OmniDrop: Layer-wise Token Pruning for Omni-modal LLMs via Query-Guidance

2026-05-14

AI Summary

This work addresses the challenge of excessive token counts in multimodal large language models caused by high-resolution audio-visual inputs, which hinders real-time inference and long-context understanding. The authors propose a training-free, layer-wise token pruning framework that operates within the LLM decoder rather than at the input stage. By leveraging textual queries to guide pruning, the method achieves task-adaptive and modality-agnostic compression. A novel temporal diversity score is introduced to preserve critical temporal context. This approach overcomes the limitations of conventional input-layer pruning strategies that rely on unreliable semantic proxies. Experiments demonstrate performance gains of up to 3.58 points across multiple audio-visual benchmarks, alongside reductions of up to 40% in prefill latency and 14.7% in memory consumption.

Abstract

Omni-modal large language models have demonstrated remarkable potential in holistic multimodal understanding; however, the token explosion caused by high-resolution audio and video inputs remains a critical bottleneck for real-time applications and long-form reasoning. Existing omni-modal token compression methods typically prune tokens at the input embedding level, relying on audio-video similarity or temporal co-occurrence as proxies for semantic relevance. In practice, such assumptions are often unreliable. To address this limitation, we propose OmniDrop, a training-free, layer-wise token pruning framework that progressively prunes audiovisual tokens within the LLM decoder layers rather than at the input-level, allowing early layers to preserve sufficient omni-modal information fusion before aggressively removing tokens in deeper layers. We further utilize text queries as guidance for modality-agnostic and task-adaptive token pruning. We also introduce a temporal diversity score that encourages balanced token survival to preserve global temporal context. Experimental results across various audiovisual benchmarks demonstrate that OmniDrop outperforms all baselines by up to 3.58 points while reducing prefill latency by up to 40% and memory usage by up to 14.7%.

Problem

• token explosion
• omni-modal LLMs
• real-time applications
• long-form reasoning
• multimodal understanding

Innovation

• layer-wise token pruning
• query-guided pruning
• omni-modal LLMs
• temporal diversity
• training-free compression