Research map / Video / Video-language understanding
A Training-Free Framework for Long Video Understanding via Video-Query-Options Similarity
Zhirong Wu, Xiaodong Wang, Langling Huang, Teng Xu, Peixi Peng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add top-N frames retrieval (ablation) | LLaVA-Video-7B / Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +4.3 and +2.0 | Improved |
| add top-k segments uniform sampling (ablation) | LLaVA-Video-7B / Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +0.2 and +0.0 | No clear effect |
| add adaptive frame sampling (AFS) (ablation) | LLaVA-Video-7B / Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +0.7 and +0.8 | Improved |
| add dynamic resolution allocation (DRA) (ablation) | Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +2.1 | Improved |
| add generated options for similarity (ablation) | LLaVA-Video-7B / Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +0.7 and +0.4 | Improved |
| add provided multiple-choice options for similarity (ablation) | LLaVA-Video-7B / Qwen2.5-VL-7B · 5 benchmarks (Table 2) | avg gain: +1.5 and +1.2 | Improved |
| use larger PE-G/14 retriever instead of CLIP-B/32 (ablation) | Ours · LVBench | accuracy: 49.8 -> 54.2 | Improved |
| use 8s segments instead of single images (length 0) (ablation) | Ours LLaVA-Video-7B · LVBench | accuracy: 51.4 -> 54.2 | Improved |
| use 64s segments instead of 8s (ablation) | Ours LLaVA-Video-7B · LVBench | accuracy: 54.2 -> 47.4 | Hurt |
| generate options for 3 rounds instead of question only (ablation) | Ours LLaVA-Video-7B · LVBench | accuracy: 50.2 -> 51.3 | Improved |
| generate options for 4 rounds instead of 3 (ablation) | Ours LLaVA-Video-7B · VideoEval-Pro MCQ | accuracy: 55.3% -> 55.1% | Hurt |
| increase option-generation rounds from 1 to 4 (ablation) | Ours LLaVA-Video-7B · VideoEval-Pro | OCA: 54.6% -> 63.8% | Improved |
Closest papers by meaning
- BOLT: Boost Large Vision-Language Model Without Training for Long-form Video Understanding (CVPR 2025 · similarity 0.92)
- Adaptive Keyframe Sampling for Long Video Understanding (CVPR 2025 · similarity 0.92)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.91)
- Streaming Long Video Understanding with Large Language Models (NeurIPS 2024 · similarity 0.91)
- MA-LMM: Memory-Augmented Large Multimodal Model for Long-Term Video Understanding (CVPR 2024 · similarity 0.91)
- Needle In A Video Haystack: A Scalable Synthetic Evaluator for Video MLLMs (ICLR 2025 · similarity 0.91)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
