A new method called T-RoPE teaches recommendation algorithms to actually track time, not just click order.
Large recommender systems have borrowed the Transformer architecture from large language models, including Rotary Position Embedding (RoPE), which encodes how far apart two tokens are in a sequence. But recommenders do not process text - they process click and purchase histories, where position is just an event counter, blind to whether two clicks happened a minute or a year apart. In a paper posted to arXiv this week, researchers propose T-RoPE, which swaps that index-based rotation for one built on actual timestamps, adding learnable coefficients, multiple time scales, and adjustments for irregular gaps between events. The authors report it beat the strongest baseline on every metric across five public benchmarks, including a 78-130% jump in one accuracy metric on a sparse dataset called PixelRec, and improved results by 13-82% on an internal e-commerce dataset with more than 6 billion interactions.
Recommendation systems increasingly borrow the same generative machinery as chatbots, so small architectural mismatches, like a position encoding designed for word order, carry through and quietly shape real products. If timestamp-aware rotation generalizes, it is a cheap fix, the authors say the added compute cost is linear, that plugs into existing Transformer-based recommenders without a redesign. The paper also reports an online A/B test in what it calls the Shop app, showing a 0.33% lift in conversion and 0.63% more orders, modest by percentage but potentially large in absolute revenue at that scale.
These are the authors' own numbers from an unpublished, non-peer-reviewed preprint, so outside replication should come before anyone rewrites a production recommender around them.