IdeaGraph: Evidence-Bearing Synergy Junctions for Cross-Field Scientific Ideation

Large language models can propose research ideas that experts judge novel, but a recent execution study reports that this advantage fades once the ideas are carried out: the bottleneck of machine-generated proposals is not novelty but feasibility. As end-to-end AI scientists scale up idea production, this gap between ideation and execution becomes a meta-scientific measurement problem: research-idea quality must be assessed before execution resources are committed. We argue that scientific ideation systems should therefore be evaluated not only by the novelty of their proposals but also by whether each proposed transfer carries inspectable evidence from both contributing research lineages. IdeaGraph introduces evidence-bearing synergy junctions: graph junctions that connect research lineages from different fields, are supported by both source corpora, and retain their surrounding typed-citation paths as inspectable generation context. The three junction types are shared research entities, induced upper topics supported by both corpora, and papers cited from both. In pairwise comparisons over 30 target papers, IdeaGraph improves judged Feasibility over target-only prompting on 27 of 30 while staying on par in Overall quality, and surpasses Chain-of-Ideas on all five criteria; a blinded ranking by ten researchers reproduces the criterion-level profile. On the seven targets with qualifying follow-up work, the proposals also align best with where each target's citation lineage later went. These results indicate that evidence-bearing synergy junctions and their bilateral literature provenance support pre-execution quality control of machine-generated research directions.