IdeaTrail Dataset Captures Full Scientific Ideation Process

Scientific research is a complex, multi-stage workflow encompassing literature search, paper reading, tool use, claim checking, cross-paper synthesis, brainstorming, and iterative writing, rather than a singular act of text generation. Existing resources typically capture only isolated components of this intricate process. This paper introduces IdeaTrail, a novel multi-turn process-trajectory dataset specifically designed for scientific ideation and proposal generation. IdeaTrail records a complete research process, from initial evidence gathering and intermediate artifact evolution to either idea selection or final proposal construction. To ensure the dataset is both realistic and grounded, it employs a Generator-Advisor synthesis loop. The Generator produces the visible trajectory through actions, observations, and artifact edits, while the Advisor, with access to the full generation context, checks for grounding, causal order, naturalness, and leakage from hidden targets. This reverse-to-forward procedure generates multi-turn research data that aligns with real scientific artifacts, effectively approximating the uncertainty, evidence use, and staged convergence inherent in actual research practice. IdeaTrail provides both a valuable dataset and a general recipe for synthesizing process-supervision data for scientific research agents.