Diana: Edge-Cloud Collaborative Long Video Understanding via CoT-Driven Iterative Retrieval

Abstract

The ubiquitous deployment of edge cameras and the emergence of multimodal large language models (MLLMs) have necessitated intelligent long video understanding at the network edge. However, current deployment paradigms face a critical trade-off. Cloud-centric approaches incur prohibitive bandwidth costs and privacy concerns, while edge-only solutions are hindered by limited computational resources. Furthermore, existing edge-cloud collaborative frameworks typically rely on single-pass, open-loop retrieval, often failing to extract sufficient evidence for complex reasoning due to semantic ambiguity. To address these challenges, we propose Diana, a chain-of-thought (CoT) driven edge-cloud collaborative system that strategically decouples perception from cognition. At the edge, Diana employs a lightweight content-aware perception pipeline to construct a hierarchical multimodal memory for efficient video indexing. On the cloud, we introduce a dynamic reasoning framework featuring a predictor for difficulty-aware query routing and a control module for CoTdriven iterative retrieval. This architecture establishes a closedloop reasoning mechanism, iteratively re-examining edge memory to resolve ambiguities. Extensive evaluations on the NExT-QA and MVBench benchmarks demonstrate that Diana achieves state-of-the-art accuracy (78.42% on NExT-QA), significantly outperforming baselines while reducing end-to-end latency by over 10× compared to cloud-centric methods.