Evaluate longer Sequence-Native Training windows

Evaluate week- or month-length user windows for Sequence-Native Training to determine the additional training speedup obtained by sharing behavior-sequence encodings across exposures beyond a single day.

Background

Sequence-Native Training (SNT) groups a user’s exposures within a time window and encodes the user behavior sequence once for all exposures in that window. The paper’s experiments cut windows at day boundaries, which preserves comparability of daily evaluation metrics but limits the amount of encoding reuse.

The authors explicitly note that extending the windows to a week or month increases the measured speedup, but they do not evaluate those configurations. The unresolved issue is therefore the practical and quantitative benefit of using longer SNT windows.

References

Windows cut by week or by month raise the speedup further, and we leave that configuration to future work.

— OneTrans-V2: Unifying Retrieval, Pre-rank, and Fine-rank with One Transformer in Industrial Recommender  (2609.28589 - Cao et al., 23 Sep 2026) in Section 6.5, subsection “Sequence-Native Training” (\S\ref{ssec:efficiency})