MEGA Hub

Allocation Before Ranking: Decoupled Token Compression for OmniLLMs

Authors

Do you know Zhenghui Guo?You can claim authorship or link another user.Do you know Yilin Yang?You can claim authorship or link another user.Do you know Yuanbin Man?You can claim authorship or link another user.Do you know Miao Yin?You can claim authorship or link another user.Do you know Weidong Shi?You can claim authorship or link another user.Do you know Rabimba Karanjai?You can claim authorship or link another user.Do you know Omprakash Gnawali?You can claim authorship or link another user.Do you know Chengming Zhang?You can claim authorship or link another user.

Abstract

Token compression in OmniLLMs is typically posed as a single saliency-ranking problem: score each multimodal token, keep the top-K. We argue this abstraction is mis-specified. The same attention score simultaneously decides two things: how much retained capacity each modality receives, and which tokens within a modality are kept. A shared top-K rule therefore inherits this audio-favoring allocation prior, spending retained capacity on audio before video tokens have a chance to compete. We propose Macer, a training-free compressor that first assigns explicit audio and video budgets, then performs allocation-normalized ranking within each modality at modality-specific shallow layers. Macer significantly reduces token cost while preserving accuracy across audio-grounded, audio--video joint, visual-dominant, and video-centric benchmarks. At 25 % retention, Macer preserves 98.7 % of full-token performance on Qwen2.5-Omni-7B and 97.3 % on Qwen2.5-Omni-3B. On Qwen2.5-Omni-7B, this 25 % setting reaches OmniZip-level performance at 45 % retention while using lower FLOPs. On OmniVinci-9B, the same allocation-before-ranking principle improves over shared top-K ranking by up to 12.9 points.

Community

00