Proceedings of the Annual Meeting of the Cognitive Science Society

Driving assistance systems require not only predicting pedestrian-side crossing intent, but also estimating how drivers interpret that intent from ego-view evidence. Existing pedestrian intent prediction methods often rely on one-hot labels or binary crossing probabilities, which cannot capture observer variation or intermediate interpretation states. We introduce \emph{Temporal Prediction of Driver-side Interpretation Distributions} (TP-DID), a causal frame-level task for predicting empirical distributions of driver-side interpretations towards pedestrian intent. To capture temporal distribution changes, we propose \textsc{Input-Modulated Mamba} (IM-Mamba), a causal state-space model that reweights current visual evidence before temporal state prediction. Experiments on PSI show that IM-Mamba improves distribution prediction, especially for observer variation, awareness-related states, and temporal changes.