Douyin and PKU's STEPS: agentic push filtering deployed at 1B-user scale, RecSys 2026 Oral

jiqizhixin · x · 2026-09-07

Douyin and Peking University present STEPS, a self-triggered agentic push recommendation system accepted as RecSys 2026 Oral.

Traditional push systems either poll at fixed intervals (wasting compute) or blindly trigger the full recall-rank pipeline on every event (50x more expensive than filtering). STEPS instead uses a filtering agent to first decide whether a push is worth sending at all; only approved pushes run the full recall–coarse-rank–fine-rank pipeline, turning push timing into active, context-aware judgment.

STEPS is deployed at full scale on Douyin's 1B+ user platform, with online A/B tests showing gains in user activity.

Original post →

More from Research

Research channel →