●builderThis method helps train agents on tasks that are currently too difficult for them to solve through random exploration alone.
●researcherYou can bypass the exploration bottleneck in agent RL by intervening during the rollout phase using evaluation rubrics.