AI@LHC 2026: what's next?

25 May 2026, 08:59 → 29 May 2026, 01:20 Asia/Shanghai

w202 (Peking University)

Chen Zhou (Peking University (CN)), Congqiao Li (Peking University (CN)), Huilin Qu (CERN), Maurizio Pierini (CERN), Qiang Li (Peking University (CN))

A precursor to BOOST 2027

We sincerely invite you to participate in the workshop “AI@LHC 2026: What's Next”, which will be held May 25–28 at Peking University, organized with the support of the CHEP–CERN Joint Program.

• Registration: Sunday, May 24

• Workshop Program: May 25–27

• Free Discussions & Collaboration Day: May 28

 

This workshop is jointly initiated and organized by scholars from CERN and Peking University who are engaged in interdisciplinary research at the intersection of LHC experiments and artificial intelligence. The meeting will feature a series of talks covering multiple topics related to AI for the LHC, including invited presentations by researchers from CERN. In addition, experts in the field of AI from industry and research institutes will be specially invited to deliver keynote talks, jointly exploring new challenges, new methods, and future directions of artificial intelligence in collider physics.

If you have recent research results related to AI applications for the LHC, we warmly welcome you to submit a presentation. We would also greatly appreciate it if you could forward this conference information to other interested colleagues, especially graduate students and postdoctoral researchers.

There is no registration fee for this workshop; however, participants are responsible for their own travel and accommodation. If you have any questions, please feel free to contact us at any time.

The program features 15 invited talks, several tutorials, and general keynotes. About 18 lightning talks (7+3 minutes), accompanied by posters, are open for submission.  

The scientific program will include invited and contributed talks on topics such as:

• AI at the LHC

• AI for event reconstruction

• AI for event generation

• Simulation-based inference for collider physics

• ML-based particle flow

• Jet tagging: from DNNs to Transformers

• Anomalous jet detection

• Jet tagging on the edge

• Anomaly detection on the edge

• Foundation models for HEP

• The future of AI on edge for real-time processing

In addition, the workshop will feature keynote talks on frontier AI and computing topics.

The workshop will bring together experts in particle physics, artificial intelligence, and computing infrastructure to discuss emerging challenges, novel methodologies, and future directions of AI in collider physics.

We warmly welcome researchers, students, and industry partners to join us for this exciting event.

---

诚挚邀请您参加研讨会 “AI@LHC 2026: What's Next”，会议将于5月25日至28日在北京大学举行。本次会议在 CHEP–CERN 联合项目的支持下组织召开。

• 报到时间： 5月24日（周日）

• 会议时间： 5月25日–27日

• 自由讨论与合作交流： 5月28日

 

本次会议由来自 欧洲核子研究中心（CERN）和北京大学、从事 LHC 实验与 AI 交叉研究的学者共同发起与组织。会议将围绕 AI for LHC 的多个专题开展系列报告，包括来自 CERN 学者的邀请报告，并特邀来自企业界、科研院所的 AI 领域专家作专题演讲，共同探讨人工智能在对撞机物理中的新挑战、新方法以及未来发展方向。如果您近期在 LHC 相关的 AI 应用方面有新的研究成果，我们诚挚欢迎您提交报告。同时也非常感谢您将会议信息转发给其他感兴趣的同行，特别是研究生和博士后。本次会议不收注册费，食宿自理。如有任何问题，欢迎随时与我们联系。

会议安排包括 15 场邀请报告、若干教程和大会主旨报告。另有约 18 个闪电报告（7+3分钟）及配套海报展示，现正开放投稿。

会议将涵盖多个专题报告，包括但不限于：

• LHC 中的人工智能应用

• 基于AI的事例重建

• 基于AI的事例生成

• 基于模拟的对撞机物理推断

• 基于机器学习的粒子流重建

• 从 DNN 到 Transformer 的喷注标记

• 异常喷注探测

• 边缘端喷注标记

• 边缘端异常探测

• 面向高能物理的基础模型

• 面向实时处理的边缘 AI 未来发展

此外，会议还将邀请多位 AI 领域专家作大会特邀报告。

本次研讨会将汇聚粒子物理、人工智能及计算基础设施领域的专家学者，共同探讨 AI 在对撞机物理中的新挑战、新方法与未来发展方向。

诚挚欢迎国内外科研人员、青年学者及学生参加本次会议。

AILHC2026new.pdf

ailhc26.jpeg

Poster_AIatLHC2026.pdf

Zoom AI@LHC

Monday 25 May

1 Morning Session 1 [Chaired by Qiang Li]

AI@LHC2026Welcome.pdf

group2.jpg

group.jpg

2 Welcome

Speaker: Qinghong Cao (Peking University)

qhc.jpg

Good morning，

 distinguished scholars, colleagues, and friends from across the globe. It is my greatest honor to stand before you today at Peking University to officially extend our warmest welcome to the "AI@LHC 2026: What's Next" workshop.
On behalf of the organizing committee, I want to express our sincere appreciation to every participant who has traveled far and wide to join us—especially our colleagues from CERN, leading research institutes, and industry partners. This gathering is a collaborative initiative supported by the CHEP–CERN Joint Program, bringing together two world-class communities in particle physics and artificial intelligence to explore the frontier of collider science.

We gather at a truly pivotal moment. The Large Hadron Collider has entered a new era of discovery, and artificial intelligence is no longer just an auxiliary tool—it is a core catalyst reshaping how we decode the universe’s most fundamental laws.

The scientific program of this workshop is designed to map the full spectrum of AI innovation for the LHC, covering multiple frontier topics:
• AI for core collider physics tasks: Including AI for event reconstruction, event generation, simulation-based inference for collider physics, and ML-based particle flow.
• Jet physics innovation: From DNN-based to Transformer-based jet tagging, anomalous jet detection, to jet tagging on the edge, pushing the boundaries of real-time processing.
• Cutting-edge AI paradigms: We will explore foundation models for high-energy physics (HEP), anomaly detection on the edge, and the future of AI on edge for real-time processing.
• Cross-domain frontier insights: The workshop also features keynote talks on frontier AI and computing topics delivered by industry and research institute experts, jointly exploring new challenges and future directions of AI in collider physics.

As we dive into these topics, I am proud to highlight the pioneering work happening right here at Peking University. Our high-energy physics team has been at the forefront of this integration:
• Development of Generalizable Jet Foundation Models: Since 2022, our team has led the development of the Global Particle Transformer (GloParT) at the CMS experiment, which is now deployed in experiments and has attracted participation from over ten international institutions. This work, along with the Particle Transformer algorithm and the JetClass/JetClass-II datasets co-developed with CERN and UCSD, has laid a critical foundation for large-scale model development in high-energy physics.
• AI for Full-Event Analysis: Our researchers have demonstrated that through precise AI engineering, the signal detection sensitivity for complex hadronic final states (such as double Higgs decay to 4b quarks) can surpass traditional methods by more than 5 times, showcasing the transformative potential of AI for LHC measurements.
• Next-Generation Jet Identification: In collaboration with IHEP and CERN, our team has developed AI-driven techniques combining the original Arbor particle flow algorithm with ParticleNet, capable of efficiently distinguishing 11 different types of jets (from 5 quarks, 5 antiquarks, and gluons). This "game-changing" technology, published in Physical Review Letters, can improve the precision of key physical measurements at future colliders by an order of magnitude.

Beyond experimental applications, Peking University is also pushing the boundaries of AI theory itself. Our interdisciplinary team recently made a landmark discovery: by applying physics’ least action principle, they identified detailed balance phenomena in LLM agent dynamics—the first time macroscopic physical laws have been found in AI generation processes without relying on specific model details. This work elevates AI research from empirical engineering to quantifiable physical science.

We also celebrate the BBT-Neutron scientific computing foundation model co-developed by PKU and partner institutions. Using innovative binary tokenization to unify multi-modal data processing, this open-source model matches the performance of specialized models like ParticleNet in jet origin identification tasks, even showing emergent phenomena not observed in traditional architectures.


Colleagues, the intersection of LHC experiments and artificial intelligence is one of the most dynamic and promising frontiers in modern science. With the solid foundation built by teams like ours at Peking University, and the collective expertise gathered in this room today, this workshop is more than a meeting—it is a launching pad for ideas that could redefine how we explore the building blocks of our universe.


So let’s embrace the next few days with curiosity, rigor, and a spirit of collaboration. I now declare the AI@LHC 2026: What's Next workshop officially open.

Thank you, and let the inspiring discussions begin!

3 Group Photo [outside open area]

4 AI at the LHC

Speaker: Maurizio Pierini (CERN)

AI@LHC-WhatNext.pdf

m1.jpg

m2.jpg

m.jpg

5 AI for event simulation

Speaker: Francesco Vaselli (Scuola Normale Superiore & INFN Pisa (IT))

AI for Event Simulation AI@LHC-1.pdf

Fr.jpg

Link to slides

10:50 Break

6 Morning Session 2 [Chaired by Maurizio Pieini]

7 AI for event reconstruction

Speaker: Farouk Mokhtar (Univ. of California San Diego (US))

F1.jpg

FM_AI@LHC_reco.pdf

8 Towards a Physical Theory of Agents

Speaker: Huaxing Zhu (ITP, PKU)

AIatLHC_ZHU.pdf

hxz.jpg

12:30 Lunch

9 Afternoon Session [Chaired by Congqiao Li]

10 Lecture + Tutorial 1: Machine Learning From Simple to Scale

Speaker: Huilin Qu (Tsung-Dao Lee Institute, Shanghai Jiao Tong University (CN))

hlq.jpg

ML4HEP_20260525.pdf

Tutorials

15:30 Break

11 Octree-based 3D Neural Representation and Learning

Keynote talk
https://wang-ps.github.io/

Speaker: Prof. Peng-Shuai Wang (Wangxuan Institute of Computer Technology, Peking University)

psw1.jpg

psw2.jpg

Tuesday 26 May

12 Morning Session 1 [Chaired by Huilin Qu]

13 Simulation-based inference for collider physics

Speaker: Jingjing Pan (KIT - Karlsruhe Institute of Technology (DE))

20260530_AI_LHC_PKU.pdf

14 Foundation Models for HEP

Speaker: Yulei Zhang (University of Washington (US))

AI_LHC_2026

ylz1.jpg

ylz2.jpg

10:10 Break

15 Morning Session 2 [Chaired by Congqiao Li]

16 Simulation-based inference for precision neutrino physics [Remote]

Speaker: Dr Arsenii Gavrikov

2044557885.jpg

Gavrikov_AI@LHC_SBI_for_MC_tuning.pdf

17 SAM 3D: 3Dfy Anything in Images [Remote]

Keynote talk
https://sites.google.com/view/weiyaowang/home

Speaker: Weiyao Wang (Thinking Machines Lab)

1118735105.jpg

video1755142039.mp4

12:20 Lunch

18 Afternoon Session [Chaired by Bingxuan Liu]

19 Tutorial 2: Introduction to Fast Machine Learning

Designing network that are hardware efficient and hardware-aware, with a particular focus on wiring ML algorithms directly into FPGA fabric, for a deterministic hardware computation.

Speakers: Christopher Edward Brown (CERN), Maciej Mikolaj Glowacki (CERN)

FastML.pdf

Introduction_to_FPGAs_PKU.pdf

ML_LHC_tutorial.ipynb

model.py

t21.jpg

t22.jpg

15:30 Break

20 New AI Infrastructure for Dr.Sai Agents at IHEP

Speaker: Hong Wang

AI Infra-Hong Wang-v3 1.pdf

w1.jpg

w2.jpg

w3.jpg

w4.jpg

21 Dr.Sai Agentic Scientific Workflow in JUNO

Modern neutrino experiments depend on complex and highly iterative analysis workflows involving reconstruction, simulation, calibration, background studies, validation, and documentation. In many cases, the bottleneck is not a single algorithm, but the efficient, reproducible, and auditable execution of expert-defined procedures. This talk presents the application of the Dr.Sai agentic scientific workflow in JUNO, focusing on system architecture and practical analysis outcomes rather than machine-learning methodology.

Dr.Sai is a project launched at IHEP to transform expert scientific procedures into structured agentic workflows. In JUNO, we adopt specification-driven development as the technical path. Scientific tasks are broken into small, testable units; atomic operations become reusable skills; and agents and researchers jointly assemble them into work plans. Requirements, validation criteria, unit tests, end-to-end tests, and acceptance tests are encoded in specification files. Together, skills, plans, and specifications form the domain-specific layer of the agentic system, and are iteratively updated during the analysis as new lessons are learned.

Two JUNO applications will be discussed. The first is the (^{8})B solar-neutrino analysis, where skill- and specification-driven workflows help organize repeated analysis cycles, configuration management, validation steps, and result traceability. The second is the acceleration of OMILREC, JUNO’s data-driven vertex and energy reconstruction framework, where agent-assisted profiling, benchmarking, and validation loops shorten the optimization cycle while preserving physics-level checks.

This work demonstrates that agentic workflows can be integrated into real, production-level neutrino analyses. The JUNO experience shows that such systems can improve reproducibility, efficiency, and scalability while keeping scientific judgment under human control.

Speaker: Xuefeng Ding (Istitute of High Energy Physics, Chinese Academy of Sciences)

2026.05.26 [Professor][PKU] Dr.Sai Agentic Scientific Workflow in JUNO.pdf

d1.jpg

d2.jpg

d3.jpg

Wednesday 27 May

22 Morning Session [Chaired by Farouk Mokhtar]

23 Jet tagging from DNN to Transformers

Speaker: Bingxuan Liu (Sun Yat-Sen University (CN))

bxl.jpg

JetTaggingEvolvement_AILHC_BingxuanLiu.pdf

24 Jet tagging on the Edge

Speaker: Christopher Edward Brown (CERN)

chris.jpg

JetTaggingOntheEdge.pdf

25 Anomaly Detection on the edge

Speaker: Maciej Mikolaj Glowacki (CERN)

What's next for AD.pdf

10:45 Break

26 Morning Session 2 [Chaired by Jingjing Pan]

27 Accelerating Multi-Latent Attention on Spatial Architectures

Multi-latent attention (MLA) shrinks the KV cache but must rebuild per-head keys and values during decode, shifting the bottleneck to on-chip movement and orchestration. Spatial architectures, which consist of many-core tiles with local memory, explicit movers, and on-chip networks, therefore reward careful dataflow and mapping as much as computing. Unlike FPGA-centric stacks with long place-and-route time, the spatial architecture platforms support faster map–measure–refine loops.

We summarize two lines of work. On Tenstorrent Wormhole architecture, we treat MLA decode as a coupled mapping and scheduling problem under decoupled read–compute–write execution. Methodologically, we use a parameterized dataflow template, a cost model calibrated with simple empirical corrections, and an auto-tuner to explore a large, regime-dependent design space systematically. Measured decode behavior changes character with context length—from compute-dominated at short sequences to a movement- and coupling-limited regime at long context—motivating architecture-aware redesign rather than incremental tuning of a single fixed mapping.

On AMD Ryzen AI (Strix Halo), we build end-to-end MLA decode through a constraint-driven co-design loop: memory layout, DMA tiling, and matrix-multiply kernels are co-optimized so the full decode step runs as a single coordinated invocation across multiple compute columns, avoiding fragile multi-stage glue that often dominates latency on tightly coupled spatial substrates.

As physics-driven AI moves toward larger contexts, tighter latency budgets, and on-detector or facility-adjacent inference, spatial architectures offer a concrete path to turn memory-efficient attention into sustained throughput. https://ceca.pku.edu.cn/en/people_/faculty_/guojie_luo/

Speaker: Guojie Luo

Accelerate-MLA-on-Spatial-Arch.pdf

Accelerate-MLA-on-Spatial-Arch.pptx

gjl.jpg

28 Discussion

12:30 Lunch

29 Afternoon Session [Chaired by Chen Zhou]

30 Agent for multiple loop calculation

The LHC now delivers percent-level precision across Higgs, top, and electroweak observables, yet the corresponding multi-loop theory predictions—essential for converting experimental measurements into constraints on the Standard Model and new physics—remain locked behind formidable expertise barriers. A full NNLO or N$^3$LO calculation requires navigating a fragmented ecosystem of specialized packages, each demanding deep domain knowledge for setup, execution, and error diagnosis. The result is a widening theory-experiment precision gap that slows the physics exploitation of LHC data. The recent emergence of AI agent systems offers a path to bridge this gap: agents can encapsulate expert computational knowledge and automate complex multi-step workflows. We present a multi-agent framework that orchestrates the full multi-loop Feynman integral pipeline through modular skills with standardized interfaces and self-healing capabilities—lowering the threshold for non-specialists to obtain state-of-the-art predictions. We further introduce research and execute agent modalities that move beyond automation toward genuine scientific reasoning, aiming to democratize precision perturbative calculations for the LHC community.

Speaker: Xiang Li (Peking University)

xl.jpg

31 AI for future Higgs factory

We present an AI-driven framework to enhance the physics reach of future electron–positron colliders through fine-grained jet understanding and holistic event analysis. We propose jet origin identification (JoI), which classifies jets into five quark species, their anti-quarks, and gluons. Using simulated $\nu\bar{\nu}H,\ H\rightarrow jj$ events at 240~GeV at the CEPC, the method achieves jet flavor tagging efficiencies of 67%–92% and jet charge flip rates of 7%–24%. Combined with the holistic approach and Advanced Color Singlet Identification (ACSI), the framework significantly improves signal–background separation and enhances the precision of key Higgs measurements by up to a factor of six. For rare and exotic Higgs decays, projected 95% confidence level upper limits on branching ratios reach $2\times10^{-4}$ to $1\times10^{-3}$, with the sensitivity to $H\rightarrow s\bar{s}$ approaching within a factor of three of the Standard Model prediction. These results demonstrate the strong potential of AI-based particle-level analysis for future collider experiments.

Speaker: Yongfeng Zhu

32 Metric Spaces for Collider Events—Towards Scientific AI for Particle Physics

As the Large Hadron Collider (LHC) generates hundreds of petabytes of data and even more with its high-luminosity upgrade, particle physics is entering a new era of data-driven discovery where Artificial Intelligence (AI) plays a pivotal role. Alongside numerous task-specific AI algorithms, recent works have introduced foundation models excelling across diverse applications. At the heart of these models is a geometric representation space for collider events that encodes the essential physics. Key questions then arise: How can we probe and refine the representation space for theoretical insights?

This talk presents a first step towards the construction and analysis of a collider space. I will introduce one special metric structure inspired by the mathematical theory of optimal transport. Explicitly-defined, physically-grounded metrics can then be compared with representation spaces implicitly generated by AI models, providing deeper theoretical insights into their behavior. In this sense, particle physics offers an ideal testbed for advancing the next generation of Scientific AI.

Speakers: Ms Tianji Cai (University of California, Santa Barbara), Tianji Cai (SLAC National Accelerator Laboratory)

tjc1.jpg

tjc2.jpg

15:35 Break

Lightening talks: Session 1

33 Potential of di-Higgs observation via a calibratable jet-free HH->4b framework

We present a calibratable, jet-free framework that enhances the search significance of the flagship LHC channel HH→4b by more than a factor of five compared to existing approaches. The method employs a mass-decorrelated discriminant to identify h1h2→4b with variable h1,2 masses and a simultaneous estimator of (mh1,mh2), both derived from multiclass classification on all-particle inputs. The HH signal response can be calibrated using ZZ→4b. Using a highly realistic simulation framework validated through multiple tests, we demonstrate the method's robustness and identify two prerequisites essential for achieving this level of sensitivity. Results indicate that with LHC Run 2 and 3 data, observation-level sensitivity to HH appears within reach, enabling constraints on κλ comparable to HL-LHC projections and offering an accelerated path to precision measurements of the Higgs trilinear coupling.

Speaker: Tianyi Yang (Peking University (CN))

AI@LHC_TianyiYang_20260527.pdf

tyy.jpg

34 Learning Detector-Level Probability Distributions with Sliced Wasserstein Distance

High-energy physics is intrinsically a field that relies on the accurate modeling and comparison of high-dimensional probability distributions, arising from complex detector responses and multi-body final states. The Sliced Wasserstein Distance (SWD) provides a powerful and computationally efficient loss function tailored for high-dimensional probability distributions. By projecting high dimensional distributions onto infinite one-dimensional subspaces, SWD enables stable and scalable optimization while preserving essential geometric information. In this work, we present the application of SWD in several key HEP tasks that require fast and precise reproduction of target distributions. These include event generation, unfolding procedures, and fast simulation of detector effects. We demonstrate that SWD-based methods offer a flexible and data-driven framework for achieving high-fidelity modeling, making them a promising tool for future analyses at LHC experiments.

Speaker: Chucheng Pan (Wuhan University (CN))

ccp1.jpg

swd_ai_lhc.pdf

35 “Hadron-in-jet” AI tagging to detect rare decays

We investigate a novel class of boosted-object signatures at the LHC, where a high-pT fat jet contains an identifiable hadron or quarkonium state originating from rare or semi-exclusive decays.

Speaker: Linrui Chen (Wuhan University)

lrc1.jpg

lrc2.jpg

WtoPia_AIatLHC2026.pdf

WtoPia_AIatLHC2026.pptx

36 Talks&Closing

AI@LHC2026Closing.pdf

closing.jpg