This paper, based on arXiv:2611.05671v1, investigates how vision‑language model (VLM) powered artificial agents perceive and exploit visual information when making decisions under uncertainty. After reviewing key concepts from visualization literacy and decision theory, we designed two complementary experiments:\n\n1. Assess models' ability to perceive uncertainty encoded in visualizations, using pie charts, bar charts, circle charts, and icon arrays;\n2. Test agents' decision performance when only image inputs are available, comparing their behavior to predictions from rational decision theory.\n\nThe experiments were run on several mainstream VLMs (including GPT‑4o, Claude, Gemini, etc.), measuring perception accuracy and decision payoff. Results show that GPT‑4o achieves the highest accuracy (≈92%) in perceiving uncertainty in pie and bar charts, while agents align best with rational theory when uncertainty is presented via circle charts, icon arrays, or textual descriptions. Decision bias grows markedly with bar charts, indicating that visualization format critically affects information decoding.\n\nThe study demonstrates that the way visual data is rendered directly shapes artificial agents' cognition and decision making, suggesting that future VLM training should incorporate multimodal uncertainty annotations to improve reliability in real scientific literature.\n\nReview