ScaleLong: A Multi-Timescale Benchmark for Long Video Understanding

The paper introduces ScaleLong, a benchmark designed to evaluate multimodal large language models on long videos through a 'within-content' multi-timescale questioning approach. It resides in the Multi-Timescale Evaluation Benchmarks leaf, which contains only two papers including this one. This sparse population suggests that systematic evaluation frameworks explicitly disentangling temporal scales within identical video content remain relatively underexplored. The benchmark features 269 videos averaging 86 minutes, with questions targeting four hierarchical timescales (clip, shot, event, story) embedded in the same content, enabling direct performance comparison across scales.

The taxonomy reveals that most research effort concentrates on architectural solutions—Hierarchical Representation and Memory Architectures, Multi-Scale Temporal Modeling Architectures, and Temporal Reasoning approaches collectively account for over half the surveyed papers. The sibling paper in this leaf (H2VU Benchmark) also addresses hierarchical video understanding but differs in design philosophy. Neighboring leaves include Long-Form Video QA and Captioning Datasets (three papers) and Domain-Specific Long Video Benchmarks (one paper), indicating that general-purpose multi-timescale evaluation remains less developed than domain-specific or single-scale benchmarks. The taxonomy's scope note explicitly distinguishes this leaf by requiring 'explicit multi-timescale evaluation design,' separating it from general long-video datasets.

Among 30 candidates examined, none clearly refute the three core contributions. The within-content multi-timescale design examined 10 candidates with zero refutable matches, suggesting this specific evaluation methodology is novel within the limited search scope. The U-shaped performance trend finding also showed no refutations across 10 candidates, though this empirical observation depends on which models were tested. The insights on visual token allocation similarly faced no direct prior work among 10 examined candidates. These statistics indicate that within the top-30 semantically similar papers, the benchmark's design choices and empirical findings appear distinctive, though exhaustive literature review might reveal additional related work.

The analysis covers a focused sample of 30 papers from semantic search, not a comprehensive survey of all video understanding benchmarks. The sparse population of the Multi-Timescale Evaluation Benchmarks leaf and absence of refuting candidates suggest the work occupies a relatively unexplored niche. However, the limited search scope means potentially relevant benchmarks from adjacent communities (e.g., video summarization, temporal action localization) may not have been fully examined. The contribution appears novel within the surveyed context, though broader validation would strengthen confidence.