RL-finetuned VLMs vulnerable to weak visual perturbations

Apple study finds RL fine-tuning improves VLMs on visual reasoning benchmarks but models remain vulnerable to weak visual perturbations. The paper examines chain-of-thought consistency under such attacks.