Research
160d agoGoldilocks RL: Tuning Difficulty for Sparse Rewards
Goldilocks, a teacher‑driven sampling method, predicts each question’s difficulty for a reinforcement‑learning student, easing navigation through sparse rewards. By adjusting task complexity in real time, the approach reduces sample inefficiency and accelerates reasoning in large language models.
The Signal
The technique, developed at Apple Machine Learning Research, shows promising gains over static curricula.