<ul><li>Self-Taught Reasoners (STaR), also known as Rejection sampling Fine-Tuning (RFT), is crucial for training self-improving reasoning Language Models (LMs).</li><li>Random observation sampling often leads to trained observation imbalance, causing over-training on solved examples and under-training on challenging ones.</li><li>AdaSTaR is a new algorithm that addresses this issue by implementing Adaptive Sampling for Diversity and Adaptive Sampling for Curriculum to ensure balanced training and adjust data difficulty based on the model's strength.</li><li>Across six benchmarks, AdaSTaR outperforms other methods with best test accuracy in all instances (6/6) and reduces training FLOPs by an average of 58.6%, showing promise for more efficient and effective self-improving LMs.</li></ul>

AdaSTaR: Adaptive Data Sampling for Training Self-Taught Reasoners

Discover more