Reproduce “RubricWorld: World Model-Guided Rubric Optimization for LLM Graders” 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
Hongji Pu, Hang Li, Yucheng Chu, Jiliang Tang
RubricWorld improves frozen essay graders by predicting how rubric edits change correction and degradation rates, using paired scores from the same answers and the current rubric state. A separate held-out check determines whether an edit is accepted. On ASAP2.0 essays, frozen gpt-5-mini reaches 46.70% exact-score accuracy after five rounds, versus 24.10% for APO; this comparison is specific to that model, dataset, metric, and protocol.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.