31 lines
1.2 KiB
Markdown
31 lines
1.2 KiB
Markdown
---
|
||
title: "Native Streaming AR Training"
|
||
created: 2026-06-20
|
||
updated: 2026-06-20
|
||
type: concept
|
||
tags: ["training", "autoregressive", "streaming", "causal"]
|
||
sources: ["https://arxiv.org/abs/2606.17800"]
|
||
---
|
||
|
||
# Native Streaming AR Training (原生流式自回归训练)
|
||
|
||
**Native Streaming AR Training** 是 [[maineCoon|MaineCoon]] 的核心训练范式：在训练和推理时使用**相同的因果逐块自回归 regime**，而非通过 teacher forcing 从非因果教师蒸馏。
|
||
|
||
## 核心原则
|
||
|
||
- **Chunk-by-chunk causal rollout**：每次预测一个 chunk，仅以已生成的 chunk 为条件
|
||
- **无 Teacher Forcing**：不从双向教师蒸馏流式行为——原生即流式
|
||
- **Train-inference matched**：训练和推理分布一致，消除 gap
|
||
|
||
## 关键组件
|
||
|
||
- [[self-resampling|Self-Resampling]]：以模型自身退化历史为条件
|
||
- [[flow-matching|Flow Matching]] loss
|
||
- [[audio-visual-representation-alignment|Cross-Modal Representation Alignment]] 加速
|
||
|
||
## 参考
|
||
- [[maineCoon|MaineCoon 论文]] Section 3.1
|
||
- [[self-resampling|Self-Resampling]]
|
||
- [[autoregressive-video-generation|自回归视频生成]]
|
||
- [[wan-streamer]] — 端到端流式全双工交互中的原生流式训练
|