ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals

A study titled "ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals" examines whether language model-generated rubrics used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading reliably reward honest answers over adversarial answers optimized to exploit them. The research focuses on the robustness of these generated rubrics to such adversarial exploitation.

Language model-generated rubrics are increasingly used as reward signals for rubric-based reinforcement learning, LLM-as-a-judge evaluation, and automated grading. Such rubrics are reliable only if they reward honest answers over adversarial answers optimized to exploit them. Yet their robustness to