HPP: Hierarchical Programmatic Probing for Long Video Understanding by Decoupling Perception and Reasoning

The article introduces Hierarchical Programmatic Probing (HPP), a framework designed to enhance long video understanding by separating semantic perception from higher-order reasoning through iterative programmatic exploration. HPP employs a coding-capable LLM to execute multi-step strategies in a structured interactive environment, utilizing components like information-density-aware hierarchical segmentation and late-interaction semantic retrieval. Validation on benchmarks such as LongVideoBench shows significant improvements in performance, indicating that this decoupling approach can effectively address the challenges of fine-grained perception and long-range relational reasoning in video analysis.