Research & Papers

TikTok's New AI Boosts Its Shopping Feed Sales by 6%

⚡Your For You page just got better at guessing what you'll buy.

Deep Dive

TikTok researchers built a new recommendation architecture called HELIX, and in online A/B tests it delivered an approximately 6% increase in e-commerce video GMV per user.

Their starting point: industrial recommendation ranking models typically scale along two modeling axes — feature interaction over heterogeneous user, item, context, and cross features, and sequence modeling over long, informative, multi-type user behavior histories. But scaling either in isolation is insufficient, the authors found, because each exhibits a limited scaling ceiling and a suboptimal scaling-law slope. They conjectured that a more favorable scaling-law slope requires jointly scaling both axes.

HELIX is their answer: a purified and unified architecture that interleaves sequence retrieval and feature interaction while enforcing one-way information flow from reusable sequence states to candidate-conditioned mix-tokens. That design preserves cross-depth communication between the two modeling axes while keeping user-side sequence computation amortizable, which enables flexible and asymmetric scaling of sequence modeling and feature interaction.

Deployed in TikTok's e-commerce recommendation system, HELIX consistently improved offline CTR AUC, CVR AUC, and other ranking metrics, according to the paper.

Key Points
  • TikTok's new recommendation AI, HELIX, raised shopping spending per user by roughly 6% in company-run live tests.
  • It works by combining two jobs at once: matching products to your profile, and reading your long history of clicks and buys.
  • The gain comes from selling more to each viewer, not from adding more viewers — a reminder that feeds are tuned to predict you.

Why It Matters

Your social feeds will get better at selling to you — meaning more impulse buys and less scrolling past ads.

📬 Get the top 10 AI stories daily