Theory Building, Evaluation, and Rationality

Model comparisons are central to evaluating different accounts of cognitive and neural processes. In complex modeling frameworks such as Dynamic Causal Modeling, candidate models often differ in their complexity, allowing more complex models to achieve better fits. Bayesian model comparison methods explicitly penalize model complexity, aiming to balance explanatory power against model flexibility. However, previous research has raised concerns about whether these penalties sufficiently compensate for different levels of complexity under all conditions, and to what extent model comparison outcomes reflect explanatory validity rather than advantages arising from increased flexibility. Here, we formalize a procedure to characterize the impact of model complexity on model fit and model comparison outcomes. By randomly generating model structures within predefined complexity classes, we estimate the baseline selection advantage attributable to flexibility. Our results demonstrate that increased complexity can systematically bias both model fit and model comparison outcomes, increasing the probability that more flexible models are selected even when they do not better reflect the true, underlying structure. These findings suggest that standard complexity penalties might not under all circumstances account for complexity-dependent biases. To address this limitation, we propose a general evaluation framework that explicitly characterizes complexity-related effects. Model comparisons are then interpreted in terms of the advantage each model shows over the fit of complexity-matched random structures. This approach enables distinguishing between genuine explanatory improvements and gains attributable to structural flexibility alone.

This is an in-person presentation on July 18, 2026 (09:00 ~ 09:20 EDT).

Users navigating hierarchical information spaces exhibit trial-and-error behavior, such as backtracking and revisits. Information Foraging Theory explains these patterns through information scent, traditionally understood as an estimate of the value of pursuing a link based on semantic cues. However, existing formulations often treat scent as a static, local evaluation, failing to account for how these estimates evolve under cognitive constraints during sequential decision-making. We propose a resource-rational account that conceptualizes information scent as a belief formed under limited cognitive resources. In this framework, "resources" refer to working memory capacity, the temporal decay (forgetting) of information traces, and the cost of sequential inspection. We argue that trial-and-error navigation behavior is not irrational noise, but as a resource-rational adaptation to limited memory and costly information acquisition. We formalize this account in a computational framework that implements scent as noisy belief evidence subject to capacity limits and forgetting. Simulations reproduce established difficulty, hierarchy-depth, and position effects. Furthermore, manipulating the effective planning horizon systematically alters structural metrics such as lostness, linking navigation organization to sequential control. Together, these findings provide a theoretical reinterpretation of information scent as resource-constrained belief formation.

This is an in-person presentation on July 18, 2026 (09:40 ~ 10:00 EDT).