Research Progress Graph: modeling authentic research trajectories, 35 agents found incomplete

ZhiyuChen4 · x · 2026-10-22

A new arXiv paper introduces the Research Progress Graph (RPG), a representation capturing how scientific ideas actually evolve from initial question to published discovery.

Problem: the idea in a paper is rarely the initial one; the intermediate loop of experiments, insights, challenges, literature reading and revisions is almost never recorded, and existing LLM research agents keep only the idea text while discarding the evidence and reasoning linking versions.
Representation: an RPG is a graph of typed research states (research question, idea update, insight, challenge, literature search, literature, code/experiment plan, experimental result), with transitions annotated by temporal index and reasoning.
Findings: the authors present 8 authentic RPGs justified by their authors and analyze 35 existing research agents — each is an incomplete RPG; none keeps the changed state, the evidence, and the linking reasoning in one record.
Implications: agents can predict the next research state, ideas can be evaluated as trajectories rather than final texts, and authentic trajectories can help build agents that emulate human researchers.

Original post →

Research Progress Graph: modeling authentic research trajectories, 35 agents found incomplete

More from coding & agent