PaperGym: Cross-Domain Retrieval Boosts LLM Ideation but Not Beyond Random Seeds

A recent investigation published on arXiv (ID 2605.11532) examines whether large language model (LLM) ideation systems gain advantages from specific cross-domain retrieval or merely from varied exposure. The authors present PaperGym, which consists of three phases: the initial tool-enhanced seed extraction using read, grep, and bash in controlled paper settings; the second phase involves cross-domain seed retrieval through paraphrasing across seven machine learning sectors; and finally, method synthesis from the retrieved seeds evaluated by rubric-based judges. While tool-enhanced extraction increased specificity and paraphrase retrieval expanded domain coverage, cross-domain retrieval did not significantly outperform a random diverse-seed control, suggesting that LLM ideation systems may not need targeted cross-domain retrieval for novelty.