SCoRe is a student-centric framework designed to improve the complex task-solving ability of Large Language Model (LLM) agents. This framework involves a student model generating training trajectories, while a teacher model corrects only the student's initial errors. This generates training data that matches the student's abilities and exposes specific weaknesses. SCoRe involves fine-tuning the student model on the corrected trajectories and short-term reinforcement learning, starting from a proven prefix prior to the initial error and assigning a target reward at that step. Through SCoRe, a 7B-parameter student model achieved agent performance equivalent to that of a 72B-parameter teacher model.
