Macer: Training-Free Token Compression for OmniLLMs

A new arXiv paper (2608.01665) introduces Macer, a training-free token compression method for OmniLLMs that decouples allocation from ranking. The authors argue that current token compression methods, which use a single saliency-ranking to keep top-K tokens, are mis-specified because the same attention score determines both cross-modal capacity allocation and within-modality token selection. This shared top-K rule inherits an audio-favoring bias, allocating retained capacity to audio before video tokens can compete. Macer addresses this by first assigning explicit audio and video budgets, then performing allocation-normalized ranking within each modality at modality-specific shallow layers. The method 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-7B (likely a typo in the original, but the paper states these figures). The paper is published on arXiv and is relevant to the field of AI and multimodal model efficiency.