Research & Papers

TikTok's New AI Picks Videos 3x Faster — And You'll Watch Longer

⚡Your For You page just got faster and better at guessing what you want.

Deep Dive

X-Rec is a new recommendation approach that learns the recommendation distribution directly in the continuous item embedding space through flow matching, then generates embedding triggers for approximate nearest neighbor retrieval. On a streaming benchmark, it substantially outperformed U2I baselines, matched the retrieval quality of SID-AR methods, and delivered 3.46x higher inference throughput than SID-AR. According to the paper, X-Rec has been deployed as a new retrieval source for a specific vertical content on TikTok, where two consecutive launches yielded significant improvements in both vertical engagement (+4.1484%) and general engagement (+0.0111%).

Key Points
  • X-Rec is a 'retrieval' system — the part of TikTok that grabs a small set of candidate videos before ranking them.
  • It runs 3.46 times faster than the previous top method, meaning less computing cost and quicker responses.
  • TikTok tested it live and saw a 4.15% jump in engagement for the content category it was applied to.

Why It Matters

Faster, sharper recommendations mean less time scrolling past things you don't want — and more of your attention captured.

📬 Get the top 10 AI stories daily