SoC：面向测试时提示词调优的语义正交校准
SoC: Semantic Orthogonal Calibration for Test-Time Prompt Tuning
对做 VLM 测试时自适应的人，这篇给出了一个可直接替换全正交约束的 Huber 型正则项：在文本提示词嵌入的原型分离上做平滑约束，既保留语义近邻又避免过度自信，可插到任何 TPT 流程里改善置信度校准。
原文依据 we propose Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论随着视觉语言模型（VLM）在医疗、自动驾驶等关键决策系统中日益普及，其不确定性估计的校准变得至关重要。然而，在 VLM 测试时提示词调优（TPT）文献中，这一维度在很大程度上仍未被充分探索，该领域此前主要关注提升判别性能。近期最优方法主张对文本提示词嵌入对施加完全正交约束，以增强可分性，进而改善校准。然而，正如本文从理论上所证明的，完全正交约束所固有的梯度会强烈地将语义相关的类别推开，最终使模型过度自信。基于这一发现，我们提出语义正交校准（SoC），这是一种基于 Huber 的正则项，在保持语义邻近性的同时实现平滑的原型分离，从而相较以往基于正交性的方法改善校准。在全面的实证验证中，我们证明 SoC 能持续提升校准性能，同时保持有竞争力的判别能力。代码见 github.com/leofillioux/SoC。
英文原文摘要 1245 字符
With the increasing adoption of vision-language models (VLMs) in critical decision-making systems such as in healthcare or autonomous driving, the calibration of their uncertainty estimates has become paramount. Yet, this dimension has been largely underexplored in the VLM test-time prompt-tuning (TPT) literature, which has prioritized on improving their discriminative performance. Recent state-of-the-art methods advocate for enforcing full orthogonality over pairs of text prompt embeddings to enhance separability, and therefore calibration. Nevertheless, as we theoretically show in this work, the inherent gradients from fully orthogonal constraints will strongly push semantically related classes away, ultimately making the model overconfident. Based on our findings, we propose Semantic Orthogonal Calibration (SoC), a Huber-based regularizer that enforces smooth prototype separation while preserving semantic proximity, thereby improving calibration compared to prior orthogonality-based approaches. Across a comprehensive empirical validation, we demonstrate that SoC consistently improves calibration performance, while also maintaining competitive discriminative capabilities. Our code is available at github.com/leofillioux/SoC.