When and Whether to Push: Douyin & Peking University’s Agentic STEPS System Wins RecSys 2026 Oral

Douyin and a Peking University research team proposed STEPS (Self‑Triggered End‑to‑end Agentic Push System), a self‑triggered push recommendation framework that treats each notification as a closed‑loop decision problem. The paper was accepted for an oral presentation at RecSys 2026 Industry Track and has been fully deployed to over 1 billion Douyin users.

Limitations of Traditional Push Systems

Conventional push pipelines rely on two passive modes: fixed‑interval polling and pre‑scheduled delivery. Fixed‑interval polling wastes compute when intervals are short and misses optimal windows when intervals are long. Pre‑scheduled delivery models user‑specific timing but cannot react to real‑time context, requiring heuristic adjustments and still operating as a static pipeline. Both approaches suffer from a fundamental bottleneck: the system is a passive executor lacking autonomous decision‑making.

Core Idea of STEPS

STEPS introduces three collaborating agents that enable the system to "wake itself up" and make autonomous decisions about when to trigger, whether to send, and how to allocate compute resources.

Planning Agent : predicts the next optimal wake‑up interval using a Decision Transformer backbone. It incorporates a gated‑RTG mechanism that multiplicatively modulates state representations with the Return‑to‑Go (RTG) signal, preventing the model from ignoring this noisy condition.

Execution Agent : upon waking, evaluates real‑time context and candidate content to make a binary send‑or‑not decision. It treats a user's interaction history as a trajectory and uses Bellman‑based value‑guided learning to weight training samples by long‑term value, aligning decisions with cumulative reward rather than short‑term clicks.

Filtering Agent : acts as a lightweight gate that dynamically decides whether to enter the expensive ranking pipeline. It reduces compute by up to 74.88% overall (85.65% for high‑frequency users) and serves as a safety guard against unreasonable wake‑up patterns.

Training and Evaluation

Offline training used six months of production logs covering the full user base. Online A/B testing ran for 14 days with random user‑device assignment, comparing STEPS against a pre‑planned production baseline and a resource‑matched fixed‑interval baseline.

Results: User Active Days increased by 0.2843%, push‑permission disable rate dropped by 1.9089%, and total compute consumption fell by 79.42%. Ablation studies showed the Planning Agent contributed most to activity and experience gains, the Execution Agent amplified those gains, and the Filtering Agent alone accounted for the majority of compute reduction.

Key Findings

The gated‑RTG mechanism successfully forces the model to attend to the RTG condition, as evidenced by a correlation‑ratio shift away from 1.0.

Push intervals became more reasonable: ultra‑short intervals (0‑20 min) decreased by 35.93%, and the distribution shifted toward longer intervals.

Push frequency balanced across users: the proportion of users receiving 1‑20 pushes per day rose by 20.31%, while high‑frequency users (>30 pushes) fell by 9.30%.

Dynamic branching reduced unnecessary ranking work by up to 85.65% for full‑attendance users, reallocating compute to higher‑value requests.

Conclusion

STEPS demonstrates a paradigm shift from passive response to an autonomous closed‑loop system, embodying three agentic traits: self‑planning, dynamic branching, and self‑evolution via reinforcement learning. This aligns with broader trends in Agentic AI, moving systems toward self‑triggered intelligent loops.