Data agents built on ReAct‑style loops can now run complex analytical workflows over structured data, yet their intermediate reasoning still proceeds almost entirely through text. Human analysts, by contrast, move among numbers, tables, and charts, using visual evidence to expose trends and anomalies that textual summaries miss. We investigate whether charts can serve as internal epistemic actions within a data agent’s reasoning loop and how to teach an agent to employ them.
We first construct \probe{}, a suite of 200 controlled tasks spanning four matched interface conditions. Experiments reveal a persistent gap between what agents can read from charts and how often they actually choose to use them. Although agents score as well from self‑generated charts as from numeric output, mere access to plotting does not guarantee consistent usage. Task slices show charts excel at temporal, pattern, and anomaly reasoning, while text leads in precise reading and conditional verification.
To address this, we propose \method{}, which distills two families of experience from conditioned trajectory rollouts. One family records how to select and verify analytical operations; the other records when to invoke text, charts, or both. A trajectory judge scores outcomes and labels intermediate steps, and a validation gate admits only experience that improves held‑out performance.
Across three held‑out benchmarks and four LLM backbones, \method{} raises analytical accuracy and the quality of visual evidence compared to matched baselines. Even backbones that never drew charts begin to use them, all without touching model parameters.
Review: This work demonstrates the feasibility of embedding visualization as an epistemic tool inside LLM reasoning loops and provides a systematic experience‑distillation framework, offering a promising direction for enhancing multimodal reasoning in data‑analysis agents.