Source-linked AI summary
OneLive: Dynamically Unified Generative Framework for Live-Streaming Recommendation
Shen Wang, Yusheng Huang, Ruochen Yang, Shuang Wen, Pengbo Xu, Jiangxia Cao, Yueyang Liu, Kuo Cai, Chengcheng Guo, Shiyao Wang, Xinchen Luo, Qiang Luo, Ruiming Tang, Shuang Yang, Zhaojie Liu, Guorui Zhou, Han Li, Kun Gai
TL;DR
Live-streaming recommendation involves evolving content, limited lifecycles, strict real-time constraints, and heterogeneous objectives that prevent direct transfer of existing generative recommendation methods. OneLive addresses these challenges with dynamic tokenization, temporal attention, efficient decoder-only generation, and multi-objective alignment. The framework was deployed on Kuaishou and serves 400 million daily live-streaming users with significant business benefits.
Problem
Evolving content, limited lifecycles, strict real-time constraints, and heterogeneous objectives limit direct application of existing generative recommendation methods to live-streaming.
Method
OneLive combines dynamic semantic-collaborative tokenization, time-aware gated attention, decoder-only generation with Sequential MTP and QK Normalization, and multi-objective reinforcement learning.
Results
OneLive is deployed on Kuaishou App, serving 400 million users daily in the live-streaming scenario and bringing significant business benefits.
Takeaways & Limitations
OneLive provides a unified generative recommendation framework tailored to live-streaming and has demonstrated effectiveness in offline and online experiments.
Abstract
from arXiv · showhide
Live-streaming recommender system serves as critical infrastructure that bridges the patterns of real-time interactions between users and authors. Similar to traditional industrial recommender systems, live-streaming recommendation also relies on cascade architectures to support large-scale concurrency. Recent advances in generative recommendation unify the multi-stage recommendation process with Transformer-based architectures, offering improved scalability and higher computational efficiency. However, the inherent complexity of live-streaming prevents the direct transfer of these methods to live-streaming scenario, where continuously evolving content, limited lifecycles, strict real-time constraints, and heterogeneous multi-objectives introduce unique challenges that invalidate static tokenization and conventional model framework. To address these issues, we propose OneLive, a dynamically unified generative recommendation framework tailored for live-streaming scenario. OneLive integrates four key components: (i) A Dynamic Tokenizer that continuously encodes evolving real-time live content fused with behavior signal through residual quantization; (ii) A Time-Aware Gated Attention mechanism that explicitly models temporal dynamics for timely decision making; (iii) An efficient decoder-only generative architecture enhanced with Sequential MTP and QK Norm for stable training and accelerated inference; (iv) A Unified Multi-Objective Alignment Framework reinforces policy optimization for personalized preferences.
1 Introduction
Live-streaming recommendation must handle evolving content, short exposure windows, strict latency, and heterogeneous objectives that limit direct transfer of conventional generative recommendation. OneLive addresses these conditions with a dynamically unified framework combining live-content tokenization, temporal modeling, efficient generation, and multi-objective alignment.
- Existing Systems: Conventional cascade architectures divide retrieval, pre-ranking, and ranking under fundamentally misaligned objectives and constraints.Retrieval emphasizes coverage and diversity, whereas later stages prioritize more precise estimation under different computational conditions.
- Generative Recommendation: Transformer-based generative recommendation unifies heterogeneous signals and end-to-end recommendation through tokenized entities and autoregressive generation.These methods improve scalability and computational efficiency but require adaptation for live-streaming complexity.
- Challenges: Live-streaming content, author activity, feedback, and interaction dynamics evolve continuously throughout each broadcast.This makes static representations unsuitable for the live-streaming setting.
- Challenges: Live-streaming recommendations must remain valid across limited exposure lifecycles while responding under stringent real-time constraints.Only ongoing broadcasts are eligible for distribution, and candidate pools and user queries change rapidly.
- OneLive: OneLive combines a Dynamic Tokenizer, Time-Aware Gated Attention, Sequential MTP with QK Normalization, and Unified Multi-Objective Alignment.The framework is designed for evolving content, temporal constraints, efficient serving, and heterogeneous user preferences.
2 Methodology
OneLive adapts unified generative recommendation to live-streaming by jointly modeling dynamic content, collaborative signals, temporal constraints, efficient generation, and training stability. Its methodology combines dynamic semantic-collaborative alignment with time-aware decoding, Sequential MTP, QK Norm, and reinforcement-learning-based optimization.
- Dynamic Tokenizer: OneLive jointly models evolving live-streaming content and immediate user feedback to produce semantically and collaboratively aligned item embeddings.The dynamic tokenizer uses multimodal content understanding followed by real-time collaborative post-alignment.
- Dynamic Tokenizer: Content-only embeddings lack collaborative signals, so OneLive refines dynamic content representations using interaction supervision before quantization.This addresses weak discriminability and changing popularity or cohort-specific preferences.
- Dynamic Tokenizer: IA Embeddings achieve higher codebook utilization and lower collision rates than MLLM Embeddings, while Res-Kmeans outperforms RQ-VAE under matched settings.The selected configuration uses Res-Kmeans with T=3 layers and codebook size N_l=8192.
- Temporal Dynamics: OneLive models live-streaming time constraints through historical-sequence, generation-anchor, and attention-gated temporal perception.Gated attention adaptively modulates weights over user history and preceding decoded sequences in both cross-attention and self-attention.
- Efficient Generation: Sequential MTP addresses inefficient autoregressive inference by retaining a full main decoder and using lightweight subsequent decoder modules for residual SID components.The design exploits the relative simplicity of predicting later residual components and reuses the main decoder’s KV cache.
- Training Stability: QK Norm applies RMSNorm to query and key vectors, constraining attention-logit magnitude and reducing softmax saturation and loss spikes during deep training.The normalization suppresses scale drift across depth and improves training stability.
- Optimization: OneLive combines next-token-style training with on-policy reinforcement learning to optimize beyond historical behavior cloning.The reinforcement-learning objective is introduced because offline next-author prediction mimics deployed historical policies.
3 Experiments
OneLive is evaluated through offline experiments, component ablations, scaling analysis, and online deployment. It consistently improves recommendation quality and demonstrates gains in inference efficiency and live-streaming business indicators.
- Overall Performance: OneLive consistently outperforms existing baselines, improving HR@128 by up to 16.70% and MRR@128 by up to 14.26%.It also surpasses the best baseline by at least 13.29% across other indicators.
- Component Ablation: Sequential MTP increases single-machine QPS by 31.8% and reduces computational graph execution latency by 55.45%.MFU reaches 22.78% on the L20 GPU, compared with 2.56% for the online ranking model.
- Component Ablation: Time-aware temporal modeling improves the inference validity rate by 14.29% through adaptive attention gating.The resulting author distribution is described as reasonable for live-streaming recommendation.
- Component Ablation: GRPO provides consistent performance and system-recognition improvements across beam sizes, whereas DPO improves candidate filtering mainly at large beam sizes.The comparison is attributed to GRPO using group responses for richer advantage and gradient feedback.
- Scaling Analysis: As model size increases, loss decreases with a stable upward performance trend, but further scaling yields diminishing marginal returns.The scaling experiment visualizes loss curves across parameter scales.
- Online A/B Test: Online deployment across Kuaishou and Kuaishou Lite produces substantial improvements in exposure, CTR, click, watch, and follow indicators.The model serves hundreds of millions of users and millions of authors across the two platforms.
4 Conclusion
The paper concludes that OneLive unifies dynamic content understanding, temporal modeling, efficient generation, and preference alignment for live-streaming recommendation. Offline and online experiments demonstrate its effectiveness, and the system has been deployed on Kuaishou App serving 400 million daily users.
- Conclusion: OneLive combines a dynamic tokenizer, time-aware gated attention, Sequential MTP, QK normalization, and multi-objective reinforcement learning.These components address content dynamics, temporal modeling, inference efficiency, and personalized preference alignment.
- Conclusion: OneLive has been deployed on Kuaishou App and serves 400 million users daily in the live-streaming scenario.The paper reports significant business benefits from this deployment.
- Motivation: Live-streaming recommendation requires real-time content understanding because live content changes dynamically and aggregated information may not match its current state.Earlier methods either focus on historical interactions or use coarse-grained multimodal information.
- Motivation: Existing discriminative methods rely on pointwise scoring, ranking, and candidate prefiltering, motivating an end-to-end generative paradigm for immediate feedback.The paper frames this shift as important for live-streaming recommendation.
- Related Work: Generative recommendation methods extend unified architectures through discrete tokenization, direct generation, preference alignment, and scenario-specific adaptations.Prior work includes encoder-decoder generation and applications to local life and query suggestion.
B.1 Experimental Settings
The experiments use large-scale Kuaishou live-streaming logs covering hundreds of millions of users, millions of authors, and billions of interactions over nearly two months.
- Dataset: The dataset contains nearly two months of Kuaishou logs from 400 million users and 3 million authors.It includes billions of user-author interactions and covers long-view and click records.
- Dataset: Data from the last day are used for testing, while data from the remaining dates are used for training.All offline and ablation experiments use this dataset.
B.1.1 Dataset.
The evaluation measures offline recall and ranking performance alongside reward-based recognition and online business effectiveness.
- Metrics: Offline evaluation uses Accuracy, Hit Rate, Mean Reciprocal Ranking, and Reward from the deployed online ranking model.Accuracy, Hit Rate, and Mean Reciprocal Ranking assess recall and ranking performance, while Reward evaluates generative-result recognition.
- Metrics: Online A/B tests use core market indicators to evaluate business effectiveness.Metric definitions are provided in Appendix C.2.
B.1.2 Evaluation Metrics.
OneLive is evaluated against competitive traditional and generative recommendation baselines. The comparison spans sequential, long-sequence, graph-based, SID-based, and industrial generative methods.
- The evaluation compares OneLive with both traditional and generative recommendation models.The traditional group includes SASRec, KuaiFormer, and a GNN variant; the generative group includes TIGER and OneRec.
- Traditional Models: Traditional baselines include SASRec, KuaiFormer, and a heterogeneous structure-aware GNN variant.
- Generative Models: Generative baselines include the SID-based method TIGER and the industrial solution OneRec.
B.2 Overall Experiemnt Analysis
The experiments show that static generative recommendation does not necessarily transfer well to dynamic live-streaming, whereas OneLive adapts to evolving content and performs consistently across user groups. Its dynamic tokenizer updates live-streaming SIDs in real time, while temporal modeling and enriched objectives support industrial adaptability.
- Overall Comparison: TIGER experiences a significant performance drop in the dynamic live-streaming setting.The passage attributes this to modeling historical interaction items rather than instant evolving content, with SID Codes performing even worse than MLLM Codes.
- Overall Comparison: The observed generative-model failure reinforces the challenge of real-time recommendation with evolving live content.
- Overall Comparison: OneLive and OneRec show strong adaptability in industrial settings.User and author side information improves generalization, while temporal modeling and enriched training objectives align with live-streaming requirements.
- Stratified Test: OneLive achieves consistent and significant improvements across low-active, high-active, and core-paid user groups.Low-active users exhibit the largest gains, although a few indicators show minor decreases.
- SID Updates: The dynamic tokenizer generates live-streaming SIDs that capture authors' current content and update in real time.This supports continuity as live-streaming content evolves.
C.1 Codebook Analysis
The codebook analysis evaluates SID quality through utilization and collision measures, while model performance is assessed with training accuracy and inference metrics. These metrics cover codebook coverage, ambiguity, prediction correctness, ranking, and online-system recognition.
- Codebook Quality: Utilization Rate measures the proportion of distinct codes used at each codebook layer.For layer L_i, it relates the number of used codes to the total codebook size.
- Codebook Quality: Collision Rate measures overlap when SIDs are associated with multiple authors or authors share the same SID.The analysis distinguishes collision rates in the SID and Author dimensions.
- Model Performance: Accuracy measures average multi-level SID prediction accuracy during training.
- Model Performance: Hit Rate measures whether the correct answer appears among the Top-k beam-search predictions, while MRR measures its rank.
- Model Performance: Reward measures online-system recognition by transferring predicted user-author pairs into a ranking model to obtain XTR scores.