Since the release of DeepSeek-R1 in January 2025, there has been a notable rise in the use of reinforcement learning based on verifiable rewards (RLVR), sparking interest in its application for enhancing products and academic research, especially in multimodal contexts. RLVR stands out because it allows models to explore and exploit actions for rewards, which is particularly beneficial when traditional differentiable loss functions are hard to define, such as in tasks requiring reasoning or dynamic decision-making. However, challenges like entropy collapse, where models become overly deterministic, highlight the need for a balanced exploration-exploitation strategy. In multimodal settings, successful RL applications often require a combination of offline RL training data, cold-start data from supervised fine-tuning, and the careful scaffolding of training data. This scaffolding ensures stable training and prevents common pitfalls like reward hacking and training instability. The text discusses various approaches to generating data for multimodal RL, including leveraging human perception with automated verification, model-generated descriptions for video prediction, and chaining specialized models for geometric tasks, while emphasizing the importance of curriculum learning to progressively increase task complexity. The iterative process of refining data acquisition and integrating human feedback is crucial, not only for starting the RL process but also for improving and maintaining the system’s performance over time.