Research map / Video / Video-language understanding
SegMo: Co-Designing Content-Aware Sparsity and Locally-Cohesive Segment Parallelism for Efficient VLM Inference
Haojuan Li, Ruohan Tang, Dongzhou Cheng, Zongpu Zhang, Jian Li, Jiaqi Wang
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 uniform frame sampling with Content-Aware Sparsification | Qwen2-VL-7B-Instruct (32 frames) · LVBench | accuracy: 32.44 -> 44.44 | Improved |
| replace uniform frame sampling with Content-Aware Sparsification | MiniCPM-o 2.6 (32 frames) · LVBench | accuracy: 34.92 -> 45.84 | Improved |
| add Locally-Cohesive Segment Parallelism on top of CAS (ablation) | SegMo (Qwen2-VL-7B-Instruct, 32 frames) · LVBench | accuracy: 44.44 -> 40.75 | Hurt |
| add CAS + LSP versus 2-GPU data-parallel baseline | Qwen2-VL-7B-Instruct · LongVideoBench (32 frames) | TTFT (s): 2.490 -> 0.701 | Improved |
| add CAS + LSP versus 2-GPU data-parallel baseline | Qwen2-VL-7B-Instruct · LVBench (64 frames) | TTFT (s): 9.460 -> 2.762 | Improved |
| add LSP on top of uniform baseline at 64 frames on Video-MME long videos | MiniCPM-o 2.6 (64 frames) · Video-MME long (30-60m) | accuracy: 59.00 -> 59.00 | No clear effect |
| balanced weight w=0.5 instead of w=0.0 (redundancy only) (ablation) | CAS · LongVideoBench (64 frames) | accuracy: 49.65 -> 51.83 | Improved |
| balanced weight w=0.5 instead of w=1.0 (relevance only) (ablation) | CAS · Video-MME (128 frames) | accuracy: 66.67 -> 69.00 | Improved |
| relevance only (w=1) instead of redundancy only (w=0) (ablation) | CAS · LongVideoBench vs Video-MME | accuracy: 49.65 -> 50.58 (LongVideoBench); 68.05 -> 66.67 (Video-MME) | Mixed |
| add Global Context Injection (ablation) | SegMo · LongVideoBench overall | accuracy: 48.46 -> 49.83 | Improved |
| add Global Context Injection (ablation) | SegMo · LongVideoBench T2A tasks | accuracy: 41.07 -> 51.78 | Improved |
Closest papers by meaning
- V-CAST: Video Curvature-Aware Spatio-Temporal Pruning for Efficient Video Large Language Models (NeurIPS 2026 · similarity 0.91)
- FastVID: Dynamic Density Pruning for Fast Video Large Language Models (NeurIPS 2025 · similarity 0.89)
- A Stitch in Time Saves Nine: Small VLM is a Precise Guidance for Accelerating Large VLMs (CVPR 2025 · similarity 0.88)
- SparseVLM: Visual Token Sparsification for Efficient Vision-Language Model Inference (ICML 2025 · similarity 0.88)
- ApET: Approximation-Error Guided Token Compression for Efficient VLMs (CVPR 2026 · similarity 0.87)
- VidChain: Chain-of-Tasks with Metric-based Direct Preference Optimization for Dense Video Captioning (AAAI 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.
