Position: What Cézanne Knew About Visual Intelligence That Vision-Language Models Miss
Abstract
This position paper argues that vision-language model benchmarks for causal reasoning contain two blind spots. First, benchmarks presuppose temporal constitution, the understanding of time as the medium through which causes bring about effects, without testing it. Second, they provide scaffolding through prompts that may substitute for internalized capability instead of accessing it. A camera records one instant from one position. Human visual intelligence works differently. Post-impressionist painter Paul Cézanne grasped this a century before cognitive science would articulate it. He painted how we see, not what is seen. We draw on his technique to distinguish camera-like processing from Cézanne-like construction of understanding through temporal integration. Preliminary evidence shows systematic disparity between fluent causal text and valid causal structure, and divergent responses to identical scaffolding manipulation. Progress requires benchmarks that go beyond output accuracy to test temporal understanding and scaffolding-invariance.