PaperGym: Rubric-Centered Evolution for Research-Plan Generation
TLDR
PaperGym turns research papers into training environments for research-plan generation, using rubrics as rewards to improve models across benchmarks.
Reasoning
Strengths include a novel rubric-centered training framework, low criterion leakage, and consistent gains across model scales and benchmarks. Weaknesses include a focus on research-plan generation rather than full scientific discovery, and no evidence of physical or real-world experimentation.
Read-first score
Read-first score 35.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 62.
Field roles
Rank sensitivity
Stability: volatile; rank range: 22.
Keyword Scores
Deep Analysis
Innovations
- PaperGym framework converts each research paper into a complete RL training environment by separating question synthesis from criteria derivation.
- Question/criteria split uses paper structure: questions from research goal and background, criteria from method and experiments, reducing criterion leakage to 3.7%.
- Criteria span methodological innovation and experimental design rather than being compressed into a single scalar per rollout.
- Rubric is used twice: as privileged context for OPSD's self-teacher and as reward for GRPO.
- Releases PaperGym-20k corpus and PaperGym-Innov/PaperGym-Design benchmarks.
Methodology
PaperGym builds training environments from research papers by synthesizing questions from the research goal and background while deriving criteria from the method and experiments. Models are trained with a two-stage rubric-centered schedule: the rubric first serves as privileged context for OPSD's self-teacher, then as the reward for GRPO. Evaluation uses Qwen3-1.7B/4B/8B models and compares against supervised fine-tuning, single-stage training, reverse ordering, and RubricHub Science across five benchmarks and ResearchQA.
Key Results
Criterion leakage drops to 3.7% versus 11.90%-34.10% in existing datasets; PaperGym training improves five-benchmark averages by +5.6, +5.0, and +4.8 points for Qwen3-1.7B/4B/8B. Models trained on PaperGym-20k win 58.1% of three-way comparisons versus 28.2% for RubricHub Science, and Qwen3-8B reaches 73.48 on ResearchQA, above Kimi K2.6.