2026 · Research

VLM robustness

CVPR 2026 workshop paper: do safety-aligned vision-language models degrade differently under common image corruptions?

First controlled base-vs-aligned evaluation of vision-language models under the ImageNet-C corruption suite, across four matched model pairs and three alignment paradigms (MPO, instruction tuning, SFT+RLHF). We identified and corrected a verbosity-scoring mismatch that biases standard VQA exact-match metrics against aligned models, and showed that alignment paradigm is a stronger predictor of robustness change than model scale — with a severity-dependent reversal under defocus blur. MPO alignment reduced corruption errors by 17% (Relative mCE: 0.830).

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