Reproduce “Training Reasoning Models to Generate Diverse Scientific Ideas” 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
Manya Wadhwa, Pranav Narayanan Venkit, Hiroaki Hayashi, Yu Li, Tuhin Chakrabarty, Chien-Sheng Wu
The authors train a model to generate three literature-grounded research ideas together, rewarding each for judged promise and distance from earlier ideas in the set. After supervised fine-tuning and reinforcement learning, their no-think Qwen3-8B scores 0.64 on Creative Utility, a judge-based quality-and-diversity measure, versus 0.38 for Claude Opus 4.8. Its judged source grounding is lower: 0.61 versus 0.98.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.