Mapping the scientific dependencies behind each major discovery

SciPaths introduces discovery pathway forecasting, a task that requires identifying enabling contributions necessary to realize a target scientific contribution and grounding them in prior work. The benchmark contains 262 expert-annotated and 2,444 silver pathways from ML and NLP papers, each documenting enabling contributions, their roles, and connections to prior work. Evaluation of frontier models shows current approaches struggle with this task, particularly in recovering methodological dependencies, though performance improves substantially when gold enabling contributions are provided—indicating that decomposition quality is the main bottleneck. The benchmark shifts focus from citation prediction and idea generation toward reasoning about scientific causality and prerequisite knowledge.