PushGen: Push Notifications Generation with LLM

TL;DR

PushGen tackles the challenge of scalable, engaging push notifications by combining a category-controllable, style-conditioned generator with a pairwise reward model trained on A/B data to optimize CTR. The two-stage framework generates diverse, evidence-grounded content and selects it through a CTR-aligned ranking objective, mitigating issues of style control and ex-ante quality estimation. Across offline and online experiments, PushGen delivers substantial CTR gains and high replacement rates of human-generated content, and has been deployed at a large scale. The work also discusses reward-hacking risks and outlines directions including VLM-backed backbones and reward-driven RL with safeguards.

Abstract

We present PushGen, an automated framework for generating high-quality push notifications comparable to human-crafted content. With the rise of generative models, there is growing interest in leveraging LLMs for push content generation. Although LLMs make content generation straightforward and cost-effective, maintaining stylistic control and reliable quality assessment remains challenging, as both directly impact user engagement. To address these issues, PushGen combines two key components: (1) a controllable category prompt technique to guide LLM outputs toward desired styles, and (2) a reward model that ranks and selects generated candidates. Extensive offline and online experiments demonstrate its effectiveness, which has been deployed in large-scale industrial applications, serving hundreds of millions of users daily.

Figures (4)

• Figure 1: Push Notifications at our Application.
• Figure 2: The overall framework of PushGen, where the <Other_Push> indicates the DeepSeek V3 671B + RAG Push results.
• Figure 3: Loss curves of different LLMs during training.
• Figure 4: Reward model classification ability analysis.