New study: Your reading highlight style is a stable trait, not a fleeting state
Highlighting patterns stay consistent for 24+ months, even across different topics.
A new arXiv preprint from researchers Kazuki Nakayashiki and Keisuke Watanabe challenges the assumption that reading behaviors are temporary. The study, titled "Trait, Not State: The Durability of Reading Identity in Social Highlighting," analyzed social highlighting patterns from heavy, long-tenured readers over a 24+ month period. The key finding: a person's highlighting style — which documents and passages they select — remains highly stable over time, acting more like a personality trait than a fluctuating state. Specifically, the fine-layer advantage (comparing a reader's profile to others within their interest neighborhood) showed no statistically detectable decline at any horizon out to 24 months. The 6-12 month retention correlation was R=1.00 (95% CI [0.85, 1.18]), and even the farthest bin was compatible with only a modest decline.
Crucially, this stability is not simply due to repeatedly highlighting the same domains. About 90% of the signal survived when excluding all sources from the reader's profile. Within-person drift was slow — a recent-half profile outperformed the old half by only +0.042. Prospectively, a personal profile built from a reader's earliest documents (median 20 months before evaluation) ranked their subsequent reads at roughly 3x the average precision (AP) of any simple non-personal prior tested. The authors operationalize "trait" as a stable signature under continued engagement, noting the scope is limited to heavy users of one platform and that exposure choice is inseparable from selection. Still, the findings have significant implications for recommendation systems: your reading identity is remarkably durable.
- Highlight signatures show no detectable decline over 6-12 months (R=1.00, 95% CI [0.85, 1.18], n=212).
- 90% of the personal signal persists even when removing all repeated domains from the profile.
- Personal profiles built from early reads (median 20 months prior) rank future reads at 3x the AP of non-personal baselines.
Why It Matters
For recommendation engines: your reading identity is predictable long-term — personalization can rely on stable traits.