Stage-adaptive Token Selection for Efficient Omni-modal LLMs

Researchers introduced SEATS, a training-free method for efficient inference in omni-modal LLMs.

SEATS prunes redundant visual and audio tokens before the model, then removes more tokens inside the LLM using layer-wise relevance patterns.

On Qwen2.5-Omni and Qwen3-Omni, it cut FLOPs by 9.3x and sped up prefill by 4.8x while retaining 96.3% of baseline performance.

The work suggests multimodal inference can be made much cheaper by adapting token selection across stages without retraining.