Image de couverture de livre

When Harmless Words Harm: A New Threat to LLM Safety via Conceptual Triggers

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Zhang, Zhaoxin, Chen, Borui, Hu, Yiming, Qu, Youyang, Zhu, Tianqing, Gao, Longxiang
Format:	Preprint
Publié:	2025
Sujets:	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

Recent Advances in Federated Learning Driven Large Language Models: A Survey on Architecture, Performance, and Security
par: Qu, Youyang, et autres
Publié: (2024)

Wisdom is Knowing What not to Say: Hallucination-Free LLMs Unlearning via Attention Shifting
par: Tan, Chenchen, et autres
Publié: (2025)

Less is More: Geometric Unlearning for LLMs with Minimal Data Disclosure
par: Tan, Chenchen, et autres
Publié: (2026)

Latent Fusion Jailbreak: Blending Harmful and Harmless Representations to Elicit Unsafe LLM Outputs
par: Xing, Wenpeng, et autres
Publié: (2025)

When Safety Fails Before the Answer: Benchmarking Harmful Behavior Detection in Reasoning Chains
par: Kakkar, Ishita, et autres
Publié: (2026)

Beyond the Safety Bundle: Auditing the Helpful and Harmless Dataset
par: Chehbouni, Khaoula, et autres
Publié: (2024)

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

Guardians and Offenders: A Survey on Harmful Content Generation and Safety Mitigation of LLM
par: Zhang, Chi, et autres
Publié: (2025)

Acquiring and Modelling Abstract Commonsense Knowledge via Conceptualization
par: He, Mutian, et autres
Publié: (2022)

Easier to Mislead Than to Correct: Harmful and Beneficial Revision in LLM Conformity
par: Qu, Jiaming, et autres
Publié: (2026)

Seeing No Evil: Blinding Large Vision-Language Models to Safety Instructions via Adversarial Attention Hijacking
par: Li, Jingru, et autres
Publié: (2026)

Dishonesty in Helpful and Harmless Alignment
par: Huang, Youcheng, et autres
Publié: (2024)

Safety Alignment of Large Language Models via Contrasting Safe and Harmful Distributions
par: Zhang, Xiaoyun, et autres
Publié: (2024)

Should LLM Safety Be More Than Refusing Harmful Instructions?
par: Maskey, Utsav, et autres
Publié: (2025)

ConKE: Conceptualization-Augmented Knowledge Editing in Large Language Models for Commonsense Reasoning
par: Zhang, Liyu, et autres
Publié: (2024)

Revisiting Backdoor Attacks on LLMs: A Stealthy and Practical Poisoning Framework via Harmless Inputs
par: Kong, Jiawei, et autres
Publié: (2025)

"They are uncultured": Unveiling Covert Harms and Social Threats in LLM Generated Conversations
par: Dammu, Preetam Prabhu Srikar, et autres
Publié: (2024)

LLM Judges Inconsistently Disagree Across Safety Criteria and Harm Categories
par: Vishnubhotla, Krishnapriya, et autres
Publié: (2026)

Owner-Harm: A Missing Threat Model for AI Agent Safety
par: Zhang, Dongcheng, et autres
Publié: (2026)

Beyond Words: Multimodal LLM Knows When to Speak
par: Liao, Zikai, et autres
Publié: (2025)

Understanding LLM Behavior When Encountering User-Supplied Harmful Content in Harmless Tasks
par: Chu, Junjie, et autres
Publié: (2026)

From Wide to Deep: Dimension Lifting Network for Parameter-efficient Knowledge Graph Embedding
par: Cai, Borui, et autres
Publié: (2023)

Compromising Honesty and Harmlessness in Language Models via Deception Attacks
par: Vaugrante, Laurène, et autres
Publié: (2025)

Understanding and Meeting Practitioner Needs When Measuring Representational Harms Caused by LLM-Based Systems
par: Harvey, Emma, et autres
Publié: (2025)

When Machine Unlearning Meets Retrieval-Augmented Generation (RAG): Keep Secret or Forget Knowledge?
par: Wang, Shang, et autres
Publié: (2024)

When Wording Steers the Evaluation: Framing Bias in LLM judges
par: Hwang, Yerin, et autres
Publié: (2026)

FFT: Towards Harmlessness Evaluation and Analysis for LLMs with Factuality, Fairness, Toxicity
par: Cui, Shiyao, et autres
Publié: (2023)

HarmMetric Eval: Benchmarking Metrics and Judges for LLM Harmfulness Assessment
par: Yang, Langqi, et autres
Publié: (2025)

HarmTransform: Transforming Explicit Harmful Queries into Stealthy via Multi-Agent Debate
par: Zhu, Shenzhe
Publié: (2025)

Causal Fingerprints of AI Generative Models
par: Xu, Hui, et autres
Publié: (2025)

Balancing Enhancement, Harmlessness, and General Capabilities: Enhancing Conversational LLMs with Direct RLHF
par: Zheng, Chen, et autres
Publié: (2024)

Reinforcement Learning Amplifies Emergent Misalignment from Harmless Rewards
par: Jørgenvåg, Magnus, et autres
Publié: (2026)

Too Helpful, Too Harmless, Too Honest or Just Right?
par: Kashyap, Gautam Siddharth, et autres
Publié: (2025)

From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios
par: Gao, Yuan, et autres
Publié: (2025)

InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidance
par: Wang, Pengyu, et autres
Publié: (2024)

Adaptive Deep Reasoning: Triggering Deep Thinking When Needed
par: Wang, Yunhao, et autres
Publié: (2025)

They Said Memes Were Harmless-We Found the Ones That Hurt: Decoding Jokes, Symbols, and Cultural References
par: Tripathi, Sahil, et autres
Publié: (2026)

Learning to Look at the Other Side: A Semantic Probing Study of Word Embeddings in LLMs with Enabled Bidirectional Attention
par: Feng, Zhaoxin, et autres
Publié: (2025)

Towards Harmless Multimodal Assistants with Blind Preference Optimization
par: Li, Yongqi, et autres
Publié: (2025)

From Judgment to Interference: Early Stopping LLM Harmful Outputs via Streaming Content Monitoring
par: Li, Yang, et autres
Publié: (2025)