How Kuaishou’s 8 Groundbreaking Papers Are Shaping AI at KDD 2025

KDD (ACM SIGKDD Conference on Knowledge Discovery and Data Mining) is the top international conference for AI data‑mining. The 2025 edition will be held in Canada from August 3‑7. Recently, Kuaishou had eight high‑quality papers accepted, focusing on recommendation systems, multi‑task learning, multimodal large models, large language models, and combinatorial optimization.

Paper 1: VLM as Policy – Common‑Law Content Moderation Framework for Short Video Platform

Link: https://arxiv.org/abs/2504.14904

Abstract: Short‑video platforms face unprecedented challenges in content moderation due to exponential growth of content. Traditional moderation suffers from high labeling cost, human bias, and difficulty adapting to hot topics. Kuaishou proposes KuaiMod, a multimodal large‑model‑based solution with three contributions: (1) a benchmark for low‑quality short‑video classification covering 4 major and 15 fine‑grained categories; (2) an automated content judgment pipeline using a two‑stage training strategy (Tag2CoT and CoT2Tag) that achieves accuracy comparable to human review and reduces user‑report rate by 20%; (3) a reinforcement‑learning‑based update mechanism leveraging user feedback for rapid adaptation to trending content. The paper received a best‑paper honor nomination.

Paper 2: Supervised Learning‑enhanced Multi‑Group Actor‑Critic for Live Stream Allocation in Feed

Link: https://arxiv.org/pdf/2412.10381

Abstract: In mixed recommendation scenarios, the system must decide whether to allocate a live‑stream slot for each user request. Existing reinforcement‑learning methods struggle with convergence and stability at industrial scale. Kuaishou introduces SL‑MGAC, which adds supervised learning to the actor‑critic framework, uses variance‑reduction techniques, and designs a multi‑group state‑decoupling module to lower prediction variance. A novel reward function avoids overly greedy live‑stream allocation. Offline policy evaluation and online A/B tests show SL‑MGAC outperforms baselines under platform constraints and offers higher stability in production.

Paper 3: Combinatorial Optimization Perspective based Framework for Multi‑behavior Recommendation (COPF)

Link: https://dl.acm.org/doi/pdf/10.1145/3690624.3709278

Abstract: Real‑world recommendation involves multiple user behaviors. Existing multi‑behavior methods fuse behaviors via graph neural networks and apply multi‑task learning, but they suffer from limited fusion perspectives and negative transfer between tasks. COPF treats multi‑behavior fusion as a combinatorial optimization problem, introducing constraints at each behavior stage (COGCN module) to improve fusion efficiency. In the prediction stage, a refined multi‑expert model (DFME module) mitigates negative transfer caused by feature and label distribution differences. Experiments on three real datasets demonstrate superior performance and validate the effectiveness of COGCN and DFME.

Paper 4: GREAT – Guiding Query Generation with a Trie for Recommending Related Search about Video at Kuaishou

Link: https://arxiv.org/abs/2507.15267

Abstract: Related‑search queries at the bottom of short‑video pages are a newly important scenario. Existing research and public datasets are scarce, and current methods rely on user behavior or embeddings, which either need long data accumulation or lack deep semantic interaction. GREAT first releases a large‑scale dataset KuaiRS for this scenario. It then uses a trie‑based dictionary to guide a large language model (LLM) in generating high‑quality queries. During training, a Next‑Token‑in‑Trie‑Prediction (NTTP) task is added; during inference, the trie steers token generation. A post‑processing filter further ensures relevance and safety. GREAT is fully deployed in Kuaishou’s photo‑to‑query service, serving billions of daily users.

Paper 5: HoME – Hierarchy of Multi‑Gate Experts for Multi‑Task Learning at Kuaishou

Link: https://arxiv.org/abs/2408.05430

Abstract: Mixture‑of‑Experts (MoE) models are widely used for multi‑task learning in recommendation systems, but they face three stability issues: expert collapse, expert degeneration, and expert under‑fitting. HoME addresses these by (1) expert normalization to align output distributions, (2) hierarchical masking to improve task sharing and reduce redundancy, and (3) feature gating to ensure each expert receives adequate gradients. Extensive offline and online experiments show a 0.954 % increase in average app usage time, and HoME is now fully deployed for Kuaishou’s short‑video services, serving 400 million users daily.

Paper 6: Improving Long‑tail User CTR Prediction via Hierarchical Distribution Alignment

Link: https://cloud.tsinghua.edu.cn/f/99648e5e5e784a53b977/

Abstract: CTR prediction for long‑tail users suffers from data scarcity and distribution imbalance. The proposed framework aligns hierarchical distributions between head and tail users, learns hierarchical residuals, and applies adaptive re‑weighting with a distribution calibration module. The method is model‑agnostic and improves long‑tail accuracy and fairness while maintaining overall performance across multiple public datasets and online experiments.

Paper 7: Mitigating Redundancy in Deep Recommender Systems – A Field Importance Distribution Perspective

Link: https://dl.acm.org/doi/abs/10.1145/3690624.3709275

Abstract: Redundant features increase computational cost and degrade performance in recommender systems. Existing feature‑selection or embedding‑size reduction methods lack a unified metric for feature‑domain contribution. This work proposes a distribution‑based feature‑domain optimization framework that learns importance scores for each domain, removes noisy features, and allocates adaptive embedding dimensions based on similarity of importance. Experiments show substantial reduction in training/inference time with improved prediction accuracy.

Paper 8: Personalized Query Auto‑Completion for Long and Short‑Term Interests with Adaptive Detoxification Generation (LaD)

Link: https://arxiv.org/abs/2505.20966

Abstract: Query auto‑completion must balance personalization with inference latency, and generated queries can contain toxic or low‑quality content. LaD uses a nested architecture: short‑term interests are encoded as token‑level inputs, while long‑term interests are pre‑computed embeddings, reducing sequence length. A Reject‑Preference Optimization (RPO) algorithm introduces a special [Reject] token trained via reinforcement learning to push low‑quality queries after it, enabling end‑to‑end detoxification. LaD has been fully deployed in Kuaishou’s search SUG service, delivering the largest AB‑test gains in two years and serving billions of users daily.

Overall, these eight papers demonstrate Kuaishou’s extensive AI research across recommendation, multimodal modeling, large language models, and system optimization, and they contribute novel methods that have already been deployed at massive scale.