Training Reasoning Models to Generate Diverse Scientific Ideas (arxiv.org)

Recent research focuses on enhancing the ideation phase of scientific inquiry by training large language models (LLMs) to generate a diverse range of high-quality scientific ideas. Traditionally, LLMs have prioritized the quality of individual ideas, often neglecting diversity. This study introduces a novel approach where the model, designed to produce multiple structured idea cards from input documents, employs a set-level reinforcement learning objective. Each idea generated is not only evaluated for quality but also for its distinctiveness from previously suggested ideas, ensuring a richer array of concepts. The significance of this work lies in its dual emphasis on quality and diversity, a crucial step for effective idea generation in research settings. The developed 8B parameter model demonstrates superior creative utility, outperforming notable models like Claude Opus 4.8 and GPT-6 Astra in terms of both creativity and uniqueness. A pilot study confirmed that the generated sets were comparable in quality to existing benchmarks while showcasing a marginal increase in unique ideas. Additionally, when subjected to proof-of-concept experiments, the model's executions did not compromise effectiveness, indicating that diversity can be achieved without sacrificing practical applicability in scientific endeavors.