arXiv new; ECCV 2026; masked image modeling · P0 · 2026-07-01

ExPLoRe：它和通用视觉自监督的关系在于：把 token distillation、CLS alignment、pixel reconstruction 变成按 patch 自适应路由的 MIM 训练信号

高相关；详见方法、贡献和实验边界。

ExPLoRe: Expert Patch-Level Loss Routing for Multi-Objective Masked Image Modeling　arXiv new; ECCV 2026; masked image modeling　原文链接

编号2606.31201 优先级P0 类别arXiv new; ECCV 2026; masked image modeling 会议arXiv new; ECCV 2026; masked image modeling 方法把 token distillation、CLS alignment、pixel reconstruction 变成按 patch 自适应路由的 MIM 训练信号来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：把 token distillation、CLS alignment、pixel reconstruction 变成按 patch 自适应路由的 MIM 训练信号。高相关；详见方法、贡献和实验边界。

**Figureure 2 · : ExPLoRe Framework Overview**Fig. 2: ExPLoRe Framework Overview. Soft Mixture of Experts (Soft-MoE) is integrated into the student encoder for patch-level adaptive loss weighting. The student encoder (ViT-Base with alternating MoE blocks at layers {1,3,5,7,9,11}) processes patches while a frozen CLIP teacher provides semantic targets. Soft-MoE dispatch weights D serve as per-patch loss coeficients: each expert weights a diferent training objective. The two routing weight types play distinct roles: dispatch weights D (normalized over patches per expert) form the loss-coupling pathway that carries loss gradients back to the router, whereas combine weights C (normalized over experts per patch) only mix expert outputs in the forward pass. Loss-coupling, where loss gradients flow through D to the router, is the key mechanism enabling learned specialization.这张图概括 ExPLoRe 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。

**Fig**Fig. A2: Silhouette coeficient across MoE blocks. Dispatch-based cluster assignments on 200 ImageNet validation images (50K tokens subsampled). The coupled model develops strong expert specialization at Block 11 (the loss block), while the detach ablation collapses to zero specialization at this block. Early blocks show similar routing structure in both models.这张图/表用于判断 ExPLoRe 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：把 token distillation、CLS alignment、pixel reconstruction 变成按 patch 自适应路由的 MIM 训练信号。

方法拆解

把 token distillation、CLS alignment、pixel reconstruction 变成按 patch 自适应路由的 MIM 训练信号

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。