通用视觉自监督研究报
A Daily Digest of Visual Self-Supervised Learning
2026-05-29 图像表征 · VFM · JEPA · 视频预训练
Visual SSL / representation · P0 · 2026-05-29

Which Pretraining Paradigm Better Serves：它和通用视觉自监督的关系在于：用 frozen-feature probing 比较 VLM 与 VGM 表征，直接回答语义与几何信息分别从哪类预训练来

高相关；详见方法、贡献和实验边界。

Which Pretraining Paradigm Better Serves Spatial Intelligence? An Empirical Comparison of Vision-Language and Video Generation Models　arXiv　原文链接

编号2605.28132 优先级P0 类别Visual SSL / representation 会议arXiv 方法用 frozen-feature probing 比较 VLM 与 VGM 表征，直接回答语义与几何信息分别从哪类预训练来 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：用 frozen-feature probing 比较 VLM 与 VGM 表征，直接回答语义与几何信息分别从哪类预训练来。 高相关；详见方法、贡献和实验边界。

Figureure 2 · : Overview of our probing framework, where (A) is semantic tagging, (B
Figureure 2 · : Overview of our probing framework, where (A) is semantic tagging, (BFigure 2: Overview of our probing framework, where (A) is semantic tagging, (B) is instance grouping, and (C) is geometry prediction. We freeze each VLM or VGM, extract temporally aligned video features, and train lightweight probes with an identical backbone architecture and task-specific heads. It is notable that the probing backbone is unified in architecture, but the three task probes are trained separately.这张图概括 Which Pretraining Paradigm Better Serves 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。
Figureure 1 · : Top: Frozen VLM and VGM features are probed on three axes for spatia
Figureure 1 · : Top: Frozen VLM and VGM features are probed on three axes for spatiaFigure 1: Top: Frozen VLM and VGM features are probed on three axes for spatial intelligence. Bottom: Results show that VLMs excel at semantics and instances, VGMs excel at geometry, and simple fusion combines their strengths, further suggesting that the two representation families are complementary.这张图/表用于判断 Which Pretraining Paradigm Better Serves 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。

核心问题

它和通用视觉自监督的关系在于：用 frozen-feature probing 比较 VLM 与 VGM 表征，直接回答语义与几何信息分别从哪类预训练来。

方法拆解

用 frozen-feature probing 比较 VLM 与 VGM 表征，直接回答语义与几何信息分别从哪类预训练来

主要贡献

高相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。