mlsanityData to 14 Oct 2026

Internal Activation Revision: Safeguarding Vision Language Models Without Parameter Update

arXiv 2501.16378 ↗First posted 24 Jan 2025cs.LG cs.AI cs.CLcited by 16

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

also in 1 more
Abstract (arXiv)

Vision-language models (VLMs) demonstrate strong multimodal capabilities but have been found to be more susceptible to generating harmful content compared to their backbone large language models (LLMs). Our investigation reveals that the integration of images significantly shifts the model's internal activations during the forward pass, diverging from those triggered by textual input. Moreover, the safety alignments of LLMs embedded within VLMs are not sufficiently robust to handle the activations discrepancies, making the models vulnerable to even the simplest jailbreaking attacks. To address this issue, we propose an internal activation revision approach that efficiently revises activations during generation, steering the model toward safer outputs. Our framework incorporates revisions at both the layer and head levels, offering control over the model's generation at varying levels of granularity. In addition, we explore three strategies for constructing positive and negative samples and two approaches for extracting revision vectors, resulting in different variants of our method. Comprehensive experiments demonstrate that the internal activation revision method significantly improves the safety of widely used VLMs, reducing attack success rates by an average of 48.94%, 34.34%, 43.92%, and 52.98% on SafeBench, Safe-Unsafe, Unsafe, and MM-SafetyBench, respectively, while minimally impacting model helpfulness.

Key facts. Cited by 16 papers in the map (data through 14 Oct 2026). Placed in MLLM jailbreaks and adversarial robustness, one of 16 directions in Multimodal LLMs / vision-language models. Builds on 2 of its 34 references that are in the map. First posted on arXiv 24 Jan 2025.

Builds on

2 of 34 references.

References · 2

  1. 2023Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyondcited by 2,224
  2. 2023MiniGPT-v2: large language model as a unified interface for vision-language multi-task learningcited by 625

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 12; 2026: 4.

16 papers.

Sort:

By area:

  1. 2025Language Models Are Capable of Metacognitive Monitoring and Control of Their Internal Activationscited by 41
  2. 2025A Comprehensive Survey of Machine Unlearning Techniques for Large Language Modelscited by 34
  3. 2025Learning to Steer: Input-dependent Steering for Multimodal LLMscited by 15
  4. 2025SAUCE: Selective Concept Unlearning in Vision-Language Models with Sparse Autoencoderscited by 13
  5. 2025DAVSP: Safety Alignment for Large Vision-Language Models via Deep Aligned Visual Safety Promptcited by 8
  6. 2026Efficient and accurate steering of Large Language Models through attention-guided feature learningcited by 8
  7. 2025Con Instruction: Universal Jailbreaking of Multimodal Large Language Models via Non-Textual Modalitiescited by 7
  8. 2025ARGUS: Defending Against Multimodal Indirect Prompt Injection via Steering Instruction-Following Behaviorcited by 7
  9. 2025VSCBench: Bridging the Gap in Vision-Language Model Safety Calibrationcited by 5
  10. 2025Activation Steering Meets Preference Optimization: Defense Against Jailbreaks in Vision Language Modelscited by 5
  11. 2025A Comprehensive Analysis for Visual Object Hallucination in Large Vision-Language Modelscited by 4
  12. 2025Secure Tug-of-War (SecTOW): Iterative Defense-Attack Training with Reinforcement Learning for Multimodal Model Securitycited by 4
4 more
  1. 2026Understanding and Defending VLM Jailbreaks via Jailbreak-Related Representation Shiftcited by 4
  2. 2025A Survey on Training-free Alignment of Large Language Modelscited by 2
  3. 2026One-shot Optimized Steering Vector for Hallucination Mitigation for VLMscited by 1
  4. 2026SPARK: Representation-Level KV Memory Alignment for Safer Vision-Language Modelscited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2024ETA: Evaluating Then Aligning Safety of Vision Language Models at Inference Time
  2. 2024Safety Fine-Tuning at (Almost) No Cost: A Baseline for Vision Large Language Models
  3. 2023FigStep: Jailbreaking Large Vision-Language Models via Typographic Visual Prompts
  4. 2025Breach in the Shield: Unveiling the Vulnerabilities of Large Language Models
  5. 2026Single-Sample Black-Box Membership Inference Attack against Vision-Language Models via Cross-modal Semantic Alignment
  6. 2024ImgTrojan: Jailbreaking Vision-Language Models with ONE Image
  7. 2025Safeguarding Vision-Language Models: Mitigating Vulnerabilities to Gaussian Noise in Perturbation-based Attacks
  8. 2025Membership Inference Attacks Against Vision-Language Models