Reproduce “PaperGym: Rubric-Centered Evolution for Research-Plan 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

Yuhan Wang, Zhengxi Lu, Yuchen Yan, Kaitao Song, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang and 1 more

PaperGym turns papers into training tasks for research-plan generation. It derives questions from goals and background and evaluation criteria from methods and experimental designs, aiming to reduce criterion leakage. Rubric-conditioned self-distillation guides token-level learning before group relative policy optimization scores plans. For Qwen3-8B, this sequence raises the five-benchmark average from 33.74 to 38.55; an LLM judge measured leakage at 3.73%.

Key insights

  • Explicitly decoupling the sources for generating research questions and evaluation criteria from scientific papers significantly reduces criterion leakage (to 3.73%), yielding cleaner and more effective training data.
  • A two-stage training curriculum combining OPSD for broad prior distribution and GRPO for targeted policy refinement, leveraging rubrics as both privileged context and reward signals, effectively uses rich feedback for complex generative tasks.
  • Designing comprehensive rubrics that cover both methodological innovation and experimental design is crucial for enabling AI models to generate well-rounded and high-quality research plans.

Summary by alphaXiv.

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