2026 · Research
VLM robustness
CVPR 2026 workshop paper: do safety-aligned vision-language models degrade differently under common image corruptions?
0.830relative mCE with MPO alignment: 17% fewer corruption errors
- My role
- Second of five authors.
- Result
- Accepted at the CVPR 2026 Workshop on Generative Models for Computer Vision.
- Built with
- PyTorch, Vision-language models, ImageNet-C, VQA
Problem
Safety alignment changes how vision-language models behave, but no one had checked whether aligned models fail differently from their base models when images are blurred, noisy or otherwise degraded.
My role
Second of five authors. I built a reproducible evaluation pipeline across GQA, VQAv2 and TextVQA, with four corruption types at three severities.
What I built
- CVPR 2026 Workshop on Generative Models for Computer Vision
- Four matched pairs: InternVL, Qwen2-VL, Gemma — GQA, VQAv2, TextVQA
- Corrected verbosity bias in VQA exact-match scoring
- Alignment paradigm predicted robustness change more than scale
Outcome
Accepted at the CVPR 2026 Workshop on Generative Models for Computer Vision. Alignment paradigm predicted robustness change better than model scale did.