STEPS: fully deployed self-triggering agentic push recommendation¶

Fidelity: core mechanism. Public data verify the mechanism only and do not simulate production traffic.

Paper information¶

Field	Details
Paper	arXiv 2608.01949
Organization	ByteDance / Douyin
First public release	2026-08-03 (arXiv v1)
Authors' code	The paper did not provide official or author code when checked on 2026-08-09.
Adapter / method	`steps`
Local reproduction code	`src/auto_research/reproductions/steps/`

Original paper¶

Motivation and method¶

Fixed frequency caps cannot adapt in real time, while periodic polling trades off cost and timing. STEPS jointly decides whether to push and when to wake next. A planning agent uses gated ordinal regression for intervals, an execution agent decides whether to send based on trajectory return, and a lightweight filter controls computation and blocks abnormal plans.

flowchart LR
 A["Student or agent rollout"] --> B["Plan wake interval, execute push, filter, and self-trigger"]
 B --> C["Auditable objective or state update"]
 C --> D["Public mini-suite evaluation"]

Key figure from the paper¶

Figure 3: Core architecture, modules, and connections. From the original paper; copyright remains with the authors.

Core equations¶

\[ p(k|s)=\sigma(b_k-f_\theta(s))-\sigma(b_{k-1}-f_\theta(s)),\quad a_t=\arg\max_aQ_\phi(s_t,a),\quad \tilde a_t=a_t\mathbf1[g(s_t)>\tau]. \]

Paper results¶

Fully deployed on Douyin, which has over one billion users. Online active days rise +0.2843%, push-permission disablement falls -1.9089%, and the filter reduces computation by 79.42%.

Local reproduction¶

Local comparison: The baseline is the shared transition + content scorer. Adding only STEPS gives +66.05% relative NDCG@10.

The run executes ordinal intervals, trajectory utility, filter safeguards, and closed-loop self-trigger scoring. MovieLens-100K, 260 users, 420 items, seed 42: NDCG@10 0.0354 → 0.0588 (+66.05%). Online numbers are paper evidence only.

auto-research reproduce --paper steps --dataset-dir data --seed 42
auto-research evolve --model rankmixer --dataset movielens-100k --direction "探索 steps 的已安装核心算子" --generations 2 --population 4

See the fixed metrics.

Reproduction boundaries¶

MovieLens has no real push permissions, contact costs, or wall-clock wake-up scheduler. It validates decision decomposition only.