Ask a medical AI the same question two different ways, and today's top models might give you two different answers.
Researchers built a training method called Prompt-Invariant RLVR, or PIRL, to fix a specific weakness in how multimodal AI models get fine-tuned with reinforcement learning. The standard approach, Reinforcement Learning with Verifiable Rewards, boosts accuracy but trains on only a narrow slice of possible phrasings, so models can stumble when a question is reworded or reformatted at deployment. The researchers traced this to two flaws: a reward signal that lumps wrong answers together with correctly-reasoned but misformatted ones, and training data too thin to capture real-world variation. PIRL tackles both with a reward that separates format from content and a regularizer that pushes the model toward consistent answers across prompts that mean the same thing.
In stress tests, PIRL's accuracy dropped by no more than 1 percent when prompts were paraphrased, versus roughly 3 percent for GRPO, a widely used RLVR baseline. That gap sounds small, but in high-stakes settings like medical visual question answering, a model three times more fragile to rewording is a real liability: a differently worded question should not change a diagnosis-adjacent answer.
It is a reminder that most "reasoning" benchmarks test models under conditions tidier than the real world, where nobody phrases a question the same way twice.