Skip to content
AI News HubLIVE
Original source2 min read

When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs

Summary

Multimodal systems that behave reasonably on average can still be unsafe for individuals whose medical, emotional, or situational context is unknown. This paper introduces MPS-Bench (5,181 scenarios from 584 images across 12 high-risk domains), shows eight frontier VLMs almost always answer directly (86-99%) and score at most 2.6/5 on personalized safety, traces failures to visual dominance during multimodal fusion, and proposes PRISM, a monitor predicting deferral need with 0.978 AUC.

SourcearXiv Computer VisionAuthor: Edward Sun, Yuchen Wu, Zixian Ma, Eric Hanchen Jiang, Yijia Xiao, Xiaoyuan Yi, Ranjay Krishna, Wei Wang, Jindong Wang, Aylin Caliskan
When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs
Report an error

The correction channel is not available yet. You can copy the article reference below for later.

Correction instructions
Read article

[Submitted on 3 Sep 2026]

Title:When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs

View a PDF of the paper titled When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs, by Edward Sun and 9 other authors

View PDF HTML (experimental)

Abstract:Vision-language models (VLMs) are increasingly deployed in high-stakes settings, where a response that is reasonable in general may still be unsafe for a particular user whose medical, emotional, or situational context is unknown to the model. We study this problem of personalized safety in multimodal systems and introduce MPS-Bench, a benchmark of 5,181 scenarios from 584 real-world images across 12 high-risk domains, each paired with a hidden user profile. Evaluating eight frontier VLMs, we find that they almost always respond directly (86-99%) rather than seek missing context, and none exceeds 2.6/5 on personalized safety. To understand why these failures arise, we analyze multimodal interactions and identify visual dominance: visual information enters text representations early and suppresses textual risk signals during multimodal fusion. Causal interventions reveal a two-stage mechanism in which visual affect is first transferred into the text stream in early layers and then shapes the final decision through this altered text representation, making late-stage internal remediation unreliable. Motivated by this mechanism, we propose PRISM, a lightweight input monitor that uses bidirectional cross-modal modulation to predict when a query is likely to require deferral. PRISM achieves 0.978 AUC and strictly dominates the safety-utility Pareto frontier across all tested models.

Comments: Published as a main conference paper at COLM 2026

Subjects:

Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI); Machine Learning (cs.LG)

Cite as: arXiv:2609.04281 [cs.CV]

(or arXiv:2609.04281v1 [cs.CV] for this version)

https://doi.org/10.48550/arXiv.2609.04281

arXiv-issued DOI via DataCite (pending registration)

Submission history

From: Edward Sun [view email] [v1] Thu, 3 Sep 2026 03:58:56 UTC (4,354 KB)

Full-text links:

Access Paper:

View a PDF of the paper titled When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMs, by Edward Sun and 9 other authors

View PDF

HTML (experimental)

TeX Source

view license

Current browse context:

cs.CV

new | recent | 2026-09

Change to browse by:

cs cs.AI cs.LG

References & Citations

NASA ADS

Google Scholar

Semantic Scholar

Loading...

Data provided by:

Bibliographic Tools

Bibliographic and Citation Tools

Bibliographic Explorer Toggle

Bibliographic Explorer (What is the Explorer?)

Connected Papers Toggle

Connected Papers (What is Connected Papers?)

Litmaps Toggle

Litmaps (What is Litmaps?)

scite.ai Toggle

scite Smart Citations (What are Smart Citations?)

Code, Data, Media

Code, Data and Media Associated with this Article

alphaXiv Toggle

alphaXiv (What is alphaXiv?)

Links to Code Toggle

CatalyzeX Code Finder for Papers (What is CatalyzeX?)

DagsHub Toggle

DagsHub (What is DagsHub?)

GotitPub Toggle

Gotit.pub (What is GotitPub?)

Huggingface Toggle

Hugging Face (What is Huggingface?)

ScienceCast Toggle

ScienceCast (What is ScienceCast?)

Demos

Demos

Replicate Toggle

Replicate (What is Replicate?)

Spaces Toggle

Hugging Face Spaces (What is Spaces?)

Spaces Toggle

TXYZ.AI (What is TXYZ.AI?)

Related Papers

Recommenders and Search Tools

Link to Influence Flower

Influence Flower (What are Influence Flowers?)

Core recommender toggle

CORE Recommender (What is CORE?)

Author

Venue

Institution

Topic

About arXivLabs

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

Which authors of this paper are endorsers? | Disable MathJax (What is MathJax?)

Key points and analysis

Article intelligence

EngineersAdvanced

Key points

  • MPS-Bench: 5,181 multimodal scenarios from 584 real-world images across 12 high-risk domains, each tied to a hidden user profile.
  • Eight frontier VLMs answer directly in 86-99% of cases and none exceeds 2.6/5 personalized safety.
  • Analysis reveals visual dominance: visual signals enter text representations early and suppress textual risk cues, undermining late-stage remediation.
  • PRISM, a lightweight monitor based on bidirectional cross-modal modulation, predicts deferral need at 0.978 AUC and dominates safety-utility Pareto frontier across all tested models.

Highlights and analysis are generated automatically and may contain errors. Check the original source.