| title | OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models |
|---|---|
| authors | Jinsen Su, Yongdong Luo, Yuexiao Ma, Yibo Hu, Meiguang Jin, Xiaowu Zheng |
| date | 2026-07-28 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2607.23193 |
| upvotes | 6 (Hugging Face Daily Papers, 2026-11-01) |
| links | abs pdf hf code |
| abstract | Existing token compression methods for omnimodal large language models typically rely on one modality to determine what to retain in the other. We show that this assumption often breaks down: for the same query, audio and video relevance often peaks at different moments. This cross-modal salience mismatch makes unidirectional guidance prone to discarding answer-critical cues under aggressive compression. We propose OmniScope, a training-free token compression framework that uses the query as a shared semantic anchor while estimating relevance separately for audio and video. OmniScope allocates modality-specific token budgets, prunes visual tokens with an anchor-delta strategy that preserves both global context and temporal changes, and merges audio tokens within each second to reduce redundancy while maintaining temporal continuity. Across four audio-video benchmarks and two Qwen2.5-Omni model scales, OmniScope achieves the best average accuracy across all compression settings. At 25% overall token retention, it delivers up to 3.53x prefill speedup and more than 15% GPU memory reduction, with only a 0.35-point drop in average accuracy. These results suggest a simple design principle for OmniLLM inference: share the query across modalities, but not the salience estimates. The code is available at https://github.com/MAC-AutoML/OmniScope. |
cite
@misc{su2026omniscope,
title={OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models},
author={Jinsen Su and Yongdong Luo and Yuexiao Ma and Yibo Hu and Meiguang Jin and Xiaowu Zheng},
year={2026},
eprint={2607.23193},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2607.23193}
}