Research map / Video / Video-language understanding
OmniZip: Audio-Guided Dynamic Token Compression for Fast Omnimodal Large Language Models
Keda Tao, Kele Shao, Bohan Yu, Weiqiang Wang, Jian Liu, Huan 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 |
|---|---|---|---|
| add audio-guided dynamic video pruning allocation (ablation) | OmniZip · WorldSense (45% retained) | avg score: 45.0 -> 45.9 | Improved |
| add audio anchor consolidation (ablation) | OmniZip · WorldSense (45% retained) | avg score: 44.8 -> 45.0 | Improved |
| use ISTC instead of DyCoke for video token compression (ablation) | OmniZip · WorldSense (45% retained) | avg score: 45.0 -> 45.9 | Improved |
| use window-wise selection instead of VisionZip global selection (ablation) | OmniZip · WorldSense (45% retained) | avg score: 44.2 -> 45.9 | Improved |
| use OmniZip attention-based audio selection instead of random audio selection (ablation) | OmniZip with ISTC · WorldSense (45% retained) | avg score: 45.1 -> 45.9 | Improved |
| apply OmniZip at 35% retention versus FastV and DyCoke at 50% retention | Qwen2.5-Omni-7B · WorldSense | avg score: 44.3 / 44.6 -> 45.3 | Improved |
| prune audio or video tokens too aggressively (ablation) | OmniZip · WorldSense sensitivity to rho_a and rho_v | avg score | Hurt |
| apply FastV-style attention-based pruning to omnimodal tokens | Qwen2.5-Omni-7B · AVUT / WorldSense | accuracy: FastV 94.3% vs OmniZip 99.1% of full at 45-50% retention | Hurt |
| use smaller model (3B vs 7B) under token compression | OmniZip · Omni QA average | relative performance: 99.1% (7B) vs 99.2% (3B) at 45% retention | No clear effect |
Later papers that compared with this method
4 later accepted papers report OmniZip as a baseline in their results tables; 4 of them report a better number on most of the results they share with it.
- OmniSIFT: Modality-Asymmetric Token Compression for Efficient Omni-modal Large Language Models (ICML 2026 · better on 6 of 6 shared results)
- AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression (NeurIPS 2026 · better on 6 of 6 shared results)
- Video-HolmesV2: Can MLLMs Reason with Spatio-Temporal Audio-Visual Evidence in Long Videos? (ECCV 2026 · better on 3 of 3 shared results)
- DASH: Dynamic Audio-Driven Semantic Chunking for Efficient Omnimodal Token Compression (ECCV 2026 · 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
- OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models (NeurIPS 2026 · similarity 0.94)
- AVOC: Enhancing Hour-Level Audio-Video Understanding in Omni-Modal LLMs via Retrieval-Inspired Token Compression (NeurIPS 2026 · similarity 0.90)
- DyCoke: Dynamic Compression of Tokens for Fast Video Large Language Models (CVPR 2025 · similarity 0.89)
- ALMTokenizer: A Low-bitrate and Semantic-rich Audio Codec Tokenizer for Audio Language Modeling (ICML 2025 · similarity 0.89)
- Omni-Interactive Universal Embedder (NeurIPS 2026 · similarity 0.88)
- One Token per Highly Selective Frame: Towards Extreme Compression for Long Video Understanding (NeurIPS 2025 · similarity 0.88)
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.
