Daily Paper Briefing

Abstract Overview

PaperGym is a framework that converts scientific papers into training environments for research-plan generation. It separates task construction from rubric derivation by synthesizing questions from a paper's research goal and background, while generating evaluation criteria from the method and experimental design sections. This decoupling is designed to prevent criterion leakage, where models earn rewards simply by restating the prompt. Using a 20,000-instance corpus across computer science, physics, and economics, the authors train models via a two-stage schedule that applies rubric-conditioned self-distillation followed by reinforcement learning with rubric-based rewards.

Novelty

The paper introduces a rubric-centered pipeline that exploits scientific paper structure to decouple prompts from evaluation criteria, lowering criterion leakage relative to prior datasets. It also establishes a two-stage training scheme that uses the exact same rubric first as privileged context for on-policy self-distillation (OPSD) and subsequently as the reward verifier for group relative policy optimization (GRPO).

Results

PaperGym-20k achieves a criterion leakage rate of 3.73%, compared to 11.90%–34.10% in existing datasets and benchmarks. The OPSD+GRPO training schedule outperforms supervised fine-tuning and single-stage baselines, boosting five-benchmark averages across Qwen3-1.7B, 4B, and 8B by +5.56, +5.04, and +4.81 points, respectively. Under identical training conditions, models trained on PaperGym-20k win 58.1% of three-way comparisons against 28.2% for RubricHub Science, and the fine-tuned Qwen3-8B reaches 73.48 on ResearchQA, surpassing Kimi K2.6 at 73.19.

Key Points

1. PaperGym constructs research questions and evaluation rubrics from disjoint sections of scientific papers to minimize prompt–criterion leakage.
2. The authors release PaperGym-20k alongside two held-out benchmarks, PaperGym-Innov and PaperGym-Design, to evaluate methodological innovation and experimental design separately.
3. A two-stage post-training pipeline applying rubric-conditioned OPSD before rubric-rewarded GRPO yields consistent gains across Qwen3 model scales and out-of-domain benchmarks.

References

• arXiv: https://arxiv.org/abs/2608.31119v1
• Fugu-MT: https://fugumt.com/fugumt/paper_check/2608.31119v1
• Hugging Face Papers: https://huggingface.co/papers/2608.31119
• GitHub: https://github.com/ZJU-REAL/PaperGym
• Project: https://zju-real.github.io/PaperGym