acceptodds
Under review as a conference paper at ICLR 2027

Context Duration in Naturalistic Smartphone Use: Evidence for Millisecond-Scale Processing

Abstract

Smartphone use is typically reported as hours of daily screen time, and content is still designed for second-to-minute attention. Recommendation and content systems have become good at predicting what a person might want, yet people often do not act on what they are shown. We call this the recommendation–action gap, and we show that it has a timescale. Using continuous on-device telemetry from 35 participants over 15 days (3.2 million screen touches; keyboard input excluded; logging only while the device was unlocked), we treat inter-touch intervals as a behavioral proxy for context duration, the time a user remains engaged with a given on-screen context. Relative to a 2022 measurement, mean daily use rose 36% (to 4 h 53 min) and daily touches rose 19% (to 5,384), yet aggregate context duration contracted 23% to 3.8 to 7.8 s, orders of magnitude below television-scale engagement (5 to 21 min). The inter-touch distribution is strongly right-skewed. Half of all touches arrive within 266 ms of the previous touch, and 75% within 1.06 s. Half of all interaction therefore sits inside the 300 ms window in which people decide to approach or avoid, well before the plays, add-to-carts, and orders that systems are trained to predict. Time on the device is accumulated through revolving-door sessions rather than sustained attention (Instagram: median 67 s sessions summing to 75 min/day; WhatsApp: median 23 s sessions summing to 38 min/day), and only the extreme tail of context duration approaches long-form media. We repeat this analysis across social media, streaming, music, food and quick commerce, and AI assistants. Decision windows differ by category: 7 to 14 s for what to view on Instagram, 4 to 7 s for what to reply to on WhatsApp, about 10 to 30 s for weekday streaming choice, 4 to 9 s for food orders, and 4 to 9 s on average across AI assistants. In every category the choice is made within seconds. Closing the recommendation–action gap needs content that can be processed in milliseconds and acted on within the measured window of each category.

Primary area:
applications to neuroscience & cognitive science
Keywords:
smartphone interaction, context duration, naturalistic behavior, rapid decision-making, inter-touch intervals, millisecond-scale content

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_PcnYOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The manuscript investigates the temporal patterns of smartphone use and argues that current recommendation systems don't operate a low millisecond time scale hence do not capture users decision processes. The manuscript includes data from 35 andriod users colelcted over 15 days leading to 3.2 million touches with significant representation of short millisecond touch.

Critical strengths

The paper addresses an interesting and practically relevant question about the temporal granularity at which recommendation systems should operate. The large number of recorded touch interactions could provide a useful empirical characterization of smartphone usage patterns across application categories.

Critical weaknesses

The authors argue to building new recommendation systems that account for this short time-scale touches on the mobile devices. It reads more like a position paper supported by data statistics. With that, in the manuscript's current state, I don't see how it would contribute to the ICLR and general ML/AI community. The work can be improved with a few exemplary tasks that this new data and insights help to solve, if indeed existing recommendation modesl fail at this temporal granularity (which has yet to be shown I believe).

Questions for authors

The work can be improved with a few exemplary tasks that this new data and insights help to solve, if indeed existing recommendation modesl fail at this temporal granularity (which has yet to be shown I believe).

The work seems to be a contribution heavily related to data. However, there is currently not a lot of details on how the data are recorded. How is the data handled, where is it stored, can it be reused? Without these information and the empirical tasks demonstration, the connection and value of short time-scale touches remain largely speculative.

As compared to 2022, the method in collecting the data remained the same as the authors claimed. So what was the conceptual novelty? It remains an open question what computational insights are gained beyond the descriptive stats.

Reviewer_Uq4KOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces the concept of recommendation-action gap in the context of smart device usage. They report different distributions for inter-touch intervals. The authors suggest ways to close the recommendation-action gap.

Critical strengths

The paper provides a range of average engagement times for different communication context on mobile phones for the given population. The authors made great effort in collecting data and analyzing the data

Critical weaknesses

The research question is “what is the temporal unit of engagement”. Related on work on measuring cognitive engagement should be surveyed more extensively. Also, a direct comparison with existing work in terms of numbers is not valid given different context, in particular the demographics of participants. Although this aspect is stated in the limitation section, the authors should be more careful in what they claim as the paper contribution.

More detailed demographics about users should be reported. Age alone does not tell readers much about the context of data collection.

Questions for authors
  1. Authors state that systems have good prediction on what people want, but people act differently. Is this a bad thing? Why do we need to close the recommendation-action gap?
Reviewer_w7g4OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces a concept that the “recommendation–action gap” is a mismatch in timescales and proposes setting the windows used for content design and recommender-system training according to measured durations for each category. The authors collected logs from 35 Android users in India covering 3.2 million screen touches over 15 days and use the interval between touches as an approximation of “context duration.”

Critical strengths
  • Good angle in terms of connecting the recommendation system and the user's behavior.

  • Real data collection. Collected data from 35 users across 15 days of using different apps.

  • Honest about limitations.

Critical weaknesses
  • There is no model, algorithm or experiment. And there are no evaluations of labels for the next screen touch. The claim is thus very weak.

  • The dataset is very small-scale. It was collected from only 35 people in India, where people might share similar habits of using Android phones.

  • Data is sparse. Some apps only have 1-2 touches per day.

Questions for authors
  • Can you evaluate on public datasets to support your claim?

  • Can you scale to other countries and other age groups?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.