KDD 2026 Highlights: 25 Kuaishou Papers Selected, 3 Oral Presentations

Introduction

KDD 2026 (the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining) will be held in Jeju, South Korea, and is recognized as a top A‑class conference in data mining. The Kuaishou technology team had 25 papers accepted, with three of them selected for oral sessions. These works span core research directions such as generative recommendation, automated bidding, semantic ID learning, multi‑behavior sequential recommendation, and more.

Selected Papers and Highlights

FlowTime: Towards Continuous Generative Watch Time Prediction via Flow‑based Personalized Priors (Oral) – Proposes a continuous generative regression framework that uses a VAE and normalizing flows to model multi‑modal watch‑time distributions. Experiments on public and industrial datasets show significant improvements in watch time, usage duration, and long‑watch rates.

Generative Recommendation for Large‑Scale Advertising (Oral) – Introduces GR4AD, a generation‑based advertising system with a unified semantic ID (UA‑SID), a lazy autoregressive decoder (LazyAR), and a value‑aware supervised learning objective (VSL) with ranking‑guided softmax preference optimization (RSPO). Offline and online tests report up to 4.2% revenue lift.

HOBA: Hierarchical On‑Policy Bidding Agents for Adaptive Online Advertising (Oral) – Presents a three‑level bidding framework that combines LLM‑generated hour‑level constraints, a causal SARSA agent for expert selection, and a pool of PID/MPC/IQL experts. Offline benchmarks and large‑scale A/B tests demonstrate stable gains in CTR and eCPM.

Atomic Intent Reasoning (AIR) – A large‑model‑enhanced recommendation pipeline that extracts fine‑grained “behavior‑intent” atoms offline and performs millisecond‑level intent inference online, yielding 3.4% GMV lift on Amazon benchmarks and 3.6% GMV lift in Kuaishou e‑commerce.

Congrats: Consistent Generative Re‑ranking with Graph‑structured Model – Designs a graph‑based generative re‑ranking model that expands decoding space and explicitly models item dependencies, achieving superior offline metrics and online quality/diversity improvements.

FatsMB: Heterogeneous Multi‑treatment Uplift Modeling for Short‑Video Recommendation – Proposes a diffusion‑based multi‑behavior framework that separates behavior‑agnostic and behavior‑specific preferences, improving recommendation relevance across clicks, likes, and purchases.

GUIDER: Generative User Interest Discovery via Explicit Reasoning with LLMs – Uses hindsight‑guided chain‑of‑thought prompting and token‑adaptive policy optimization to generate explicit interest sets, delivering consistent gains in accuracy, session length, and interest diversity.

HRPO: Hierarchical Residual Policy Optimization for Generative Recommendation – Introduces a residual‑credit credit‑to‑go signal for each token in the semantic ID tree, enabling stable PPO‑style updates and delivering multi‑percentage improvements in IAA revenue.

SaFRO: Satisfaction‑Aware Fusion via Dual‑Relative Policy Optimization for Short‑Video Search – Builds a satisfaction‑aware reward model and a dual‑relative policy optimizer to align short‑term ranking with long‑term user satisfaction, achieving notable CTR and retention gains.

Other notable works include KLAN (personalized landing‑page navigation), OneLive (dynamic generative framework for live‑streaming recommendation), PlatformBid (a unified auto‑bidding benchmark with the new BidFlow method), ProductWebGen (multimodal product webpage generation benchmark), SMES (scalable multi‑task recommendation with expert sparsity), QuaSID (qualification‑aware semantic ID learning), and S²GR (stepwise semantic‑guided reasoning in latent space).

Real‑World Impact

All of the above solutions have been deployed on Kuaishou’s platform, serving billions of daily active users. Reported online improvements include higher watch time, increased ad revenue, better search satisfaction, more effective push notifications, and higher e‑commerce conversion rates. The team emphasizes the bridge between academic breakthroughs and large‑scale industrial practice.

Conclusion

These 25 papers demonstrate Kuaishou’s commitment to advancing AI research and translating it into production‑grade systems. The upcoming KDD 2026 conference will showcase these results, and the team looks forward to further collaborations and innovations.