SAERL study reports fewer training steps for mathematical reasoning

Researchers report that SAERL, a method for organizing reinforcement learning data using a model's internal features, reached target mathematical reasoning accuracy with 42% fewer training steps on Qwen2.5-Math-1.5B. The preprint also reports a 3.1% improvement in average accuracy over GRPO on that model. The method uses sparse autoencoders, which separate internal activity into interpretable features, to select samples, order them by difficulty and group them into training batches. Its curriculum is constructed once using small models and transfers to larger ones. The reported experiments cover mathematical reasoning across model scales and reinforcement learning algorithms.