Image de couverture de livre

Diagnosing and Repairing Unsafe Channels in Vision-Language Models via Causal Discovery and Dual-Modal Safety Subspace Projection

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Fu, Jinhu, Lou, Yihang, Si, Qingyi, Zhang, Shudong, Bai, Yan, Su, Sen
Format:	Preprint
Publié:	2026
Sujets:	Computer Vision and Pattern Recognition Artificial Intelligence
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

Learning to Edit Knowledge via Instruction-based Chain-of-Thought Prompting
par: Fu, Jinhu, et autres
Publié: (2026)

Bridging the Gap in Vision Language Models in Identifying Unsafe Concepts Across Modalities
par: Qu, Yiting, et autres
Publié: (2025)

Safe Inputs but Unsafe Output: Benchmarking Cross-modality Safety Alignment of Large Vision-Language Model
par: Wang, Siyin, et autres
Publié: (2024)

Safe Vision-Language Models via Unsafe Weights Manipulation
par: D'Incà, Moreno, et autres
Publié: (2025)

Safe Semantics, Unsafe Interpretations: Tackling Implicit Reasoning Safety in Large Vision-Language Models
par: Cai, Wei, et autres
Publié: (2025)

Diagnosing Causal Reasoning in Vision-Language Models via Structured Relevance Graphs
par: Pratama, Dhita Putri, et autres
Publié: (2026)

Selective Vision-Language Subspace Projection for Few-shot CLIP
par: Zhu, Xingyu, et autres
Publié: (2024)

Probing the Safety Response Boundary of Large Language Models via Unsafe Decoding Path Generation
par: Wang, Haoyu, et autres
Publié: (2024)

ClawSafety: "Safe" LLMs, Unsafe Agents
par: Wei, Bowen, et autres
Publié: (2026)

Annotating and Auditing the Safety Properties of Unsafe Rust
par: Rao, Zihao, et autres
Publié: (2025)

UnsafeChain: Enhancing Reasoning Model Safety via Hard Cases
par: Tomar, Raj Vardhan, et autres
Publié: (2025)

Dual-Use AI Face Swap Apps Are Mostly Unsafe: A Systematic Safety Audit
par: Daffalla, Alaa, et autres
Publié: (2026)

Beyond the Safety Tax: Mitigating Unsafe Text-to-Image Generation via External Safety Rectification
par: Meng, Xiangtao, et autres
Publié: (2025)

Hierarchical Dual-Subspace Decoupling for Continual Learning in Vision-Language Models
par: Qin, Mengxin, et autres
Publié: (2026)

Promoting Online Safety by Simulating Unsafe Conversations with LLMs
par: Hoffman, Owen, et autres
Publié: (2025)

Fearless Unsafe. A More User-friendly Document for Unsafe Rust Programming Base on Refined Safety Properties
par: Cui, Mohan, et autres
Publié: (2024)

Online Self-Calibration Against Hallucination in Vision-Language Models
par: Chen, Minghui, et autres
Publié: (2026)

Are Large Language Models Table-based Fact-Checkers?
par: Zhang, Hanwen, et autres
Publié: (2024)

EigenShield: Causal Subspace Filtering via Random Matrix Theory for Adversarially Robust Vision-Language Models
par: Darabi, Nastaran, et autres
Publié: (2025)

Breaking the Trade-Off Between Faithfulness and Expressiveness for Large Language Models
par: Yang, Chenxu, et autres
Publié: (2025)

DLEN: Dual Branch of Transformer for Low-Light Image Enhancement in Dual Domains
par: Xia, Junyu, et autres
Publié: (2025)

S-GRPO: Early Exit via Reinforcement Learning in Reasoning Models
par: Dai, Muzhi, et autres
Publié: (2025)

GuardTrace-VL: Detecting Unsafe Multimodel Reasoning via Iterative Safety Supervision
par: Xiang, Yuxiao, et autres
Publié: (2025)

Dialectics of Alignment: Harnessing Unsafe Knowledge for Dynamic Safety Routing
par: Hashemzadeh, Maryam, et autres
Publié: (2026)

Object Attribute Matters in Visual Question Answering
par: Li, Peize, et autres
Publié: (2023)

Multimodal Hypothetical Summary for Retrieval-based Multi-image Question Answering
par: Li, Peize, et autres
Publié: (2024)

Model Unlearning via Sparse Autoencoder Subspace Guided Projections
par: Wang, Xu, et autres
Publié: (2025)

Image Safeguarding: Reasoning with Conditional Vision Language Model and Obfuscating Unsafe Content Counterfactually
par: Bethany, Mazal, et autres
Publié: (2024)

Dual Prompt Learning for Adapting Vision-Language Models to Downstream Image-Text Retrieval
par: Wang, Yifan, et autres
Publié: (2025)

Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training
par: Yuan, Youliang, et autres
Publié: (2024)

iGSP:Implicit Gradient Subspace Projection for Efficient Continual Learning of Vision-Language Models
par: Cui, Xuezhi, et autres
Publié: (2026)

Adversary-Aware DPO: Enhancing Safety Alignment in Vision Language Models via Adversarial Training
par: Weng, Fenghua, et autres
Publié: (2025)

Sparse Models, Sparse Safety: Unsafe Routes in Mixture-of-Experts LLMs
par: Jiang, Yukun, et autres
Publié: (2026)

Asymmetric Visual Semantic Embedding Framework for Efficient Vision-Language Alignment
par: Liu, Yang, et autres
Publié: (2025)

Causality-based Cross-Modal Representation Learning for Vision-and-Language Navigation
par: Wang, Liuyi, et autres
Publié: (2024)

Towards Understanding Unsafe Video Generation
par: Pang, Yan, et autres
Publié: (2024)

MOSAIC: Module Discovery via Sparse Additive Identifiable Causal Learning for Scientific Time Series
par: Fan, Shicheng, et autres
Publié: (2026)

Aligning Information Capacity Between Vision and Language via Dense-to-Sparse Feature Distillation for Image-Text Matching
par: Liu, Yang, et autres
Publié: (2025)

Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models
par: Xu, Shicheng, et autres
Publié: (2024)

Subspace Alignment for Vision-Language Model Test-time Adaptation
par: Zeng, Zhichen, et autres
Publié: (2026)