Researchers have made significant strides in enhancing the reasoning capabilities of multimodal large language models (MLLMs) through reinforcement learning with verifiable rewards (RLVR). However, the conventional approach to RLVR has a major flaw: it relies on outcome-driven optimization, where both perception and reasoning are updated using a shared reward based solely on the final answer. This methodology obscures credit assignment, often leading to improvements in reasoning at the expense of perception. A new approach, perception-reasoning coevolution, has been proposed to address this issue, allowing for more nuanced and effective optimization of MLLMs. By decoupling perception and reasoning, this method enables more accurate credit assignment and improved overall performance. The development of more advanced MLLMs through RLVR has significant implications for security, as it can both expand capabilities and introduce new risks1. This matters to practitioners because it highlights the need for careful consideration of the potential security consequences of emerging MLLM technologies.