DEEPRUBRIC: Evidence-Tree Rubric Supervision for Efficient Reinforcement Learning of Deep Research Agents
DEEPRUBRIC creates a framework to build high-quality training data for deep research agents by reversing traditional methods. Instead of letting an LLM guess evaluation criteria, DEEPRUBRIC first determines what an evidence-backed report should be judged on and then generates precise query-rubric pairs from that structure. This evidence-first approach leads to comparable performance with significantly fewer reinforcement learning GPU hours.