COLLABORATIVE LLM-SLMARCHITECTURE FORUNSTRUCTURED DATA ANALYTICS

The Hong Kong University of Science and Technology (Guangzhou)

Data Science and Analytics Thrust

PhD Qualifying Examination

By Mr LIN Teng

Abstract

The exponential growth of unstructured data across industries has intensified the demand for scalable and efficient analytics solutions. While large language models (LLMs) demonstrate unparalleled capabilities in semantic understanding and cross-modal reasoning, their practical deployment faces critical challenges, including computational latency from ultra-scale LLMs, inefficiencies in cross-modal retrieval, cost-accuracy tradeoffs, and semantic fragmentation across heterogeneous modalities. This survey systematically investigates the emerging paradigm of collaborative LLM-SLM architectures that synergize the contextual depth of LLMs with the agility of small-scale language models (SLMs) to address these limitations. Central to this architecture is a closed-loop workflow comprising Parsing SchedulingExecution-Feedback (PSEF) phases, enabling granular task decomposition, context-aware model dispatching, and iterative optimization. The survey further identifies unresolved challenges, including semantic consistency in cross-modal alignment, ethical implications of automated decision pipelines, and scalability constraints in heterogeneous data environments. By synthesizing cutting-edge research from peer-reviewed studies, this work provides a roadmap for advancing collaborative architectures, emphasizing the need for adaptive training protocols, resource-aware optimization, and domain-specific benchmarking. Our analysis concludes that LLM-SLM collaboration represents a critical evolution in unstructured data analytics, balancing performance, sustainability, and operational practicality for real-world deployment.

PQE Committee

Chair of Committee: Prof. WANG Wei

Prime Supervisor: Prof. TANG Nan

Co-Supervisor: Prof. LUO Yuyu

Examiner: Prof. YANG Weikai