Image de couverture de livre

Analysing the Safety Pitfalls of Steering Vectors

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Li, Yuxiao, Fastowski, Alina, Zaradoukas, Efstratios, Prenkaj, Bardh, Kasneci, Gjergji
Format:	Preprint
Publié:	2026
Sujets:	Cryptography and Security Computation and Language
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

Injecting Falsehoods: Adversarial Man-in-the-Middle Attacks Undermining Factual Recall in LLMs
par: Fastowski, Alina, et autres
Publié: (2025)

From Confidence to Collapse in LLM Factual Robustness
par: Fastowski, Alina, et autres
Publié: (2025)

Reinforcement Unlearning via Group Relative Policy Optimization
par: Zaradoukas, Efstratios, et autres
Publié: (2026)

Graph Inverse Style Transfer for Counterfactual Explainability
par: Prenkaj, Bardh, et autres
Publié: (2025)

Understanding Knowledge Drift in LLMs through Misinformation
par: Fastowski, Alina, et autres
Publié: (2024)

RAZOR: Sharpening Knowledge by Cutting Bias with Unsupervised Text Rewriting
par: Yang, Shuo, et autres
Publié: (2024)

Not All Features Deserve Attention: Graph-Guided Dependency Learning for Tabular Data Generation with Language Models
par: Zhang, Zheyu, et autres
Publié: (2025)

CURE: Controlled Unlearning for Robust Embeddings -- Mitigating Conceptual Shortcuts in Pre-Trained Language Models
par: Kocak, Aysenur, et autres
Publié: (2025)

EvalxNLP: A Framework for Benchmarking Post-Hoc Explainability Methods on NLP Models
par: Dhaini, Mahdi, et autres
Publié: (2025)

Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control
par: Cyberey, Hannah, et autres
Publié: (2025)

Attention Mechanisms Don't Learn Additive Models: Rethinking Feature Importance for Transformers
par: Leemann, Tobias, et autres
Publié: (2024)

Probabilistic Aggregation and Targeted Embedding Optimization for Collective Moral Reasoning in Large Language Models
par: Yuan, Chenchen, et autres
Publié: (2025)

Compiling Activation Steering into Weights via Null-Space Constraints for Stealthy Backdoors
par: Yin, Rui, et autres
Publié: (2026)

Interpretable LLM Guardrails via Sparse Representation Steering
par: He, Zeqing, et autres
Publié: (2025)

AdaSteer: Your Aligned LLM is Inherently an Adaptive Jailbreak Defender
par: Zhao, Weixiang, et autres
Publié: (2025)

Trojan Activation Attack: Red-Teaming Large Language Models using Activation Steering for Safety-Alignment
par: Wang, Haoran, et autres
Publié: (2023)

Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety
par: Zhang, Yuyou, et autres
Publié: (2025)

Mitigating Fine-tuning based Jailbreak Attack with Backdoor Enhanced Safety Alignment
par: Wang, Jiongxiao, et autres
Publié: (2024)

EnchTable: Unified Safety Alignment Transfer in Fine-tuned Large Language Models
par: Wu, Jialin, et autres
Publié: (2025)

Cross-Task Defense: Instruction-Tuning LLMs for Content Safety
par: Fu, Yu, et autres
Publié: (2024)

Be Your Own Red Teamer: Safety Alignment via Self-Play and Reflective Experience Replay
par: Wang, Hao, et autres
Publié: (2026)

Model-Agnostic Lifelong LLM Safety via Externalized Attack-Defense Co-Evolution
par: Zhang, Xiaozhe, et autres
Publié: (2026)

Moral Lenses, Political Coordinates: Towards Ideological Positioning of Morally Conditioned LLMs
par: Yuan, Chenchen, et autres
Publié: (2026)

PandaGuard: Systematic Evaluation of LLM Safety against Jailbreaking Attacks
par: Shen, Guobin, et autres
Publié: (2025)

S-Eval: Towards Automated and Comprehensive Safety Evaluation for Large Language Models
par: Yuan, Xiaohan, et autres
Publié: (2024)

Logit-Gap Steering: A Forward-Pass Diagnostic for Alignment Robustness
par: Li, Tung-Ling, et autres
Publié: (2025)

Overriding Safety protections of Open-source Models
par: Kumar, Sachin
Publié: (2024)

ML-Bench&Guard: Policy-Grounded Multilingual Safety Benchmark and Guardrail for Large Language Models
par: Zhao, Yunhan, et autres
Publié: (2026)

Bag of Tricks for Subverting Reasoning-based Safety Guardrails
par: Chen, Shuo, et autres
Publié: (2025)

SafeAligner: Safety Alignment against Jailbreak Attacks via Response Disparity Guidance
par: Huang, Caishuang, et autres
Publié: (2024)

Root Defence Strategies: Ensuring Safety of LLM at the Decoding Level
par: Zeng, Xinyi, et autres
Publié: (2024)

Fingerprint Vector: Enabling Scalable and Efficient Model Fingerprint Transfer via Vector Addition
par: Xu, Zhenhua, et autres
Publié: (2024)

RAG Safety: Exploring Knowledge Poisoning Attacks to Retrieval-Augmented Generation
par: Zhao, Tianzhe, et autres
Publié: (2025)

TWGuard: A Case Study of LLM Safety Guardrails for Localized Linguistic Contexts
par: Chu, Hua-Rong, et autres
Publié: (2026)

One Word at a Time: Incremental Completion Decomposition Breaks LLM Safety
par: Arif, Samee, et autres
Publié: (2026)

Understanding and Mitigating Over-refusal for Large Language Models via Safety Representation
par: Zhang, Junbo, et autres
Publié: (2025)

Overlooked Safety Vulnerability in LLMs: Malicious Intelligent Optimization Algorithm Request and its Jailbreak
par: Gu, Haoran, et autres
Publié: (2026)

GradSafe: Detecting Jailbreak Prompts for LLMs via Safety-Critical Gradient Analysis
par: Xie, Yueqi, et autres
Publié: (2024)

Do Prompts Guarantee Safety? Mitigating Toxicity from LLM Generations through Subspace Intervention
par: Singh, Himanshu, et autres
Publié: (2026)

Bias Amplification in RAG: Poisoning Knowledge Retrieval to Steer LLMs
par: Wang, Linlin, et autres
Publié: (2025)