Imagen de Portada

Benign Samples Matter! Fine-tuning On Outlier Benign Samples Severely Breaks Safety

Fuente: arXiv

Guardado en:

Detalles Bibliográficos
Autores principales:	Guan, Zihan, Hu, Mengxuan, Zhu, Ronghang, Li, Sheng, Vullikanti, Anil
Formato:	Preprint
Publicado:	2025
Materias:	Machine Learning Computation and Language
Acceso en línea:	Acceder al recurso
Etiquetas:	Agregar Etiqueta Sin Etiquetas, Sea el primero en etiquetar este registro!

Ejemplares similares

UFID: A Unified Framework for Input-level Backdoor Detection on Diffusion Models
por: Guan, Zihan, et al.
Publicado: (2024)

Are LLMs Ready for Neural-integrated Mechanistic Modeling? A Benchmark and Agentic Framework
por: Guan, Zihan, et al.
Publicado: (2026)

Outlier-weighed Layerwise Sampling for LLM Fine-tuning
por: Li, Pengxiang, et al.
Publicado: (2024)

Large Language Models Lack Temporal Awareness of Medical Knowledge
por: Guan, Zihan, et al.
Publicado: (2026)

What is in Your Safe Data? Identifying Benign Data that Breaks Safety
por: He, Luxi, et al.
Publicado: (2024)

What Makes and Breaks Safety Fine-tuning? A Mechanistic Study
por: Jain, Samyak, et al.
Publicado: (2024)

Improving Sampling Methods for Fine-tuning SentenceBERT in Text Streams
por: Garcia, Cristiano Mesquita, et al.
Publicado: (2024)

No Free Lunch: Retrieval-Augmented Generation Undermines Fairness in LLMs, Even for Vigilant Users
por: Hu, Mengxuan, et al.
Publicado: (2024)

Memento: Fine-tuning LLM Agents without Fine-tuning LLMs
por: Zhou, Huichi, et al.
Publicado: (2025)

GRASS: Gradient-based Adaptive Layer-wise Importance Sampling for Memory-efficient Large Language Model Fine-tuning
por: Tian, Kaiyuan, et al.
Publicado: (2026)

Blending Supervised and Reinforcement Fine-Tuning with Prefix Sampling
por: Huang, Zeyu, et al.
Publicado: (2025)

Layer-wise Importance Matters: Less Memory for Better Performance in Parameter-efficient Fine-tuning of Large Language Models
por: Yao, Kai, et al.
Publicado: (2024)

Few Tokens, Big Leverage: Preserving Safety Alignment by Constraining Safety Tokens during Fine-tuning
por: Wang, Guoli, et al.
Publicado: (2026)

RoLoRA: Fine-tuning Rotated Outlier-free LLMs for Effective Weight-Activation Quantization
por: Huang, Xijie, et al.
Publicado: (2024)

Why LLM Safety Guardrails Collapse After Fine-tuning: A Similarity Analysis Between Alignment and Fine-tuning Datasets
por: Hsiung, Lei, et al.
Publicado: (2025)

Multitask Mayhem: Unveiling and Mitigating Safety Gaps in LLMs Fine-tuning
por: Jan, Essa, et al.
Publicado: (2024)

Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models
por: Goel, Anmol, et al.
Publicado: (2026)

Advancing Parameter Efficiency in Fine-tuning via Representation Editing
por: Wu, Muling, et al.
Publicado: (2024)

Objective Matters: Fine-Tuning Objectives Shape Safety, Robustness, and Persona Drift
por: Vennemeyer, Daniel, et al.
Publicado: (2026)

PMSS: Pretrained Matrices Skeleton Selection for LLM Fine-tuning
por: Wang, Qibin, et al.
Publicado: (2024)

SEE: Continual Fine-tuning with Sequential Ensemble of Experts
por: Wang, Zhilin, et al.
Publicado: (2025)

CFSafety: Comprehensive Fine-grained Safety Assessment for LLMs
por: Liu, Zhihao, et al.
Publicado: (2024)

Fundamental Safety-Capability Trade-offs in Fine-tuning Large Language Models
por: Chen, Pin-Yu, et al.
Publicado: (2025)

Learning Rate Matters: Vanilla LoRA May Suffice for LLM Fine-tuning
por: Lee, Yu-Ang, et al.
Publicado: (2026)

Efficient Ensemble for Fine-tuning Language Models on Multiple Datasets
por: Li, Dongyue, et al.
Publicado: (2025)

LoX: Low-Rank Extrapolation Robustifies LLM Safety Against Fine-tuning
por: Perin, Gabriel J., et al.
Publicado: (2025)

SEAL: Safety-enhanced Aligned LLM Fine-tuning via Bilevel Data Selection
por: Shen, Han, et al.
Publicado: (2024)

A Classical View on Benign Overfitting: The Role of Sample Size
por: Park, Junhyung, et al.
Publicado: (2025)

REAL Sampling: Boosting Factuality and Diversity of Open-Ended Generation via Asymptotic Entropy
por: Chang, Haw-Shiuan, et al.
Publicado: (2024)

Vaccine: Perturbation-aware Alignment for Large Language Models against Harmful Fine-tuning Attack
por: Huang, Tiansheng, et al.
Publicado: (2024)

CoMoE: Contrastive Representation for Mixture-of-Experts in Parameter-Efficient Fine-tuning
por: Feng, Jinyuan, et al.
Publicado: (2025)

Topic Modeling with Fine-tuning LLMs and Bag of Sentences
por: Schneider, Johannes
Publicado: (2024)

On the Loss of Context-awareness in General Instruction Fine-tuning
por: Wang, Yihan, et al.
Publicado: (2024)

DataShield: Safety-degrading Data Filtering for LLM Benign Instruction Fine-Tuning
por: Zhang, Junbo, et al.
Publicado: (2026)

Learn Hard Problems During RL with Reference Guided Fine-tuning
por: Wu, Yangzhen, et al.
Publicado: (2026)

Threshold Filtering Packing for Supervised Fine-Tuning: Training Related Samples within Packs
por: Dong, Jiancheng, et al.
Publicado: (2024)

Two-stage LLM Fine-tuning with Less Specialization and More Generalization
por: Wang, Yihan, et al.
Publicado: (2022)

A General Framework to Enhance Fine-tuning-based LLM Unlearning
por: Ren, Jie, et al.
Publicado: (2025)

ResRL: Boosting LLM Reasoning via Negative Sample Projection Residual Reinforcement Learning
por: Lin, Zihan, et al.
Publicado: (2026)

Dissecting Outlier Dynamics in LLM NVFP4 Pretraining
por: Dong, Peijie, et al.
Publicado: (2026)