Cover Image

Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning?

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Yin, Qingyu, Leong, Chak Tou, Yang, Linyi, Huang, Wenxuan, Li, Wenjie, Wang, Xiting, Yoon, Jaehong, YunXing, XingYu, Gu, Jinjin
Format:	Preprint
Published:	2025
Subjects:	Artificial Intelligence Cryptography and Security
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Why Safeguarded Ships Run Aground? Aligned Large Language Models' Safety Mechanisms Tend to Be Anchored in The Template Region
by: Leong, Chak Tou, et al.
Published: (2025)

Induction Head Toxicity Mechanistically Explains Repetition Curse in Large Language Models
by: Wang, Shuxun, et al.
Published: (2025)

Constrain Alignment with Sparse Autoencoders
by: Yin, Qingyu, et al.
Published: (2024)

Finding RELIEF: Shaping Reasoning Behavior without Reasoning Supervision via Belief Engineering
by: Leong, Chak Tou, et al.
Published: (2026)

Defensive Refusal Bias: How Safety Alignment Fails Cyber Defenders
by: Campbell, David, et al.
Published: (2026)

Scaling over Scaling: Exploring Test-Time Scaling Plateau in Large Reasoning Models
by: Wang, Jian, et al.
Published: (2025)

Seeing Isn't Believing: Mitigating Belief Inertia via Active Intervention in Embodied Agents
by: Wang, Hanlin, et al.
Published: (2026)

STeCa: Step-level Trajectory Calibration for LLM Agent Learning
by: Wang, Hanlin, et al.
Published: (2025)

E2CL: Exploration-based Error Correction Learning for Embodied Agents
by: Wang, Hanlin, et al.
Published: (2024)

TokenSkip: Controllable Chain-of-Thought Compression in LLMs
by: Xia, Heming, et al.
Published: (2025)

Evaluating Parameter Efficient Methods for RLVR
by: Yin, Qingyu, et al.
Published: (2025)

The End of an Era...Falling off a Cliff
by: Zmuda, Allison
Published: (2010)

Merlin's Whisper: Enabling Efficient Reasoning in Large Language Models via Black-box Persuasive Prompting
by: Xia, Heming, et al.
Published: (2025)

SPA-RL: Reinforcing LLM Agents via Stepwise Progress Attribution
by: Wang, Hanlin, et al.
Published: (2025)

Mitigating Unhelpfulness in Emotional Support Conversations with Multifaceted AI Feedback
by: Wang, Jiashuo, et al.
Published: (2024)

Bringing the Mountain to Mohammed without Falling off the Cliff of Unmanageable Technology.
by: Algermissen, Virginia, et al.
Published: (1992)

No Two Devils Alike: Unveiling Distinct Mechanisms of Fine-tuning Attacks
by: Leong, Chak Tou, et al.
Published: (2024)

Instruct Once, Chat Consistently in Multiple Rounds: An Efficient Tuning Framework for Dialogue
by: Wang, Jian, et al.
Published: (2024)

KNN-SSD: Enabling Dynamic Self-Speculative Decoding via Nearest Neighbor Layer Set Optimization
by: Song, Mingbo, et al.
Published: (2025)

Detection Is Cheap, Routing Is Learned: Why Refusal-Based Alignment Evaluation Fails
by: Frank, Gregory N.
Published: (2026)

When Refusals Fail: Unstable Safety Mechanisms in Long-Context LLM Agents
by: Hadeliya, Tsimur, et al.
Published: (2025)

Deeper Insights Without Updates: The Power of In-Context Learning Over Fine-Tuning
by: Yin, Qingyu, et al.
Published: (2024)

Probing the Difficulty Perception Mechanism of Large Language Models
by: Lee, Sunbowen, et al.
Published: (2025)

Subtle Errors in Reasoning: Preference Learning via Error-injected Self-editing
by: Xu, Kaishuai, et al.
Published: (2024)

Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training
by: Yuan, Youliang, et al.
Published: (2024)

Are Video Reasoning Models Ready to Go Outside?
by: He, Yangfan, et al.
Published: (2026)

Safety Alignment of Large Language Models via Contrasting Safe and Harmful Distributions
by: Zhang, Xiaoyun, et al.
Published: (2024)

Deactivating Refusal Triggers: Understanding and Mitigating Overrefusal in Safety Alignment
by: Xue, Zhiyu, et al.
Published: (2026)

CREMA: Generalizable and Efficient Video-Language Reasoning via Multimodal Modular Fusion
by: Yu, Shoubin, et al.
Published: (2024)

Towards Robust Temporal Reasoning of Large Language Models via a Multi-Hop QA Dataset and Pseudo-Instruction Tuning
by: Tan, Qingyu, et al.
Published: (2023)

Can Video LLMs Refuse to Answer? Alignment for Answerability in Video Large Language Models
by: Yoon, Eunseop, et al.
Published: (2025)

AutoPal: Autonomous Adaptation to Users for Personal AI Companionship
by: Cheng, Yi, et al.
Published: (2024)

Midgap states induced by Zeeman field and $p$ wave superconductor pairing
by: Jin, Yuanjun, et al.
Published: (2023)

Oyster-I: Beyond Refusal -- Constructive Safety Alignment for Responsible Language Models
by: Duan, Ranjie, et al.
Published: (2025)

From Threat to Tool: Leveraging Refusal-Aware Injection Attacks for Safety Alignment
by: Chae, Kyubyung, et al.
Published: (2025)

Safety is Not Only About Refusal: Reasoning-Enhanced Fine-tuning for Interpretable LLM Safety
by: Zhang, Yuyou, et al.
Published: (2025)

Probing Cultural Awareness in LLMs: A Case Study of Cross-Culture Aesthetic Stylistics
by: Wang, Jiashuo, et al.
Published: (2026)

Protein as a Second Language for LLMs
by: Chen, Xinhui, et al.
Published: (2025)

Ontology Neural Network and ORTSF: A Framework for Topological Reasoning and Delay-Robust Control
by: Oh, Jaehong
Published: (2025)

Video-Skill-CoT: Skill-based Chain-of-Thoughts for Domain-Adaptive Video Reasoning
by: Lee, Daeun, et al.
Published: (2025)