Shopify's T-RoPE makes rotary embeddings time-aware, lifting sequential recommenders up to 130%
_reachsumit · x · 2026-09-28
Shopify researchers propose T-RoPE, a time-aware rotary position embedding for sequential generative recommenders. Standard RoPE encodes only interaction order; the authors prove it stays time-translation invariant even on timestamps and cannot distinguish seasonal contexts.
T-RoPE rotates attention by real timestamps across multiscale frequency banks, adding learnable temporal coefficients, shifted query alignment, and non-stationary key rotation while preserving the RoPE interface. Across five public benchmarks it ranks first on every metric: +78–130% HR@10 on sparse PixelRec, +8–12% on Amazon Books, and +13–82% over an HSTU + Time RAB backbone on an industrial dataset with 6B+ interactions.
More from Research
- CS papers near 18,000/month, up 10x since 2016, led by complexity and algorithms — burny_tech · 2026-10-25
- ICML Paper: Empowerment-Based Representations Are 1000x More Sensitive to Control-Relevant Features — burny_tech · 2026-10-25
- ETH Zurich teaches G1 robot to climb ladders with just 10 minutes of human data — Distinct-Question-16 · 2026-10-25
- New soft actuator weighs 15g, lifts 400g, costs under $3 to build — IanPritchard · 2026-10-25
- Three different arXiv papers are all named LeWAM, and researchers are begging for mercy — burny_tech · 2026-10-25
- Tinker's StableReinforce reverse-engineered: it's REINFORCE plus score centering — andersonbcdefg · 2026-10-25