Research map / Video / Video-language understanding
Selective Structured State-Spaces for Long-Form Video Understanding
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| replace uniform token processing with adaptive token selection using an S4-feature-guided mask generator | ViS4mer on LVU · LVU | accuracy across tasks: up to 9.6% | Helped |
| add S5 adaptive token selection | ViS4mer on COIN · COIN | accuracy: 88.41 -> 90.42 | Helped |
| add S5 adaptive token selection | ViS4mer on Breakfast · Breakfast | accuracy: 85.10 -> 90.14 | Helped |
| add LSMCL pretraining to S5 (ablation) | S5 on Breakfast · Breakfast | accuracy: 90.14 -> 90.70 | Helped |
| add LSMCL pretraining to S5 (ablation) | S5 on COIN · COIN | accuracy: 90.42 -> 90.81 | Helped |
| use S4 features rather than ViT features as the mask-generator input (ablation) | S5 on LVU · LVU | average improvement across nine tasks: +6.7 | Helped |
| replace adaptive selection with random masking (ablation) | S5 on LVU · LVU | accuracy across tasks | Hurt |
| increase the S5 masking ratio beyond 50% (ablation) | S5 on LVU · LVU | average improvement across nine tasks | Hurt |
| add another S5 block in place of an S4 block (ablation) | S5 pipeline on LVU · LVU | average improvement across nine tasks: 0.5% | Helped |
| increase LSMCL masking ratio beyond 50% (ablation) | S5 with LSMCL on LVU · LVU | average improvement across nine tasks | No clear effect |
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- FlexSelect: Flexible Token Selection for Efficient Long Video Understanding (NeurIPS 2025 · similarity 0.88)
- MovieChat: From Dense Token to Sparse Memory for Long Video Understanding (CVPR 2024 · similarity 0.87)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.87)
- A Simple Recipe for Contrastively Pre-training Video-First Encoders Beyond 16 Frames (CVPR 2024 · similarity 0.87)
- SEAL: Semantic Attention Learning for Long Video Representation (CVPR 2025 · similarity 0.87)
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.