Get Started
Home
Topics
Search
Library
Research questionHow should vision-language models answer valid parts of compound queries while withholding unsafe or unanswerable parts?A single visual-language query can combine content that should be answered with content that should be corrected, withheld, or treated as unknown. All-or-nothing response behavior can therefore both provide unsafe answers and refuse valid components.
AI
Alignment & Safety
Computer Vision
Evaluation & Benchmarks
Multimodal Models
Latest papersRecent research connected to this question, newest first.Knowing What Not to Answer: Selective Non-Compliance in Vision-Language ModelsThe source introduces and evaluates a benchmark covering false premises, visual inaccessibility, universal unknowns, task feasibility, and safety at both query and component levels. It also studies fine-tuning with selective-response examples; the evidence is limited to the evaluated vision-language models and benchmark tasks.research paper · Sep 4, 2026
Related questions
How can vision-language models correct unsafe generations token by token without disrupting safe reasoning?How can vision-language models decide when missing user context requires deferring rather than answering?How can vision-language models express visually grounded answers with context-appropriate information structure?How should causal VLMs preserve access to questions placed before image tokens?