We introduce MindFlow, a framework that formalizes research idea generation as a Mind Flow. A Mind Flow is represented as a graph composed of modular thinking operators and is modeled by a probabilistic mind supernet. Given a research topic, a controller dynamically samples thinking flows to produce candidate ideas. Optimization is performed via a tournament‑based relative ranking, allowing the controller to progressively favor higher‑quality flows in this open‑ended setting. An evaluation protocol is also proposed, jointly assessing problem finding and problem solving, thus moving beyond title‑ or abstract‑only judgments. Experiments across diverse topics demonstrate MindFlow’s superiority as an explicit, controllable, and optimizable research idea innovator.
Review: MindFlow abstracts ideation into a searchable graph, giving large language models flexible reasoning paths and highlighting the promise of controllable generation.