AIAny
Icon for item

PaperGym: Rubric-Centered Evolution for Research-Plan Generation

Turns each research paper into a training environment to generate verifiable research plans by synthesizing questions from goals/background and deriving evaluation criteria from methods/experiments. Key features: four-stage extraction that reduces criterion leakage to 3.7%, a two-stage rubric-centered training (self-distillation then GRPO), and the PaperGym-20k corpus with two held-out benchmarks.

Introduction

Why this matters

Research planning is inherently open-ended and hard to evaluate automatically because plans lack single gold answers. PaperGym's core insight is to treat each published paper as a complete environment: synthesize the open-ended question from the paper's goal and background, and extract a multi-criterion rubric from the method and experimental design. This decoupling sharply reduces spurious signal (criterion leakage) and lets training optimize against verifiable, atomic criteria rather than a single scalar reward.

Key Findings
  • Decoupled data pipeline: a four-stage extraction (Research Goal, Background, Research Method, Experimental Design) enforces that inputs and targets come from disjoint sections, cutting criterion leakage to about 3.7% versus 11.9%–34.1% in prior datasets. This makes rubric-based rewards more trustworthy.
  • Rubric-centered two-stage training: first use the rubric as privileged context for a self-distillation teacher to convert sparse criteria into dense token-level guidance; then use the rubric as the reward for GRPO to refine full-plan policies. This schedule outperforms supervised fine-tuning and single-stage variants across model scales.
  • Resources and empirical wins: PaperGym-20k (20k instances across CS, physics, economics) plus two held-out benchmarks (PaperGym-Innov and PaperGym-Design). Models trained on PaperGym show consistent gains (e.g., Qwen3-8B reaches 73.48 on ResearchQA) and win more human comparisons than models trained on alternative rubric datasets.
Who it's for — and trade-offs

Great fit if you evaluate or train LLMs/agents that must propose long-horizon scientific plans, design experiments, or optimize methodological novelty; the dataset and rubric format help convert qualitative research goals into verifiable training signals. Look elsewhere if you need turnkey experiment execution (PaperGym focuses on plan generation and evaluation, not running experiments) or if your domain lacks structured papers with separable goal/background and method/design sections. The pipeline depends on clean paper text (LaTeX sources were used) and on reliably extractable structure, so messy or short-form reports may yield weaker rubrics.

Information

  • Websitearxiv.org
  • OrganizationsZhejiang University, Apple
  • AuthorsYuhan Wang, Zhengxi Lu, Yuchen Yan, Kaitao Song, Wenqi Zhang, Weiming Lu, Jun Xiao, Yueting Zhuang, Yongliang Shen
  • Published date2026/08/31

More Items

Analyzes on-policy distillation for LLM fine-tuning, shows teacher token-level supervision is often noisy and not the main driver of gains, and introduces OPSA, a supervision-free, entropy-adaptive method that suppresses low-probability tokens to improve downstream accuracy.

Trains LLM agents to proactively edit and manage their working context for long-horizon tasks using an expanded toolset (planning, long-term memory, soft offloading) and a fine-grained RL algorithm that identifies critical edits and assigns action-level credit. Improves accuracy while keeping contexts compact on long-context QA and deep search.

Proposes VLAct, a representation-centric continued pre-training method for Vision-Language-Action models that preserves VLM priors and enforces cross-embodiment action semantics to turn limited robot trajectories into transferable visual-action representations; shows strong gains and sample efficiency on multiple VLA benchmarks using modest compute.