Flowcharts are typically presented as images, driving the trend of using vision-language models (VLMs) for end-to-end flowchart understanding. However, two key challenges arise: (i) limited controllability, since users can only modify input images, and (ii) lack of explainability, since it is difficult to trace VLM errors to specific causes.
We propose TextFlow (NAACL 2025), addressing these issues with two stages: a Vision Textualizer, which generates textual representations from flowchart images, and a Textual Reasoner, which performs question-answering based on those text representations. TextFlow offers three advantages: users can select the type of text representation (Graphviz, Mermaid, PlantUML), or convert it into an executable graph object; it improves explainability by attributing errors to visual or textual processing components; and it promotes modularization, allowing advanced LLMs to be used in the reasoner stage when VLMs underperform end-to-end. Experiments on the FlowVQA and FlowLearn benchmarks demonstrate TextFlow’s state-of-the-art performance and robustness.