Rubrics Enhance Open-Ended AI Generation as Privileged Information

This research explores a novel approach to improving open-ended text generation in large language models (LLMs) by leveraging rubrics as "privileged information" (PI) during on-policy self-distillation (OPSD). Unlike traditional methods that use rubrics as scalar rewards for reinforcement learning (RL) or rely on hard reference completions, this study demonstrates that rubrics provide a much richer, dense signal when used as soft PI. The findings indicate that distilling towards soft rubric PI is more effective than distilling towards hard reference completions, as rubrics capture the preference structure across a range of valid responses rather than over-constraining the model to a single correct answer. This method, termed RuPI, significantly outperforms rubric-as-reward RL and reference-PI distillation across Qwen and Llama model families on benchmarks like HealthBench and ResearchQA, leading to higher quality, more nuanced open-ended generations.