FRecSys '26
ACM SIGOPS ATC 2026 Workshop

FRecSys '26

The 1st Workshop on Frontier Recommender Systems

View Program View Call for Papers

About the Workshop

The 1st Workshop on Frontier Recommender Systems (FRecSys), co-located with ACM SIGOPS ATC 2026, brings together the systems and recommendation communities to rethink the infrastructure behind modern recommender systems. These systems shape online experiences at global scale while meeting demanding requirements for latency, cost, reliability, and freshness.

Emerging paradigms—including generative, foundation-model-based, multimodal, agentic, and interactive recommendation—are reshaping what recommender systems compute and how they are built and served. FRecSys offers a forum to explore the infrastructure needed for these frontier workloads and to advance rigorous, reproducible research through shared insights, benchmarks, and evaluation methods.

Program

Sunday, November 15, 2026 · Regency Ballroom I, Hyatt Regency Hong Kong, Shatin · All times are Hong Kong Time (UTC+8)

  1. 08:00–08:30

    Registration & Badge Pickup

  2. 08:30–08:40

    Opening Remarks

    General Co-Chairs

  3. 08:40–09:05

    Invited Talk

    Minhui Xie (Nanjing University)

  4. 09:05–09:10

    Best Paper Award

    Announcement and award presentation

  5. 09:10–10:30
    Paper Session 1

    Generative Recommendation Inference & Serving

    Each paper: 8-minute talk + 2-minute Q&A

    1. 09:10
      OneLA: Scaling Linear-Attention Decoding to Large Beams in Generative Recommendation Xiangrui Yang (The University of Hong Kong); Cheng Peng, Yunfeng Zhao, Liang Zeng, Ao Hu (Kuaishou Technology); Jiawei Yang (Peking University); Shengzhe Wang, Jingshan Lv, Xiao Liang, Chen Yang, Jiaqiang Liu (Kuaishou Technology); Yiming Qiu (The University of Hong Kong)
    2. 09:20
      Token-Native Serving Architecture for Generative Recommendation with Semantic IDs Chuyu Qiu, Xuan Qi, Lujia Zhang, Chuanwei Yi, Faran Ahmad, Charlie Zhang, Hongbo Qin, Liangbei Xu, Jinho Hwang, Emma Lin, Yuxi Hu, Yongxiong Ren, Rui Jian, Jiyan Yang, Kai Ren (Meta)
    3. 09:30
      RIG: GPU Inference Framework for Long-Context, Short-Decode, Large-Beam Generative Recommendation Bin Chai, Jerry Chen (NVIDIA)
    4. 09:40
      Taming Wide Beams: Request-Adaptive Inference for Generative Recommendation Enrui Zhao (Renmin University of China); Minhui Xie (Nanjing University); Yunfeng Zhao (Kuaishou Technology); Puqing Wu, Yaxin Ma, Hongqi Shan (Renmin University of China); Yu Kong, Liang Zeng, Jiaqiang Liu, Xiao Liang (Kuaishou Technology); Yunpeng Chai (Renmin University of China)
    5. 09:50
      FSBO: Fusing Vocabulary Projection and Beam Selection for Efficient Generative Recommendation Chunfeng Du, Zeqing Wu, Wentao Li (Zhengzhou University); Yuan Liu (Dawning Information Industry Co., Ltd.); Zhihao Zhang (Tsinghua University)
    6. 10:00
      FlagCD: Exact Request-Constrained Decoding over Billion-Scale Item Catalogs Jiawei Yang (Peking University); Xiangrui Yang (The University of Hong Kong); Cheng Peng (Kuaishou Technology); Zihao Zheng (Peking University); Liang Zeng, Yunfeng Zhao, Lei Yan, Yuting Deng, Jiaqiang Liu, Xiao Liang (Kuaishou Technology)
    7. 10:10
      ServerlessANN: Low-Latency Serverless Vector Search with CXL Memory Pools Bing Tian, Yanlin Huang, Haikun Liu, Jianhao Mu, Zhuohui Duan (Huazhong University of Science and Technology); XiaoCheng Zhong, Yiping Sun, Xiao Chen (Xiaohongshu)
    8. 10:20
      NaviPerf: Architecture-Aware Performance Prediction for LLM Serving on Ascend NPUs Xiao Zhou, Jinghao Wang, Zeying Wang, Yifeng Zhang (Beihang University); Qicheng Liang (Xiaohongshu); Tianyu Wo, Chunming Hu, Renyu Yang (Beihang University)
  6. 10:30–11:00

    Coffee Break

  7. 11:00–11:50
    Paper Session 2

    Training Systems & Agentic Optimization

    Each paper: 8-minute talk + 2-minute Q&A

    1. 11:00
      RecIS: Unifying Sparse and Dense Computation for Efficient Recommendation Training Hua Zong (Alibaba Group); Shuai Yang (University of Chinese Academy of Sciences); Zhengxiong Zhou, Qingtao Zeng, Jiawei Liu, Hao Wan, Hongbin Zhou, Jiahao Wang, Hao Zhang, Bencheng Yan, Yuheng Huang, Huimin Yi, Jiamang Wang, Han Zhu, Jian Xu, Bo Zheng, Lin Qu (Alibaba Group)
    2. 11:10
      SparseRec: Predictive GPU Sharing and Data-Aware Scheduling for Recommendation Model Training Jinghao Wang, Xinlei Zheng, Xiao Zhou, Yihui Zhang, Zeying Wang (Beihang University); Qicheng Liang (Xiaohongshu); Tianyu Wo, Chunming Hu (Beihang University); Xiaoyang Sun (University of Leeds)
    3. 11:20
      KAISync: Characterizing and Optimizing Weight Synchronization in Co-Located RL Post-Training Yuxi Luo (Beihang University); Xiaoyang Sun (Kuaishou Technology); Zeying Wang, Cangzhou Yuan (Beihang University); Hengrui Ju (Kuaishou Technology); Renyu Yang (Beihang University)
    4. 11:30
      TALOS: Torch Acceleration through a Loop-Driven Optimization Skill for Recommendation Systems Ruotong Wang, Weida Yang, Zihua Wu (NVIDIA)
    5. 11:40
      RobustSGPO: Search-Space Control for Agent Harness Evolution Zibo Zhao, Jijun Shi, Mo Zhou, Zhongyuan Wang, Shifu Bie, Yunfei Zhang, Xuanting Zhou, Xiangyu Wu (Kuaishou Technology)
  8. 11:50–12:00

    Closing Remarks & Open Discussion

  9. 12:00–13:30

    Lunch & Networking

Call for Papers

The 1st Workshop on Frontier Recommender Systems (FRecSys) aims to provide a forum for the systems and recommendation communities to rethink the infrastructure foundations of modern recommender systems. Recommender systems are now core production infrastructure for online services, shaping user experience, content distribution, advertising, and e-commerce at global scale. They operate under strict latency, cost, reliability, and freshness constraints, while continuously adapting to changing users, content, models, policies, and business requirements.

For many years, recommender-system infrastructure has been shaped by retrieval-and-ranking pipelines, large embedding tables, feature stores, and DLRM-style serving architectures. These systems remain essential and continue to raise important systems challenges. At the same time, rapid advances in AI are creating new recommendation paradigms, including generative recommendation, foundation-model-based recommendation, multimodal recommendation, agentic recommendation, chain-of-thought recommendation, and interactive user modeling. These emerging workloads change both what recommender systems compute and how they must be served. They introduce new requirements for model execution, user-context management, online adaptation, system observability, resource scheduling, evaluation, and deployment at scale.

FRecSys seeks papers that examine how recommendation infrastructure should evolve for these frontier workloads. We welcome early ideas, position papers, preliminary research results, industry experience reports, benchmark proposals, and lessons from production deployments. The workshop aims to bring together researchers and practitioners from systems, machine learning, recommendation, databases, networking, hardware, and cloud infrastructure to identify open problems and community resources for the next generation of recommendation systems.

A further goal of the workshop is to encourage the development of shared artifacts for rigorous and reproducible research. Today, many important systems questions in recommender systems are difficult to study because public benchmarks, traces, and evaluation methodologies often fail to capture the scale, heterogeneity, dynamism, and operational constraints of production environments. FRecSys therefore particularly encourages submissions that expose realistic production constraints, characterize workloads that are difficult to capture in existing public benchmarks, formalize open systems problems for emerging recommendation platforms, or propose community resources such as representative traces, workload specifications, benchmark suites, and evaluation guidelines.

Main Topics

Topics of interest include, but are not limited to:

  • Agentic, interactive, and reasoning-enhanced recommendation systems.
  • Infrastructures for generative and agentic recommendation.
  • Retrieval, ranking, and generative recommendation systems.
  • Approximate nearest neighbor (ANN) search, vector indexing, and retrieval infrastructure.
  • Multimodal, graph-based, sequence-based, and foundation-model-based recommenders.
  • Training and serving systems for large-scale recommenders.
  • Batching, scheduling, GPU acceleration, and tail-latency control.
  • Embedding systems, feature stores, user-context management, and caching.
  • Online learning, continual training, model refresh, and feature freshness.
  • Reliable model deployment, rollout, rollback, and production experimentation.
  • Streaming logs, feedback loops, user simulation, and benchmark construction.
  • Reliability, observability, debugging, and incident response for recommendation services.
  • Co-design across algorithms, accelerators, and serving platforms.
  • Workload characterization, production traces, evaluation methodology, and reproducibility.

The half-day workshop will include an invited talk, research paper presentations, and a closing discussion on open systems problems and possible community artifacts. See the full program. This is the first edition of the workshop.

Publication & Open Access

The proceedings of FRecSys ’26 will be published in the ACM Digital Library as part of the International Conference Proceedings Series (ICPS).

Camera-Ready Format

Accepted papers must use the ACM two-column sigconf format for publication in ICPS. For LaTeX, use \documentclass[sigconf]{acmart} with the official ACM proceedings template (Overleaf template).

Include author names and affiliations, remove the anonymous and review options, and include the rights statement and bibliographic metadata supplied by ACM after completing the eRights form.

Important note to authors about ACM's new open access publishing model

ACM has introduced a new open access publishing model for the International Conference Proceedings Series (ICPS). Authors based at institutions that are not yet part of the ACM Open program and do not qualify for a full geographic waiver will be required to pay an article processing charge (APC) to publish their ICPS article in the ACM Digital Library. To determine whether or not an APC will be applicable to your article, please follow the detailed guidance here.

Further information may be found on the ACM website, as follows:

Please direct all questions about the new model to icps-info@acm.org.

Review Process

Double-Anonymous Review

FRecSys uses a double-anonymous review process: authors’ identities are hidden from reviewers, and reviewers’ identities are hidden from authors. Submissions must not reveal author names, affiliations, or other identifying information.

Three Reviews per Paper

Each submission is expected to receive three independent reviews. Reviewers are drawn from the Program Committee listed below, where their names and affiliations are provided.

One-Month Review Period

Reviewers will have approximately one month to evaluate submissions. The author-notification date has been updated to October 12, 2026.

Final Decisions

Submissions will receive an accept or reject decision directly; the workshop will not issue minor- or major-revision decisions. Authors of accepted papers may address reviewer feedback when preparing the camera-ready version.

Committees & Organizers

General Co-Chairs

Steering Committee

  • Yuan XieHong Kong University of Science and Technology
  • Ceyu XuHong Kong University of Science and Technology

Program Committee

Publication & Web Chair