myWiki/concepts/native-streaming-ar-training.md

title, created, updated, type, tags, sources

title	created	updated	type	tags	sources
Native Streaming AR Training	2026-06-20	2026-06-20	concept	training autoregressive streaming causal	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 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：以模型自身退化历史为条件
• flow-matching loss
• audio-visual-representation-alignment 加速

参考

• maineCoon Section 3.1
• self-resampling
• autoregressive-video-generation
• wan-streamer — 端到端流式全双工交互中的原生流式训练