Reproduce “Co-Evolving Policy and Rubric via Capability Awareness for Open-Ended Generation” 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

Chengsheng Zhang, Wenshen Xu, Yuchuan Wu, Zhiyuan Yu, Junfeng Fang, Tianlong Wang, Bingxue Qiu, Yongbin Li and 1 more

CoCA trains models for open-ended answers by comparing query-only and rubric-guided responses from the same policy, then adapting rubric criteria and routing training based on the differences. On Qwen3-8B, it reports a mean of 66.41 across seven benchmarks, versus 63.98 for rubric-based RL. The comparison uses more default rollouts, and diagnosis depends on finite samples and LLM judges.

Summary by alphaXiv.

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