LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL

LLM-as-a-Tutor mitigates reward signal collapse in non-verifiable RL by dynamically generating training prompts that match evolving policy capabilities. The framework uses a single model to act as both an examiner, performing pairwise comparisons to detect non-discriminative prompts, and a generator to adapt task difficulty.