Compositional Unalignment: Auditing Safety in Large Vision-Language Models
cs.LG, cs.AI, cs.CL
Submitted: 2026-11-04
Updated: 2026-11-04
Terminology
Sources
- Qwen3-VL Technical Report
- Qwen2.5-VL Technical Report
- SCANS: Mitigating the Exaggerated Safety for LLMs via Safety-Conscious Activation Steering
- Model Tampering Attacks Enable More Rigorous Evaluations of LLM Capabilities
- Harnessing Textual Refusal Directions for Multimodal Safety
- Faithfulness to Refusal: A Causal Audit of Neuron Selectors
- Perfect Detection, Failed Control: The Geometry of Knowing vs. Steering in Language Models
- Understanding the Effects of Safety Unalignment on Large Language Models
- When Behavioral Safety Evaluation Fails: A Representation-Level Perspective
- What's in the Image? A Deep-Dive into the Vision of Vision Language Models
- A Single Neuron Is Sufficient to Bypass Safety Alignment in Large Language Models
- AtP*: An efficient and scalable method for localizing LLM behaviour to components
- LoRA Fine-tuning Efficiently Undoes Safety Training in Llama 2-Chat 70B
- Security Tensors as a Cross-Modal Bridge: Extending Text-Aligned Safety to Vision in LVLM
- Safety Layers in Aligned Large Language Models: The Key to LLM Security
- Inference-Time Vulnerability Beyond Shallow Safety: Alignment Along Generation Trajectories
- Senses Wide Shut: A Representation-Action Gap in Omnimodal LLMs
- Detection Without Correction: A Robust Asymmetry in Activation-Based Hallucination Probing
- Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs
- Decodability is Not Causality: Dissociating Probe Readouts from Behavioral Drivers via SAE Decomposition
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis