Research map / Video / Video-language understanding
State-Space Hierarchical Compression with Gated Attention and Learnable Sampling for Hour-Long Video Understanding in Large Multimodal Models
Geewook Kim, Minjoon Seo
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 |
|---|---|---|---|
| replace average-pooling queries with learnable gated patch aggregation (GPA) (ablation) | BIMBA-style bi-Mamba compressor · LVBench/MLVU/VideoMME average | average accuracy: 45.44 -> 52.22 | Improved |
| add time-axis aggregator (TAA) on top of GPA (ablation) | MambaMia · LVBench/MLVU/VideoMME average | average accuracy: 52.22 -> 53.56 | Improved |
| use Mamba V1 instead of V2 (ablation) | MambaMia · LVBench/MLVU/VideoMME average | average accuracy: 53.56 -> 53.13 | No clear effect |
| use bi-directional instead of uni-directional compressor blocks (ablation) | MambaMia · LVBench/MLVU/VideoMME average | average accuracy: 51.86 -> 53.56 | Improved |
| use delta-based frame sampling instead of uniform 50% sampling (ablation) | MambaMia final model · LVBench/MLVU/VideoMME average | average accuracy: 56.30 -> 56.95 | Improved |
| use delta-based frame sampling instead of uniform 50% sampling (ablation) | MambaMia final model · LVBench | accuracy: 43.38 -> 44.61 | Improved |
| add delta-based secondary sampling (halving tokens) versus none (ablation) | MambaMia final model · LVBench/MLVU/VideoMME average | average accuracy: 57.27 -> 56.95 | No clear effect |
| replace vanilla MLP projector with proposed compression | Pythia-2.8B LMM · LVBench/MLVU/VideoMME average | average accuracy: 32.78 -> 41.71 | Improved |
| replace vanilla MLP projector with proposed compression | Mamba2-2.7B LMM · LVBench/MLVU/VideoMME average | average accuracy: 34.14 -> 39.85 | Improved |
| use a Mamba2 LLM instead of Pythia with vanilla projector | Vanilla LMM · LVBench/MLVU/VideoMME average | average accuracy: 32.78 -> 34.14 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- VAMBA — Vamba: Understanding Hour-Long Videos with Hybrid Mamba-Transformers (ICCV 2025) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.90)
- LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding (ICML 2025 · similarity 0.88)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.87)
- Less Is More, but Where? Dynamic Token Compression via LLM-Guided Keyframe Prior (NeurIPS 2025 · similarity 0.87)
- CoRDS: Coreset-Based Representative and Diverse Selection for Streaming Video Understanding (NeurIPS 2026 · similarity 0.87)
- InfiniPot-V: Memory-Constrained KV Cache Compression for Streaming Video Understanding (NeurIPS 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.
