Metraipolis › archive › dc_c4290759060e
Depth-Dependent Safety Geometry in Vision-Language Models: Layer-Wise Localization and Bidirectional Refusal Control — arXiv:2610.32182
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.32182
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-26T19:53:50.114Z
- SHA-256 of the archived text
acdd461d117388891bd2d4fab8a7a70f71ea65c12b00295ab005649c6d6d2a4e- Stable URL of this page
https://metraipolis.com/documento/dc_c4290759060e
How to cite
Depth-Dependent Safety Geometry in Vision-Language Models: Layer-Wise Localization and Bidirectional Refusal Control — arXiv:2610.32182 — arXiv (en), https://arxiv.org/abs/2610.32182 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_c4290759060e (Metraipolis archive dc_c4290759060e, sha256 acdd461d117388891bd2d4fab8a7a70f71ea65c12b00295ab005649c6d6d2a4e)
BibTeX
@misc{dcc4290759060e,
title = {Depth-Dependent Safety Geometry in Vision-Language Models: Layer-Wise Localization and Bidirectional Refusal Control — arXiv:2610.32182},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.32182},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_c4290759060e; sha256 acdd461d117388891bd2d4fab8a7a70f71ea65c12b00295ab005649c6d6d2a4e},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.32182. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Vision-Language Models (VLMs) are widely deployed, yet the internal mechanisms behind their safety behavior remain poorly understood. We present a lightweight inference-time framework for localizing and manipulating safety-sensitive representations in frozen VLMs, without retraining or modifying parameters. On Qwen3-VL-4B, layer-wise activation divergence analysis reveals non-uniform safety geometry across depth. Causal interventions show that early vision layers support strong bidirectional refusal control through activation steering along a learned safety direction, with unsafe rejection spanning 6%-96%, whereas later layers respond more to activation magnitude scaling and exhibit a substantial safety-selectivity trade-off. Deeper layers remain well separated along the safety direction yet are largely insensitive to steering, distinguishing representational separability from causal controllability. InternVL2.5-4B shows qualitatively similar depth-dependent activation geometry despite differences in layer location and magnitude. Our results expose a white-box, representation-level refusal-bypass attack surface in VLM vision encoders and motivate inference-time stabilization of safety-sensitive representations. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno). Categorie arXiv: Computer Vision and Pattern Recognition (cs.CV) ; Computation and Language (cs.CL); Cryptography and Security (cs.CR); Machine Learning (cs.LG). Pagina del lavoro: https://arxiv.org/abs/2610.32182 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-26T19:45:48.391Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).