PDF 摘要：中文翻译尚未生成，请点击“显示原文依据”查看英文内容。

Vision-language models for autonomous driving (AD-VLMs) have been widely adopted to integrate visual perception, language reasoning, and decision-making, yet they still pose serious safety risks under perturbations, misleading instructions, and long-tail traffic interactions. Existing safeguards have shown effectiveness in mitigating unsafe behaviors. However, these safeguards usually introduce external constraints explicitly, making the repair dependent on limited predefined rules or prompt regeneration, which weakens generalization to unseen scenarios and increases generation latency. To address these, we propose LATENTGUARD, a latent world logic repair framework that models safety logic inside the hidden representation space of AD-VLMs. Specifically, LATENTGUARDlearns logic-induced feature shifts from unsafe outputs and their counterfactual safe alternatives, enabling the model to repair unsafe hidden states toward safer world-consistent representations. To further improve generalization, we introduce conditional latent logic modeling and future-consistency regularization, which encourage repaired states to follow safer latent world evolution beyond the current output. Experiments on VRU-Accident and NA VSIM show that LATENTGUARDreduces unsafe response rate from 32.6% to 18.8% and improves PDMS from 82.3 to 84.6. Closed-loop and human-in-the-loop case studies further validate that LATENTGUARDcan repair unsafe AD-VLMs responses in realistic driving interactions while preserving driving behavior.

全文第 2 页依据：中文翻译尚未生成，请点击“显示原文依据”查看英文内容。

Ourcontributionsare summarized as: • We formulate AD-VLMs safety repair as alatent world logic repairproblem, moving beyond explicit rule enumeration, prompt regeneration, and output correction. • We propose LATENTGUARD, a latent world logic repair framework that learns logic-induced feature shifts, regenerates hidden-state repair signals, and regularizes them with future consistency. • Comprehensive experiments on accident-centric scenarios, common driving scenarios, and physical-world case studies demonstrate the effectiveness and generalization.

全文第 1 页依据：中文翻译尚未生成，请点击“显示原文依据”查看英文内容。

To address these, we propose LA- TENTGUARD, a latent world logic repair framework that models safety logic inside the hidden representation space of AD-VLMs.

全文第 1 页依据：中文翻译尚未生成，请点击“显示原文依据”查看英文内容。

To address these, we propose LA- TENTGUARD, a latent world logic repair framework that models safety logic inside the hidden representation space of AD-VLMs.

全文第 2 页依据：中文翻译尚未生成，请点击“显示原文依据”查看英文内容。

We evaluate LATENTGUARDon VRU- Accident and NA VSIM, where LATENTGUARDreduces unsafe response rate from 32.6% to 18.8% and improves PDMS from 82.3 to 84.6.