Can AI automatically write grading rubrics for training better language models?

Training language models on open-ended tasks like healthcare advice or creative writing requires nuanced evaluation—not just right/wrong answers. ARES automates this by converting raw documents into question-answer pairs, then generates custom rubrics tailored to each question that score responses across multiple dimensions. The system validates itself (checking questions are self-contained, answers faithful) and conditions generation on domain and persona to boost diversity. On benchmarks spanning healthcare, instruction-following, and other domains, rubric-based RL trained with ARES outperforms standard fine-tuning and binary-reward methods.