Can better questions and rubrics improve AI reasoning?

Training AI systems with rubric-based reinforcement learning hits a wall: vague queries produce vague rubrics, while overly specific ones demand references the model can't satisfy. QUBRIC solves this by jointly refining both. It converts open-ended questions into concrete scenarios, derives rubrics from teacher-policy mismatches, and filters for signal-rich pairs. The result: a 5.5-point gain on ArenaHard, plus strong transfer to reasoning tasks like legal and moral judgment—all without needing human verification of every answer.