safetyfactbullishVision-Language Models can be trained to reason about AI-generated image edits using reinforcement learning without explicit reasoning supervisionComputer Vision02 Aug 2026http://arxiv.org/abs/2607.28464v1