The idea presented in a published paper is rarely the project's original conception. Between them lies a closed‑loop process where researchers run experiments, extract insights, pinpoint challenges, survey literature, and revise the idea, each step driven by explicit evidence and reasoning. We call the path an idea follows through this loop, together with its supporting evidence and reasoning, an authentic research trajectory, yet it is seldom recorded. Current LLM‑based research agents typically generate an idea once, then refine it against reviews, literature, or experimental outcomes, or search over candidate ideas and experiments. In these workflows the textual idea evolves while the evidence and reasoning linking successive versions are discarded. To address this we introduce the Research Progress Graph (RPG), a representation that captures the evolution of an idea as typed research states linked by cognitive transitions, each annotated with a temporal index and the underlying reasoning, making the evolution an explicit object of study. We ground the RPG in authentic trajectories whose steps are justified by authors in annotations and papers. Mapping the outputs of thirty‑five research agents onto RPGs reveals that each is an incomplete RPG: none retains the changed state, the evidence that caused the change, and the reasoning connecting them as a single evolving record. Modeling ideation with RPGs reshapes how agents are designed, evaluated, and trained: agents predict the next research state instead of a complete idea, ideas are assessed as trajectories rather than final texts, and authentic trajectories serve directly as training data. These shifts move research agents from generating static ideas to modeling how ideas evolve.
Review