Abstract

Mobile operating systems dispatch hundreds of notifications daily to active users, and the cumulative interruption cost has become measurable in terms of task completion time, cognitive load, and self-reported frustration. Notification coalescing, the practice of grouping, deferring, or suppressing alerts based on inferred user state and contextual urgency, is one of the primary interface-level levers platforms have available to reduce this cost. Android’s notification channels system and iOS Focus Modes represent current production implementations, but both rely heavily on manual configuration by users who rarely engage with settings menus. This piece examines what automatic, adaptive coalescing systems must model and where current designs fall short.

The Interruption Cost Problem

Research on interruption and resumption in knowledge work, including foundational studies by Gloria Mark at UC Irvine, established that task resumption after an interruption takes on average 23 minutes for complex tasks. Mobile notifications represent a qualitatively different interruption pattern from desktop computing contexts: they arrive in shorter bursts, occur during physical transitions, and compete with embodied attention demands such as walking or driving. A 2024 ESM study of 200 smartphone users conducted through the University of Nottingham’s Mixed Reality Lab found that participants experienced an average of 63 notification-driven attention shifts per day, but considered fewer than 15 of those shifts as justified in retrospect. The gap between actual and retrospectively justified interruptions is the design target for coalescing systems.

What Adaptive Coalescing Must Model

A coalescing system that improves on manual Focus Modes needs at minimum three inference layers. First, it must model task engagement: whether the user is currently focused on an activity that would be disrupted by an alert versus between tasks with available attentional bandwidth. Signals available without intrusive sensing include screen touch frequency, app foreground dwell time, accelerometer motion patterns, and calendar schedule data. Second, it must model notification urgency, which is not simply sender priority but a combination of sender relationship, message content, and time sensitivity. Third, it must model user preference drift, since the same user has substantially different coalescing preferences on a Tuesday morning versus a Saturday evening.

Google’s adaptive notification system, introduced in Android 12 and refined through subsequent releases, uses on-device ML to learn per-app interruption preferences based on whether users dismiss or interact with notifications. The system improves over weeks but does not model task engagement state or urgency content; it learns frequency preferences, not contextual appropriateness.

Interface Design for Coalesced Delivery

Deferring a notification is only useful if the deferred batch reaches the user in a form that is actually processed. Coalesced delivery interfaces must solve a secondary design problem: presenting a summary of held notifications that communicates urgency variance across the batch without overwhelming the user with a wall of items. Current Android notification shade grouping by app is a partial solution but does not convey within-group urgency ordering. iOS summary notifications show a count and a few sender names, which reduces visual complexity at the cost of urgency signaling.

A more expressive model, proposed by researchers at MIT CSAIL in a 2025 workshop paper, would apply extractive summarization to the text content of coalesced notifications before delivery, surfacing a one-sentence digest and flagging the highest-urgency item for immediate action. The approach requires on-device NLP inference but is feasible on current NPU hardware. User studies with the prototype showed 34 percent reduction in time-to-action on high-urgency items compared to conventional batched display, alongside a 22 percent reduction in self-reported interruption annoyance.

Configuration Burden and Default Behavior

The persistent failure mode of coalescing systems is opt-in design. Sophisticated controls buried in settings menus reach the fraction of users motivated to engage with them, leaving the majority on aggressive default notification delivery. Platform vendors face a tension: coalescing by default risks suppressing notifications users would have wanted, generating support complaints and trust erosion. The MIT CSAIL proposal suggests a confidence-gated approach where automatic coalescing applies only to notification types where the model’s urgency prediction confidence exceeds a threshold, falling back to normal delivery when the system is uncertain. This preserves user trust by not silencing potentially important alerts while still reducing interruption volume for the highest-confidence low-urgency categories.

The interaction design of coalescing systems is ultimately a negotiation between system judgment and user agency. Getting that balance right at scale, across billions of users with heterogeneous preferences, remains an open problem that current deployments have not fully solved.