LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks
The paper replaces scalar RL rewards with coached textual feedback for open-ended tasks.
Experiential Learning turns an LLM judge’s assessment into “experiential knowledge” that can guide a teacher model and then be distilled into the policy. The authors argue this preserves fine-grained preferences that rubric scores flatten away. In their experiments across two policy families, EL beats rubric-based RL on held-out and unseen open-ended tasks, generalizes better, and reduces reward hacking. HF Daily Papers' note
Experiential Learning turns an LLM judge’s assessment into “experiential knowledge” that can guide a teacher model and then be distilled into the policy. The authors argue this preserves fine-grained preferences that rubric scores flatten away. In their experiments across two policy families, EL beats rubric-based RL on held-out and unseen open-ended tasks, generalizes better, and reduces reward hacking. HF Daily Papers' note
score 5