Kuaishou's PushDualGen boosts play rate 8.5% with interpretable AI recommendations
A lightweight LLM recommender serving 1B users with traceable decisions and 37.7% lower dissatisfaction.
Kuaishou, the Chinese short-video giant, has published details of PushDualGen, a new AI recommendation system now running in production for push notifications to nearly one billion users. The paper tackles a core problem in generative recommendation: LLMs that output semantic IDs (SIDs) can personalize content end-to-end, but their internal reasoning is a black box, making it hard for engineers to trace why a specific item was pushed. Prior work like OneRec-Thinking added chain-of-thought (CoT) before generating SIDs, yet that approach is too inference-heavy for industrial-scale traffic.
PushDualGen instead uses a two-stage generation pipeline: first, the model outputs a semantic ID for the recommended item, then it generates a short, natural-language "copy" that explains the recommendation. The explanation is skippable during inference, giving engineers a cheap, interpretable audit trail while adding minimal latency. This design keeps the model lightweight enough for Kuaishou's massive online push system, unlike full CoT-based approaches.
Online A/B tests show meaningful gains: effective video play rate increased relatively by 8.50%, and user dissatisfaction dropped 37.70%. The authors also note that PushDualGen improves the content ecosystem over the long term by giving more exposure to long-tail videos—niche content that often gets ignored by traditional recommenders. The deployment proves that interpretable generative recommenders can scale to billion-user platforms without sacrificing performance, offering a practical template for other large-scale push and feed systems.
- PushDualGen generates a semantic ID (SID) followed by a skippable natural-language copy, enabling interpretable recommendations without full chain-of-thought (CoT) inference.
- Deployed in Kuaishou's production push system serving ~1 billion users; online A/B test shows effective play rate up 8.50% and dissatisfaction down 37.70%.
- The lightweight design avoids the high inference cost of OneRec-Thinking and increases exposure for long-tail videos, improving long-term content diversity.
Why It Matters
Interpretable generative recommendation at billion-user scale proves transparency can coexist with performance, setting a new standard for industrial push systems.