Reproduce “Rubric-Aware On-Policy Self-Distillation for LLM Personalization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yilun Qiu, Xiaoyan Zhao, Chengbing Wang, Cilin Yan, Rui Zu, Wanyang Zhang, Xiaolong Jiang, Jiayin Cai and 1 more
GRASP trains a rubric-free student model to match a rubric-informed teacher’s next-token distributions across the student’s own generated prefixes, using forward KL divergence. A reliability filter retains examples whose teacher responses cover all rubric aspects. On LaMP-QA, Gemma2-9B scores 0.5255 versus PlanPers’s 0.4612, but GRASP does not lead every category, including Society & Culture for Qwen3-4B.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.