Aero Realtime:实现低延迟流式多模态生成的完全对齐输入输出流
文章背景与核心概要
现有的流式多模态模型大多受限于“先预填充(prefill)后解码(decode)”的轮次模式,导致模型无法在生成过程中自然地处理新的输入,即缺乏双工(duplex)能力。为了解决这一问题,研究人员提出了 Aero Realtime,这是一个 4B 参数规模的流式多模态模型,旨在实现真正的实时交互。
Aero Realtime 的核心创新在于引入了共享时间网格(shared temporal grid),将视频、音频和文本输出对齐到约 80 毫秒的时间槽中。这种设计使得输入和输出能够同步推进,通过单一的自回归目标函数同时学习“何时响应”和“生成什么内容”。该模型在 NVIDIA A6000 GPU 集群上表现优异,中位处理延迟仅为 84 毫秒,证明了完全对齐的输入输出建模在实现主动式、硬件高效的多模态交互方面的可行性。
Aero Realtime:实现低延迟流式多模态生成的完全对齐输入输出流
执行摘要
Aero Realtime is a novel 4B streaming multimodal model featuring a duplex architecture designed for low-latency, real-time generation. Unlike traditional streaming models that rely on rigid turn-based prefill-then-decode patterns or fragmented polling mechanisms, Aero Realtime aligns video, audio, and textual outputs onto a shared temporal grid. This allows inputs and outputs to advance synchronously, enabling a single autoregressive objective to handle response timing and content generation simultaneously. Operating on four NVIDIA A6000 workstation GPUs, the model achieves a median processing lag of 84 ms and a P95 lag of 173 ms over continuous video streams, proving the viability of fully aligned input-output modeling for proactive, hardware-efficient multimodal interactions.
Aero Realtime 是一款新型 4B 参数流式多模态模型,采用双工架构设计,专为低延迟实时生成而生。与依赖僵化的“先预填充后解码”模式或碎片化轮询机制的传统流式模型不同,Aero Realtime 将视频、音频和文本输出对齐到一个共享的时间网格上。这使得输入和输出能够同步推进,从而允许单一的自回归目标函数同时处理响应时机和内容生成。该模型在四块 NVIDIA A6000 工作站 GPU 上运行,在连续视频流中实现了 84 毫秒的中位处理延迟和 173 毫秒的 P95 延迟,证明了完全对齐的输入输出建模在主动式、硬件高效的多模态交互中的可行性。
论文元数据
- arXiv ID: arXiv:2608.08469 [cs.AI]
- Submitted: August 9, 2026
- Primary Subject: Artificial Intelligence (
cs.AI)- Authors:
- Kaichen Zhang
- Wei Huang
- Keming Wu
- Bo Li
- Xiaojuan Qi
- DOI: 10.48550/arXiv.2608.08469
- arXiv ID: arXiv:2608.08469 [cs.AI]
- 提交日期: 2026年8月9日
- 主要学科: 人工智能 (
cs.AI) - 作者:
- Kaichen Zhang
- Wei Huang
- Keming Wu
- Bo Li
- Xiaojuan Qi
- DOI: 10.48550/arXiv.2608.08469
摘要
Existing streaming multimodal models process observations incrementally but still follow a turn-based prefill-then-decode pattern, making them non-duplex: new observations cannot naturally enter an active generation stream. Proactive alternatives use micro-turn polling or external response gates, which fragment continuous interaction, decouple response timing from language generation, and complicate KV-cache-friendly serving.
We introduce Aero Realtime, a 4B streaming multimodal model with a duplex architecture for realtime generation. Aero Realtime aligns video, audio, and textual output on a shared temporal grid, where each approximately 80-ms audio slot predicts either a lexical token or a silence token. This allows input and output to advance together, enabling one autoregressive objective to learn both when to respond and what to generate.
During inference, Aero Realtime appends only the newest multimodal slot, carries forward the previous output state, and reuses the KV cache for efficient incremental execution. We further provide a complete training and serving recipe, including realtime QA construction, slot-aligned supervision, hardware-aware distributed training, and resumable inference.
On four NVIDIA A6000 workstation GPUs, Aero Realtime maintains 84-ms median and 173-ms P95 processing lag over 20 minutes of a continuously streamed video, remaining within 200 ms of the source timeline. These results demonstrate the feasibility of fully aligned input-output modeling for duplex, proactive, and hardware-aligned multimodal interaction.
现有的流式多模态模型虽然能增量处理观测数据,但仍遵循“先预填充后解码”的轮次模式,导致其不具备双工能力:新的观测数据无法自然地进入活跃的生成流中。主动式替代方案通常使用微轮询或外部响应门控,这会使连续交互碎片化,将响应时机与语言生成解耦,并使 KV 缓存友好的服务变得复杂。
我们引入了 Aero Realtime,这是一个具有双工架构的 4B 流式多模态模型,用于实时生成。Aero Realtime 将视频、音频和文本输出对齐在共享的时间网格上,其中每个约 80 毫秒的音频槽预测一个词汇标记或静音标记。这使得输入和输出能够共同推进,从而使单一的自回归目标能够同时学习何时响应以及生成什么内容。
在推理过程中,Aero Realtime 仅附加最新的多模态槽,延续之前的输出状态,并重用 KV 缓存以实现高效的增量执行。我们进一步提供了完整的训练和服务方案,包括实时问答构建、槽对齐监督、硬件感知分布式训练以及可恢复推理。
在四块 NVIDIA A6000 工作站 GPU 上,Aero Realtime 在 20 分钟的连续流式视频中保持了 84 毫秒的中位处理延迟和 173 毫秒的 P95 延迟,始终保持在源时间轴的 200 毫秒以内。这些结果证明了完全对齐的输入输出建模在双工、主动式和硬件对齐的多模态交互中的可行性。
关键贡献与架构亮点
- Duplex Architecture: Bypasses conventional turn-based constraints, permitting continuous, bidirectional flow of multimodal inputs and outputs.
- Shared Temporal Grid: Synchronizes video, audio, and text streams into ~80-ms slots, where each audio slot predicts lexical or silence tokens.
- Unified Autoregressive Objective: Learns both when to respond and what to generate within a single training objective.
- Efficient Inference & Serving: Leverages KV cache reuse by appending only the newest multimodal slot and carrying forward previous output states.
- Comprehensive Recipe: Includes realtime QA construction, slot-aligned supervision, hardware-aware distributed training, and resumable inference pipelines.
- 双工架构: 绕过了传统的轮次限制,允许多模态输入和输出的连续、双向流动。
- 共享时间网格: 将视频、音频和文本流同步为约 80 毫秒的时间槽,每个音频槽预测词汇或静音标记。
- 统一自回归目标: 在单一训练目标内同时学习“何时响应”和“生成什么”。
- 高效推理与服务: 通过仅附加最新的多模态槽并延续之前的输出状态,利用 KV 缓存重用机制。
- 完整方案: 包含实时问答构建、槽对齐监督、硬件感知分布式训练以及可恢复推理流水线。
访问与资源
- Full-Text Links: View PDF | HTML Version | TeX Source
- License: Creative Commons Attribution 4.0
- 全文链接: 查看 PDF | HTML 版本 | TeX 源码
- 许可协议: 知识共享署名 4.0
