Generative Recommendation at Kuaishou: Systematic Evolution and Salon Highlights

Generative Recommendation Evolution at Kuaishou

As generative techniques deepen into recommendation, search, and advertising systems, Kuaishou’s architecture is shifting from isolated model innovations to a unified base, infrastructure, and generation pipeline. The salon explored how large‑model capabilities can be integrated into core industrial workflows while balancing effectiveness, efficiency, and engineering scalability.

1. OneReason: When Recommendation Systems Learn to Reason

Kuaishou’s OneRec series demonstrated scaling benefits of next‑token prediction, but complex user intents and long‑term interest dynamics exposed capability limits. To move beyond pure generation, the OneReason project introduced a chain‑of‑thought (CoT) data pipeline tailored to recommendation tasks, enabling models to perform genuine reasoning. The team also built a reasoning data flywheel in the reinforcement‑learning stage, iteratively improving CoT quality and achieving substantial gains in reasoning ability for recommendation scenarios.

2. Pool‑Rec: Heterogeneous Compute Pooling for Recommendation Estimation

Scaling generative recommendation models demands higher compute efficiency. Kuaishou’s computing‑engine lead presented Pool‑Rec, a system that unifies heterogeneous GPU resources into an elastic pool. By upgrading infrastructure, scheduling, and inference engines, Pool‑Rec achieves unified, elastic provisioning of AZ‑level resources, markedly improving model‑compute utilization (MFU) and supporting large‑scale deployment of OneRec models.

3. Industrial‑Scale Generative Search & Decision

In e‑commerce, user queries are often ambiguous and context‑rich, causing information loss in traditional cascade search. The generative search framework leverages stronger semantic understanding to model complex intents and item matching. Building on OneSearch V1/V2, Kuaishou outlined a roadmap from query generation to item generation, then to internalized reasoning and decision‑making, emphasizing interest orchestration, unified search‑recommendation pipelines, and closed‑loop main‑chain integration.

4. GR4AD: Generative Advertising Recommendation from Token to Revenue

Advertising recommendation must align model understanding with commercial goals such as conversion efficiency and real‑time feedback. GR4AD extends the OneRec decoder‑only architecture with multi‑modal semantic principal‑component IDs for richer ad and user representations. It introduces Lazy Auto‑Regressive (Lazy AR) decoding and value‑aware learning to better align model outputs with revenue metrics. A dedicated real‑time training and inference service ensures stable, large‑scale online deployment, delivering noticeable business gains on Kuaishou’s ad platform.

LLM‑Rec Challenge

Kuaishou and ACM SIGIR 2026 jointly launched the "Kuaishou Explorer LLM‑Rec Challenge" (official site: https://ks-llmrec.streamlake.com/). The competition invites full‑time university students worldwide to combine recommendation models with large language models, aiming to accelerate the deep fusion of LLMs and recommendation technology and move promising ideas from research to production.

Conclusion

From OneReason to Pool‑Rec, and from OneSearch V2 to GR4AD, the salon showcased Kuaishou’s roadmap: scaling generative models toward reasoning‑enhanced capabilities, continuously optimizing training and inference efficiency, and industrializing generative AI across recommendation, search, and advertising. The LLM‑Rec challenge further invites the community to co‑define the next generation of intelligent recommendation systems.