The paper presents a framework called SOAR, aimed at helping large language models (LLMs) overcome learning plateaus. When faced with datasets that yield low success rates, reinforcement learning approaches often stall due to a lack of training signals. SOAR tackles this problem by enabling a model to generate a self-directed curriculum. A teacher model produces synthetic problems for a student model, which then learns from these challenges. The teacher receives rewards based on the student’s improvement on difficult problems, rather than relying on traditional proxy rewards. The researchers conducted experiments using challenging mathematical benchmarks where initial success rates were at zero. They found that their method is effective in unlocking learning even under sparse, binary rewards. By leveraging the latent capabilities of pretrained models, SOAR can create useful problems that guide learning. The study also highlights a significant insight: the quality and clarity of the generated questions matter more than whether the questions can be solved correctly. This approach avoids issues seen in previous self-play methods, such as instability and diversity collapse, by grounding the curriculum in the actual progress of the student model. Ultimately, SOAR suggests that a model can generate stepping stones for learning without needing to solve the hardest problems upfront. This opens up new avenues for enhancing model training without the need for curated data, positioning SOAR as a promising method for advancing machine learning capabilities.