Visual SSL / representation · P2 · 2026-06-06

Cosine Misleads：它和通用视觉自监督的关系在于：诊断 latent visual reasoning 中 cosine/MSE 对齐指标可能反向误导，提醒看 load-bearing latent 而非表面相似度

中高相关；详见方法、贡献和实验边界。

Cosine Misleads: Auxiliary Losses Reshape Vision Language Models, Not Their Latents　arXiv　原文链接

编号2606.05753 优先级P2 类别Visual SSL / representation 会议arXiv 方法诊断 latent visual reasoning 中 cosine/MSE 对齐指标可能反向误导，提醒看 load-bearing latent 而非表面相似度来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：诊断 latent visual reasoning 中 cosine/MSE 对齐指标可能反向误导，提醒看 load-bearing latent 而非表面相似度。中高相关；详见方法、贡献和实验边界。

**Figureure 2 · : PRISM overview**Figure 2: PRISM overview. Two inference-time diagnostics for the LVR family. Axis 1 trains linear probes (Alain and Bengio, 2018) at two positions in the model: (a) the answer-decoding state the LM head reads, and (b) the feedback variable the autoregressive loop re-injects. We report the decodability gap $G = a c c _ { p r o b e } ( a ) - a c c _ { p r o b e } ( b )$ , which summarizes how much more answer-decodable the post-latent state is than the latent. Axis 2 perturbs the LVR-injected hidden states (latents) in generation (truncation, noise, swap) and measures the change in accuracy; small |∆acc| means the latent is bypassed. Section 5 shows the two axes are tied: G predicts each variant’s response to latent perturbation.这张图概括 Cosine Misleads 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。

**Cosine misleads: r = 0.94 across variants Figure 1: Cosine misleads**Cosine misleads: r = 0.94 across variants Figure 1: Cosine misleads. Cosine alignment between LVR-position hidden states and their teacher-forced visual targets is negatively correlated with V∗Bench accuracy across all five trained variants (Pearson r=−0.94). Progressive variants (P-LVR-2, P-LVR-3) reach the highest cosine but the lowest accuracy.这张图/表用于判断 Cosine Misleads 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：诊断 latent visual reasoning 中 cosine/MSE 对齐指标可能反向误导，提醒看 load-bearing latent 而非表面相似度。

方法拆解

诊断 latent visual reasoning 中 cosine/MSE 对齐指标可能反向误导，提醒看 load-bearing latent 而非表面相似度

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。