Get Started
Home
Topics
Search
Library
Research questionHow can vision-language models decide when missing user context requires deferring rather than answering?A response that is generally reasonable can be unsafe for a particular user when relevant medical, emotional, or situational context is unavailable. Visual information may also enter the model’s text representation early and suppress textual signals that should prompt caution.
AI
Alignment & Safety
Evaluation & Benchmarks
Multimodal Models
Latest papersRecent research connected to this question, newest first.When Seeing Overrides Knowing: Visual Dominance and Deferral-Based Method for Personalized Safety in VLMsThe evidence concerns eight frontier VLMs evaluated on MPS-Bench: 5,181 scenarios drawn from 584 real-world images across 12 high-risk domains, each paired with a hidden user profile. The study measures direct responses, deferral behavior, and personalized safety, analyzes visual dominance through multimodal interventions, and tests an input monitor for predicting when deferral may be needed.research paper · Sep 3, 2026
Related questions
How can vision-language models express visually grounded answers with context-appropriate information structure?How can speech language models protect one user’s private context when responding to another speaker?How should vision-language models answer valid parts of compound queries while withholding unsafe or unanswerable parts?How can vision-language models correct unsafe generations token by token without disrupting safe reasoning?