PaperGym: Rubric-Centered Evolution for Research-Plan Generation

arXiv 2026 58 method

TLDR

PaperGym turns research papers into reinforcement learning environments with rubric-based rewards, improving research-plan generation across models and benchmarks.

Reasoning

The paper introduces a novel framework for research-plan generation and demonstrates clear gains over baselines with multiple models and benchmarks. However, the abstract does not provide details on limitations, real-world deployment, or failure cases, and some broader claims rely on benchmark-specific results.

Read-first score

Read-first score 58, weighted from topical fit, citation, graph, method, reproducibility, and recency signals.

Recency 8%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 25%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,dataset,evaluation,experiment

Topical relevance 42%

48.3

Uses existing LLM keyword relevance scores normalized to 0-100. AI scientist,automated scientific discovery,autonomous research agent,automated research,literature review agent,survey generation,automated experimentation,experiment design agent,AI for scientific research,paper writing agent,research automation,scientific discovery agent

Reproducibility 25%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=dataset

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 28.

Keyword Scores

AI scientist

AI for scientific research

automated scientific discovery

automated research

research automation

scientific discovery agent

autonomous research agent

experiment design agent

automated experimentation

paper writing agent

literature review agent

survey generation

Deep Analysis

Innovations

PaperGym turns each research paper into a complete training environment by exploiting paper structure: questions are synthesized from the research goal and background, while criteria are derived from the method and experiments.
This separation reduces criterion leakage to 3.7%, compared with 11.90% to 34.10% in existing datasets, limiting reward hacking by paraphrase.
A rubric-centered training schedule uses the rubric twice: first as privileged context for an OPSD self-teacher, then as the reward signal for GRPO.
The work releases the pipeline, the 20,000-instance corpus PaperGym-20k, and the benchmarks PaperGym-Innov and PaperGym-Design.

Methodology

PaperGym constructs training environments from individual papers by synthesizing the question from the research goal and background and deriving evaluation criteria from the method and experiments, with criteria spanning methodological innovation and experimental design. Training follows a two-stage rubric-centered schedule in which rubrics serve first as privileged context for an OPSD self-teacher and then as the reward for GRPO. The approach is evaluated on Qwen3-1.7B/4B/8B against supervised fine-tuning, either stage alone, the reverse ordering, and RubricHub Science, using five-benchmark averages, three-way win rates, and ResearchQA.

Key Results

Criterion leakage falls to 3.7%, versus 11.90% to 34.10% in existing datasets. The proposed schedule improves five-benchmark averages by +5.6, +5.0, and +4.8 points across Qwen3-1.7B/4B/8B; models trained on PaperGym-20k win 58.1% of three-way comparisons against 28.2% for RubricHub Science, and the trained Qwen3-8B reaches 73.48 on ResearchQA, above Kimi K2.6.

Resources

Paper