arXiv replacement/update + TMLR accepted · P1 · 2026-06-19

Beyond the Linear Separability Ceiling：它和通用视觉自监督的关系在于：用线性可分性上界诊断 VLM 是感知失败还是推理失败，并用 contrastive objective 修正视觉 manifold

中高相关；详见方法、贡献和实验边界。

Beyond the Linear Separability Ceiling: Aligning Representations in VLMs　arXiv replacement/update + TMLR accepted　原文链接

编号2507.07574 优先级P1 类别arXiv replacement/update + TMLR accepted 会议arXiv replacement/update + TMLR accepted 方法用线性可分性上界诊断 VLM 是感知失败还是推理失败，并用 contrastive objective 修正视觉 manifold 来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：用线性可分性上界诊断 VLM 是感知失败还是推理失败，并用 contrastive objective 修正视觉 manifold。中高相关；详见方法、贡献和实验边界。

**Figureure 2 · : Illustration of the framework and its classifications, explaining PC**Figure 2: Illustration of the framework and its classifications, explaining PCA scenarios observed empirically. The LSC probe (dashed line) is the linear boundary between positive (cP ) and negative $\left( c _ { N } \right)$ centroids. (a) LSC probe: the query $( v _ { Q } )$ is correctly classified by the linear probe. (b) Alignment gap: classified when generative accuracy is on average not statistically superior to the LSC $\left( p \ge 0 . 0 5 \right)$ . The LSC probe succeeds, but the generative model fails. A non-linear model is expected to outperform its linear probe; failure to do so suggests a misalignment. (c) Surpassing the ceiling: classified when generative accuracy is on average statistically superior to the LSC $\left( p < 0 . 0 5 \right)$ . The LSC probe fails, but the generative model succeeds.这张图概括 Beyond the Linear Separability Ceiling 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。

**Figure A.3: Illustration of contrastive loss hyperparameter scan results for Pixtral**Figure A.3: Illustration of contrastive loss hyperparameter scan results for Pixtral. Plot displays model accuracy after every epoch (0-indexed) for selected hyperparameters.这张图/表用于判断 Beyond the Linear Separability Ceiling 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：用线性可分性上界诊断 VLM 是感知失败还是推理失败，并用 contrastive objective 修正视觉 manifold。

方法拆解

用线性可分性上界诊断 VLM 是感知失败还是推理失败，并用 contrastive objective 修正视觉 manifold

主要贡献

中高相关；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。