OmniSIFT: Modality-Asymmetric Token Compression for Efficient Omni-modal Large Language Models

TL;DR

OmniSIFT is a novel token compression framework for Omni-modal LLMs that reduces computational load by selectively pruning video and audio tokens, maintaining high performance with fewer parameters.

Contribution

It introduces a two-stage, modality-asymmetric token compression method optimized end-to-end, significantly improving efficiency while preserving or enhancing model performance.

Findings

OmniSIFT introduces only 4.85M parameters for Qwen2.5-Omni-7B.

It outperforms all compression baselines with 25% of original tokens.

OmniSIFT maintains lower latency than training-free baselines.

Abstract

Omni-modal Large Language Models (Omni-LLMs) have demonstrated strong capabilities in audio-video understanding tasks. However, their reliance on long multimodal token sequences leads to substantial computational overhead. Despite this challenge, token compression methods designed for Omni-LLMs remain limited. To bridge this gap, we propose OmniSIFT (Omni-modal Spatio-temporal Informed Fine-grained Token compression), a modality-asymmetric token compression framework tailored for Omni-LLMs. Specifically, OmniSIFT adopts a two-stage compression strategy: (i) a spatio-temporal video pruning module that removes video redundancy arising from both intra-frame structure and inter-frame overlap, and (ii) a vision-guided audio selection module that filters audio tokens. The entire framework is optimized end-to-end via a differentiable straight-through estimator. Extensive experiments on five representative benchmarks demonstrate the efficacy and robustness of OmniSIFT. Notably, for Qwen2.5-Omni-7B, OmniSIFT introduces only 4.85M parameters while maintaining lower latency than training-free baselines such as OmniZip. With merely 25% of the original token context, OmniSIFT consistently outperforms all compression baselines and even surpasses the performance of the full-token model on several tasks.