mlsanityData to 7 Oct 2026

Unraveling and Mitigating Safety Alignment Degradation of Vision-Language Models

arXiv 2410.09047 ↗First posted 11 Oct 2024cs.CL cs.AI cs.LGcited by 23

On the map: Multimodal LLMs / vision-language models → MLLM jailbreaks and adversarial robustness

also in 2 more
Abstract (arXiv)

The safety alignment ability of Vision-Language Models (VLMs) is prone to be degraded by the integration of the vision module compared to its LLM backbone. We investigate this phenomenon, dubbed as ''safety alignment degradation'' in this paper, and show that the challenge arises from the representation gap that emerges when introducing vision modality to VLMs. In particular, we show that the representations of multi-modal inputs shift away from that of text-only inputs which represent the distribution that the LLM backbone is optimized for. At the same time, the safety alignment capabilities, initially developed within the textual embedding space, do not successfully transfer to this new multi-modal representation space. To reduce safety alignment degradation, we introduce Cross-Modality Representation Manipulation (CMRM), an inference time representation intervention method for recovering the safety alignment ability that is inherent in the LLM backbone of VLMs, while simultaneously preserving the functional capabilities of VLMs. The empirical results show that our framework significantly recovers the alignment ability that is inherited from the LLM backbone with minimal impact on the fluency and linguistic capabilities of pre-trained VLMs even without additional training. Specifically, the unsafe rate of LLaVA-7B on multi-modal input can be reduced from 61.53% to as low as 3.15% with only inference-time intervention. WARNING: This paper contains examples of toxic or harmful language.

Key facts. Cited by 23 papers in the map (data through 7 Oct 2026). Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 35 of its 43 references that are in the map. First posted on arXiv 11 Oct 2024.

Builds on

35 of 43 references.

Closest ancestors · 6

Same direction, newest first

  1. 2024Safety Alignment for Vision Language Models22 May
  2. 2024Images are Achilles' Heel of Alignment: Exploiting Visual Vulnerabilities for Jailbreaking Multimodal Large Language Models14 Mar
  3. 2024Eyes Closed, Safety On: Protecting Multimodal LLMs via Image-to-Text Transformation14 Mar
  4. 2024Red Teaming Visual Language Models23 Jan
  5. 2024MLLM-Protector: Ensuring MLLM's Safety without Hurting Performance5 Jan
  6. 2023Can Language Models be Instructed to Protect Personal Information?3 Oct

Other references · 29

  1. 2021Learning Transferable Visual Models From Natural Language Supervisioncited by 33,456
  2. 2014Microsoft COCO: Common Objects in Contextcited by 17,498
  3. 2022Training language models to follow instructions with human feedbackcited by 15,712
  4. 2023Judging LLM-as-a-Judge with MT-Bench and Chatbot Arenacited by 8,393
  5. 2023Direct Preference Optimization: Your Language Model is Secretly a Reward Modelcited by 7,916
  6. 2023Visual Instruction Tuningcited by 7,792
  7. 2017Deep reinforcement learning from human preferencescited by 3,846
  8. 2023InstructBLIP: Towards General-purpose Vision-Language Models with Instruction Tuningcited by 2,855
  9. 2023MiniGPT-4: Enhancing Vision-Language Understanding with Advanced Large Language Modelscited by 2,689
  10. 2022Learn to Explain: Multimodal Reasoning via Thought Chains for Science Question Answeringcited by 1,881
  11. 2023Inference-Time Intervention: Eliciting Truthful Answers from a Language Modelcited by 1,259
  12. 2023Representation Engineering: A Top-Down Approach to AI Transparencycited by 1,239
17 more
  1. 2021A General Language Assistant as a Laboratory for Alignmentcited by 1,023
  2. 2023ShareGPT4V: Improving Large Multi-Modal Models with Better Captionscited by 1,000
  3. 2023"Do Anything Now": Characterizing and Evaluating In-The-Wild Jailbreak Prompts on Large Language Modelscited by 457
  4. 2023Beyond Hallucinations: Enhancing LVLMs through Hallucination-Aware Direct Preference Optimizationcited by 223
  5. 2022Extracting Latent Steering Vectors from Pretrained Language Modelscited by 217
  6. 2024Aligning Modalities in Vision Large Language Models via Preference Fine-tuningcited by 197
  7. 2024Multi-Modal Hallucination Control by Visual Information Groundingcited by 196
  8. 2024Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Modelscited by 160
  9. 2023Inspecting and Editing Knowledge Representations in Language Modelscited by 140
  10. 2023Silkie: Preference Distillation for Large Visual Language Modelscited by 128
  11. 2024On Prompt-Driven Safeguarding for Large Language Modelscited by 109
  12. 2023DRESS: Instructing Large Vision-Language Models to Align and Interact with Humans via Natural Language Feedbackcited by 99
  13. 2024mDPO: Conditional Preference Optimization for Multimodal Large Language Modelscited by 92
  14. 2024InferAligner: Inference-Time Alignment for Harmlessness through Cross-Model Guidancecited by 78
  15. 2024Strengthening Multimodal Large Language Model with Bootstrapped Preference Optimizationcited by 75
  16. 2024Enhancing Large Vision Language Models with Self-Training on Image Comprehensioncited by 70
  17. 2024Multi-modal Preference Alignment Remedies Degradation of Visual Instruction Tuning on Language Modelscited by 26

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2024: 2; 2025: 14; 2026: 7.

23 papers.

Sort:

By area:

  1. 2025A Survey of Safety on Large Vision-Language Models: Attacks, Defenses and Evaluationscited by 66
  2. 2025BadVLA: Towards Backdoor Attacks on Vision-Language-Action Models via Objective-Decoupled Optimizationcited by 37
  3. 2024AutoTrust: Benchmarking Trustworthiness in Large Vision Language Models for Autonomous Drivingcited by 35
  4. 2025A Survey of Multimodal Hallucination Evaluation and Detectioncited by 26
  5. 2025Panacea: Mitigating Harmful Fine-tuning for Large Language Models via Post-fine-tuning Perturbationcited by 20
  6. 2025Understanding and Rectifying Safety Perception Distortion in VLMscited by 17
  7. 2025Unveiling the Basin-Like Loss Landscape in Large Language Modelscited by 13
  8. 2025Attention! Your Vision Language Model Could Be Maliciously Manipulatedcited by 12
  9. 2024Defending LVLMs Against Vision Attacks through Partial-Perception Supervisioncited by 7
  10. 2025EcoAlign: An Economically Rational Framework for Efficient LVLM Alignmentcited by 5
  11. 2025Bridging the Gap in Vision Language Models in Identifying Unsafe Concepts Across Modalitiescited by 4
  12. 2025Bootstrapping LLM Robustness for VLM Safety via Reducing the Pretraining Modality Gapcited by 3
11 more
  1. 2026Surgery: Mitigating Harmful Fine-Tuning for Large Language Models via Attention Sinkcited by 3
  2. 2025Dynamic Token Reweighting for Robust Vision-Language Modelscited by 1
  3. 2025LLaVAShield: Safeguarding Multimodal Multi-Turn Dialogues in Vision-Language Modelscited by 1
  4. 2025SGM: Safety Glasses for Multimodal Large Language Models via Neuron-Level Detoxificationcited by 1
  5. 2026Safety Geometry Collapse in Multimodal LLMs and Adaptive Drift Correctioncited by 1
  6. 2025VisuoAlign: Safety Alignment of LVLMs with Multimodal Tree Searchcited by 0
  7. 2026Safety-Potential Pruning for Enhancing Safety Prompts Against VLM Jailbreaking Without Retrainingcited by 0
  8. 2026Two Birds, One Projection: Harmonizing Safety and Utility in LVLMs via Inference-time Feature Projectioncited by 0
  9. 2026Guardrail-Agnostic Societal Bias Evaluation in Large Vision-Language Modelscited by 0
  10. 2026SPARK: Representation-Level KV Memory Alignment for Safer Vision-Language Modelscited by 0
  11. 2026SafeMol: Dual-Modality Safety Alignment for Molecular Multimodal Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024Cross-Modal Safety Mechanism Transfer in Large Vision-Language Models
  2. 2025VLM-Guard: Safeguarding Vision-Language Models via Fulfilling Safety Alignment Gap
  3. 2024PSA-VLM: Enhancing Vision-Language Model Safety through Progressive Concept-Bottleneck-Driven Alignment
  4. 2024SPA-VL: A Comprehensive Safety Preference Alignment Dataset for Vision Language Model
  5. 2024ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time
  6. 2026SafeRI: Recognition and Intervention for Token-Level Safety Intervention in Large Vision Language Models
  7. 2025MSR-Align: Policy-Grounded Multimodal Alignment for Safety-Aware Reasoning in Vision-Language Models
  8. 2024Layer-wise Alignment: Examining Safety Alignment Across Image Encoder Layers in Vision Language Models