Reproduce “RubricEM: Meta-RL with Rubric-guided Policy Decomposition beyond Verifiable Rewards” 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

Gaotang Li, Bhavana Dalvi Mishra, Zifeng Wang, Jun Yan, Yanfei Chen, Chun-Liang Li, Long T. Le, Rujun Han and 4 more

RubricEM trains research agents in four stages—Plan, Research, Review, and Answer—using judge-scored rubrics for intermediate feedback and a reflection policy that turns attempts into guidance. A matched 600-step ablation finds both additions improve scores over baseline. Its 8B model scores 55.5 across four long-form benchmarks, versus 53.6 for DR Tulu-8B; these scores do not establish intermediate judgments are correct.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.