---
title: "OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.25669v1"
permalink: "https://feed.astralor.com/items/itm_9020d75c6eb01083fa7c3589"
published: "2026-07-28T12:50:31.000Z"
score: 81
channel: "ai"
tags: ["OmniDelta","Qwen2.5-Omni","Token Compression","Omni-modal Large Language Model","Skill-Driven Budget Allocation"]
---

# OmniDelta: Skill-Driven Budget Allocation for Token Compression in OmniLLMs

## AI commentary

> This paper hits a key pain point in deploying OmniLLMs: the overhead of long token sequences. OmniDelta requires no additional training, yet significantly speeds up inference while maintaining accuracy by cleverly reallocating token budgets. The 'skill pool' idea is particularly noteworthy, incorporating query intent into allocation for more intelligent compression than attention-based methods. Highly relevant for engineering teams working on efficient multimodal deployment.

## At a glance

Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, ignoring the preceding budget-allocation problem. This paper proposes OmniDelta, a training-free, skill-driven framework that constructs audio and video skill pools for intent-aware inter-modal allocation and content-aware intra-modal allocation. On Qwen2.5-Omni-7B at 25% token retention, OmniDelta reduces GPU memory by 22% and achieves 1.64× end-to-end speedup, establishing a new accuracy-efficiency Pareto frontier.

## Key points

- First systematic study of the token budget allocation problem in OmniLLMs, showing that direct query-to-audio/video similarity is unreliable
- Proposes OmniDelta, a training-free framework that constructs skill pools for intent-aware inter-modal and content-aware intra-modal allocation
- On Qwen2.5-Omni-7B at 25% token retention, reduces GPU memory by 22% and achieves 1.64× end-to-end speedup, outperforming existing methods
- Method is compatible with existing pruning strategies, preserving the total retained-token ratio while reallocating budget

## Full text

Emerging Omni-modal Large Language Models (OmniLLMs) enable unified understanding of text, audio, and video, but their long audio-video token sequences introduce substantial memory and inference costs. Existing compression methods mainly focus on selecting important tokens under fixed budgets, leaving the preceding budget-allocation problem underexplored. We show that direct query-to-audio/video similarity is unreliable for inter-modal budget allocation, and that uniform intra-modal budgets can miss key evidence while retaining redundant content. To address these limitations, we propose OmniDelta, a training-free, skill-driven framework that couples intent-aware inter-modal allocation with content-aware intra-modal allocation. OmniDelta first constructs audio and video skill pools to shift the fixed retained-token budget according to query demand, then reallocates modality budgets over audio segments and video frames using local complexity and temporal redundancy. The resulting local budgets can be combined with existing pruning strategies, preserving the total retained-token ratio while changing where the budget is spent. Experiments on four audio-video benchmarks with two Qwen2.5-Omni models show that OmniDelta establishes a new accuracy-efficiency Pareto frontier across pruning ratios. At 25% token retention on Qwen2.5-Omni-7B, OmniDelta reduces GPU memory by 22.0% and achieves a 1.64x end-to-end speedup over full-token inference.

---

Verify against the original before citing; curated by Astral Feed: https://feed.astralor.com/items/itm_9020d75c6eb01083fa7c3589