张祥雨：多模态研究、推理训练与在线学习

核心内容

本期由李广密参与主持，对话阶跃星辰首席科学家张祥雨，回顾视觉研究向语言与多模态模型演进的经历。发布方介绍特别提到模型规模、推理能力与训练目标之间的关系。

章节进一步涉及强化学习、思维链、视觉理解与生成、长上下文、多模型协作和在线学习。节目中的时间预测和技术解释应保留嘉宾与录制时期的语境，不能当作已成立的普遍结论。

Co-host Li Guangmi speaks with StepFun chief scientist Xiangyu Zhang about the path from computer vision to language and multimodal models. The publisher highlights questions about model scale, reasoning, and training objectives.

The chapters also cover reinforcement learning, chains of thought, visual understanding and generation, long context, model collaboration, and online learning. Forecasts and technical explanations should be read as the guest’s views in the episode’s original context, not as established general conclusions.

介绍依据 · 发布方说明 · 完整内容待核读

编辑导览，依据原始来源 ↗，整理于 2026-10-20。中英文均由本站概括整理。内容介绍与下方原话、主张的逐条核对分开标注。

仅索引这一场已经收录，还没有抽原话。