OmniScope: Modality-Decoupled Token Compression for Omnimodal Large Language Models
Audio and video tokens should be compressed separately because their relevant moments often do not line up.
OmniScope uses the query as a shared anchor, then estimates audio and video relevance independently. It assigns modality-specific token budgets, prunes visual tokens to keep context and temporal change, and merges audio tokens within each second. Across four audio-video benchmarks and two Qwen2.5-Omni scales, it reports the best average accuracy across compression settings. At 25% token retention, the paper reports up to 3.53x prefill speedup and more than 15% lower GPU memory use, with a 0.35-point average accuracy drop. HF Daily Papers' note
OmniScope uses the query as a shared anchor, then estimates audio and video relevance independently. It assigns modality-specific token budgets, prunes visual tokens to keep context and temporal change, and merges audio tokens within each second. Across four audio-video benchmarks and two Qwen2.5-Omni scales, it reports the best average accuracy across compression settings. At 25% token retention, the paper reports up to 3.53x prefill speedup and more than 15% lower GPU memory use, with a 0.35-point average accuracy drop. HF Daily Papers' note
score 5