Get Started
Home
Topics
Search
Library
Research questionHow can vision-language models maintain visual recognition when modalities are missing and source training data is unavailable?At deployment, incomplete modality inputs can shift a vision-language model’s predictions and degrade recognition accuracy. This is particularly restrictive when privacy, storage, or accessibility constraints prevent access to the original source training data.
AI
Computer Vision
Evaluation & Benchmarks
Image & Video Processing
Machine Learning
Multimodal Models
Research Paper
Statistical Machine Learning
Technology
Latest papersRecent research connected to this question, newest first.Test-Time Logit Prompting for Source-Free Missing Modality AdaptationApplies to vision-language models performing visual recognition with missing-modality inputs when source training data cannot be accessed. The reported evidence comes from diverse vision-language benchmarks, with improvements of up to 8%; the abstract does not specify particular deployment modalities or access assumptions beyond the source-free setting.research paper · Sep 2, 2026
Related questions
How can vision-language models resist multimodal jailbreaks that adapt their strategies and transfer across defenses?How can models adapt to low-resource languages without damaging source-language and related-task performance?How can we predict and interpret vision-language model failures to support timely human intervention?How can large vision-language models reduce visual hallucinations without extra training or decoding passes?