Research map / Architectures and efficiency / Attention and transformers
OmniSparse: Training-Aware Fine-Grained Sparse Attention for Long-Video MLLMs
Feng Chen, Yefei He, Shaoxuan He, Yuanyu He, Jing Liu, Lequan Lin, Akide Liu, Zhaoyang Li, Jiyuan Zhang, Zhenbang Sun, Bohan Zhuang, Qi Wu
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 |
|---|---|---|---|
| train with OmniSparse instead of full attention | LLaVA-Video-7B · ActivityNet-QA | accuracy: 57.4 -> 57.6 | No clear effect |
| train with OmniSparse instead of MOBA | LLaVA-Video-7B 256k · ActivityNet-QA | accuracy: 55.4 -> 57.6 | Improved |
| scale training context from 256k to 1M | OmniSparse · ActivityNet-QA | accuracy: 57.6 -> 58.2 | Improved |
| use OmniSparse in place of FastV (training-free) | LLaVA-Video-7b · average of 4 benchmarks | average score: 59.9 -> 60.8 | Improved |
| use OmniSparse in place of VisionZip | LLaVA-Video-7b · average of 4 benchmarks | average score: 42.5 -> 60.8 | Improved |
| add query selection (ablation) | OmniSparse · VideoMME | accuracy / FLOPs reduction: FLOPs -54.4%, VideoMME 64.7 unchanged | Improved |
| add KV selection (ablation) | OmniSparse · VideoMME | accuracy / FLOPs reduction: FLOPs -51.5%, VideoMME 64.7 unchanged | Improved |
| combine query and KV selection (ablation) | OmniSparse · VideoMME | FLOPs reduction: 54.4% -> 77.9% | Improved |
| combine KV pruning and regrouping for cache (ablation) | OmniSparse · VideoMME | cache reduction: 29.7% / 51.5% -> 64.1% | Improved |
| keep first-head queries (rather than remove) (ablation) | OmniSparse · VideoMME | accuracy: 0.4% drop if removed, 3% extra FLOPs saving | Improved |
| use OmniSparse instead of FlashAttention | LLaVA-Video-7b · 64k input | TTFT: 15.45 -> 10.06 | Improved |
| train with sparse attention (training-aware) vs training-free | OmniSparse · long-video benchmarks | FLOPs and memory reduction: additional 13.1% FLOPs and 13.4% memory reduction | Improved |
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.
- MOBA — MoBA: Mixture of Block Attention for Long-Context LLMs (NeurIPS 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
- Bidirectional Sparse Attention for Faster Video Diffusion Training (NeurIPS 2026 · similarity 0.87)
- Full Attention Strikes Back: Transferring Full Attention into Sparse within Hundred Training Steps (NeurIPS 2026 · similarity 0.87)
- SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference (NeurIPS 2026 · similarity 0.87)
- Faster Video Diffusion with Trainable Sparse Attention (NeurIPS 2025 · similarity 0.86)
- Attention-aware Inference Optimizations for Large Vision-Language Models with Memory-efficient Decoding (CVPR 2026 · similarity 0.86)
- MMInference: Accelerating Pre-filling for Long-Context Visual Language Models via Modality-Aware Permutation Sparse Attention (ICML 2025 · similarity 0.86)
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.
