通用视觉自监督研究报
A Daily Digest of Visual Self-Supervised Learning
2026-06-11 图像表征 · VFM · JEPA · 视频预训练
arXiv recent-list补漏 · P3 · 2026-06-11

Thinking Without Images：它和通用视觉自监督的关系在于：通过 on-policy self-distillation 内化局部视觉证据查看过程，和 VLM 视觉推理/自蒸馏有关

中相关；详见方法、贡献和实验边界。

Thinking Without Images: Internalizing Visual Manipulation with On-Policy Self-Distillation　arXiv recent-list补漏　原文链接

编号2606.08719 优先级P3 类别arXiv recent-list补漏 会议arXiv recent-list补漏 方法通过 on-policy self-distillation 内化局部视觉证据查看过程，和 VLM 视觉推理/自蒸馏有关 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：通过 on-policy self-distillation 内化局部视觉证据查看过程，和 VLM 视觉推理/自蒸馏有关。 中相关；详见方法、贡献和实验边界。

Figureure 2 · : A concrete training instance of Imagine-OPD
Figureure 2 · : A concrete training instance of Imagine-OPDFigure 2: A concrete training instance of Imagine-OPD. Given an original image and a question, the student model generates its own imagination trace. A teacher then evaluates the same trajectory with privileged cropped evidence views and answer, providing on-policy supervision through a token-level KL objective.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Thinking Without Images 的方法或实验，请结合正文精读段落一起看。
Figureure 1 · : From “Thinking with Images” to “Thinking with Imagination”
Figureure 1 · : From “Thinking with Images” to “Thinking with Imagination”Figure 1: From “Thinking with Images” to “Thinking with Imagination”. “Thinking with Images” interleaves text reasoning with explicit tool calls and cropped images, improving access to local evidence but incurring repeated tool interaction and visual encoding. Thinking with Imagination replaces these intermediate images with an internal imagination process that can carry forward the same fine-grained visual evidence without actually invoking tools.这张图来自论文 PDF 的结构化抽取。当前用于辅助理解 Thinking Without Images 的方法或实验，请结合正文精读段落一起看。

核心问题

它和通用视觉自监督的关系在于：通过 on-policy self-distillation 内化局部视觉证据查看过程，和 VLM 视觉推理/自蒸馏有关。

方法拆解

通过 on-policy self-distillation 内化局部视觉证据查看过程，和 VLM 视觉推理/自蒸馏有关

主要贡献

中相关；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。