterative reasoning workflows over large amounts of multimodal data. In this work, we envision NSDS, the next-generation NeuralSymbolic Data System, which supports dynamic, recursive, and multi-stage reasoning with explainability and lineage auditing over different types of data across heterogeneous programming environments. In particular, NSDS is a hybrid framework combining both agentic and deterministic execution and provides a natural solution for automating task decomposition, iterative refinement, and coordinated execution across neural and symbolic components. We further analyze the key technical challenges in designing such a system, including task decomposition, hardware heterogeneity, multi-agent coordination, unified data storage, and lineage auditing, etc. Finally, we outline the architectural principles that can enable the next-generation neural-symbolic data system.