ICML paper analyzes how combining supervised fine-tuning and reinforcement learning drives compositional generalization in language models · Digg