Beyond the Post Hoc User Study: Modeling Visual Decision-Making with Active Inference

Bridging the Mechanistic Gap in Visualization

Empirical user studies are the gold standard for evaluating data visualizations, yet they are often post hoc—they measure performance after a design is finished rather than explaining the cognitive mechanisms that lead to success or failure. This paper addresses this gap by introducing a computational approach based on Active Inference, a probabilistic framework that models perception and action as a process of minimizing uncertainty and surprise. By treating chart reading as a dynamic, goal-oriented search, the authors move beyond static performance metrics toward executable simulations that can predict how human attention, memory, and bias interact with visual encodings.

Modeling Dual-Process Cognition

The authors implement a dual-process theory of decision-making, distinguishing between Type 1 (Fast, heuristic) and Type 2 (Slow, analytic) processing. They instantiate these as two distinct Active Inference agents tasked with estimating the average of two bars in a chart. The Fast agent relies on quick, heuristic visual impressions, making it susceptible to tick-salience bias. The Slow agent engages in more effortful, sequential analysis, which makes it vulnerable to working-memory decay as it attempts to hold intermediate values. By formalizing these strategies, the model generates inspectable cognitive traces, including fixation sequences and evolving belief uncertainty.

Toward Predictive Design

This architecture provides a scaffold for a more generative science of visualization. Rather than replacing human subjects, these synthetic agents allow researchers to formalize and test hypotheses about how specific visual features trigger cognitive failures. The authors propose that future empirical studies—using eye tracking, response times, and error distributions—can be used to parameterize and refine these models. This shift toward in silico evaluation could allow designers to anticipate and mitigate potential interpretation errors early in the design process, rather than discovering them only after expensive user testing.