ReasFlow: AI Agents Aid Mathematical Scientific Discovery

While AI agents have advanced in empirical scientific discovery, theory-driven fields like applied mathematics, which demand rigorous proofs and synthesis of domain knowledge, remain largely unexplored by automated systems. This paper introduces ReasFlow, an end-to-end autonomous agent system specifically designed for reasoning-centric scientific discovery. ReasFlow operates on a collaborative model, positioning the human expert as a Principal Investigator who guides an agent capable of executing rigorous derivations, much like a graduate student. The system features a robust internal verification loop that audits logical coherence and corrects errors before human review. It also includes an automated knowledge retrieval and self-improvement mechanism that proactively surfaces both declarative facts and overlooked procedural heuristics, significantly reducing the need for expert intervention. The system unifies various research stages, including literature synthesis, algorithm design, theorem proving, experimentation, and manuscript preparation. ReasFlow was deployed to autonomously generate five complete research papers with theoretical and empirical content from minimal prompts, consistently achieving high evaluation scores against state-of-the-art baselines. This work, publicly accessible via the ReasLab platform, provides a collaborative workspace for AI-assisted theoretical research.