# Deferred Audio Pruning with Local Audio-Visual Dynamics for Omni-LLMs

arXiv: https://arxiv.org/abs/2608.08794 · HTML: https://mlsanity.com/paper/2608.08794/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Audio and omni-modal token compression, one of 12 directions in Efficient LLM inference and serving.
- Builds on 17 of its 32 references that are in the map.
- First posted on arXiv 9 Aug 2026.

## Placed in

- Efficient LLM inference and serving / [Audio and omni-modal token compression](https://mlsanity.com/inference/audio-and-omni-modal-token-compression/)

## Builds on (most cited first, 17 of 17)

- [LLaVA-OneVision: Easy Visual Task Transfer](https://mlsanity.com/paper/2408.03326/) (2024)
- [Sinkhorn Distances: Lightspeed Computation of Optimal Transportation Distances](https://mlsanity.com/paper/1306.0895/) (2013)
- [Similarity of Neural Network Representations Revisited](https://mlsanity.com/paper/1905.00414/) (2019)
- [Video-MME: The First-Ever Comprehensive Evaluation Benchmark of Multi-modal LLMs in Video Analysis](https://mlsanity.com/paper/2405.21075/) (2024)
- [LongVU: Spatiotemporal Adaptive Compression for Long Video-Language Understanding](https://mlsanity.com/paper/2410.17434/) (2024)
- [LongVILA: Scaling Long-Context Visual Language Models for Long Videos](https://mlsanity.com/paper/2408.10188/) (2024)
- [Objects that Sound](https://mlsanity.com/paper/1712.06651/) (2017)
- [VideoChat-Flash: Hierarchical Compression for Long-Context Video Modeling](https://mlsanity.com/paper/2501.00574/) (2024)
- [WorldSense: Evaluating Real-world Omnimodal Understanding for Multimodal LLMs](https://mlsanity.com/paper/2502.04326/) (2025)
- [Baichuan-Omni-1.5 Technical Report](https://mlsanity.com/paper/2501.15368/) (2025)
- [Daily-Omni: Towards Audio-Visual Reasoning with Temporal Alignment across Modalities](https://mlsanity.com/paper/2505.17862/) (2025)
- [Audio-centric Video Understanding Benchmark without Text Shortcut](https://mlsanity.com/paper/2503.19951/) (2025)
- [AVE-CLIP: AudioCLIP-based Multi-window Temporal Transformer for Audio Visual Event Localization](https://mlsanity.com/paper/2210.05060/) (2022)
- [UniComp: Rethinking Video Compression Through Informational Uniqueness](https://mlsanity.com/paper/2512.03575/) (2025)
- [EchoingPixels: Aliasing-Resistant Joint Token Reduction for Audio-Visual LLMs](https://mlsanity.com/paper/2512.10324/) (2025)
- [OmniSelect: Dynamic Modality-Aware Token Compression for Efficient Omni-modal Large Language Models](https://mlsanity.com/paper/2605.18041/) (2026)
- [Keep What Audio Cannot Say: Context-Preserving Token Pruning for Omni-LLMs](https://mlsanity.com/paper/2605.11605/) (2026)

## Built on by (most cited first, 2 of 2)

- [Omni-Streaming Thinking](https://mlsanity.com/paper/2609.15128/) (2026)
- [Learning to Reason with Compressed Context: Ground-Truth-Free Adaptation of OmniLLMs via Self-Distillation](https://mlsanity.com/paper/2609.39953/) (2026)

## Abstract (arXiv)

Omni-modal LLMs jointly process audio, video, and text, but long multimodal sequences incur substantial prefill and KV-cache costs. Existing omni-modal compression methods primarily focus on pre-LLM token reduction, leaving modality-specific compression across the LLM boundary underexplored. We propose A-PACK, a two-stage framework that defers audio pruning until query-conditioned multimodal interactions emerge. Our analysis shows that audio exhibits higher task-relevant information density and representational diversity per token than video. We further find that local audio-visual dynamics provide a more effective cue for visual selection than token-wise matching. We therefore preserve audio and compress video with local dynamics before the LLM, then progressively prune low-relevance audio and visual tokens and their KV-cache entries inside the LLM. Across four benchmarks on Qwen2.5-Omni-7B/3B, A-PACK achieves the strongest average performance among the evaluated prior methods while reducing prefill FLOPs by up to 78% and improving decoding throughput by up to 2.21x.