Using LLMs to Model Notification Timing from Context-Sensitive Features

Managing user attention to find opportune moments for interruptions is challenging. Recent works have begun to address the ability of multimodal large language models to detect user interruptibility and schedule communication. However, existing works focused on binary decisions, less frequent context switches, and did not work in real-time. We present an LLM-based background process that aims at detecting interruptibility and scheduling notifications in office working environments based on diverse visual information and interaction parameters. A user study with (N=21) indicated that the system could not fully achieve its intended potential, since interruptibility is a complex interplay of dynamic and highly individual factors. Still, we demonstrate how subsequent experiments with different modeling approaches can improve performance. Notably, providing the model with the history of users’ previous ratings led to the strongest improvements, while removing all contextual features resulted in the best overall alignment. These findings highlight strong temporal effects in self-reported interruptibility and suggest limits of multimodal inference alone. We conclude by discussing the implications for the design of hybrid attention management systems.