non-verifiable-rl

LLM-as-a-Tutor introduces a framework that extends LLM's role from judge to tutor by dynamically adjusting prompt difficulty through pairwise comparison and constraint addition, improving instruction-following performance in reinforcement learning.