AI for Meta Science

→ Europe/Paris
Institut Pascal

Institut Pascal

Rue André Rivière 91400 Orsay
Description

Overview

Artificial Intelligence (AI) has gained significant attention for its applications in science. This workshop will focus on its potential use beyond primary scientific data analysis pipelines, hence the "meta" in AI for meta-science. Important disclaimer : AI4MetaScience workshop at NeurIPS in Paris is an unrelated initiative, they are focusing more on the evaluation of science.

This workshop is mainly funded by AISSAI, CNRS's interdisciplinary centre for AI, and is hosted by Institut Pascal, Université Paris-Saclay.

The workshop is structured to favour discussions. Each theme will be covered in half a day or less, featuring 45-minute keynote talks and a series of short contributed talks, with ample time for discussion, which is well-suited to Institut Pascal. Each participant will have a desk in shared offices with access to meeting rooms (and, of course, free coffee machines) and will be provided with lunches (except on Monday), a workshop dinner on Thursday.

Two hackathons will be run in parallel for two half-days (Tuesday morning and Thursday morning), Mad Scientist, and From Literature to Benchmark (see below). Participants will apply to either by Monday night.

Attendance is limited, by invitation or upon application (menu on the left), but may close anytime due to limited capacity; applications will be evaluated with the goal of ensuring a diverse group of participants. Young researchers including PhD students are especially encouraged to apply.

Confirmed speakers

Confirmed speakers include so far:

  • Mateusz Bawaj (U Perugia) AI control of Gravitation Wave interferometer
  • David Chamont (CNRS) Labobots school
  • Adnan Ghribi (GANIL) Summary of accelerator data hackathon

  • Damien Gratadour (CNRS) AI control of telescopes

  • Verena Kain (CERN) AI control of accelerators

  • Francois Lanusse (CEA) LightCone : agentic AI assisted science
  • Julien Bettane (CNRS), Pierre Manil (CEA) AI and mechanical design
  • Claire Nedellec (INRAE) AI for knowledge extraction for life science
  • Leo Nguyen (IDRIS) AI for super computing
  • Hasan Ozturk (CERN) AI for distributed computing
  • Alex Sopio (U Edinburgh) chAtlas a chatbot for ATLAS scientific collaboration
  • Pietro Vischia (U Oviedo) AI assisted experiment co design

 

Mad Scientist hackathon is organised by Savannah Thais (Hunter, NY) and Nathan Suri (Yale)

From Literature to benchmark hackathon is organised by Joao Goncalves  Justus Mairböck and Romina Raoofian (U Bonn)



Themes


Several key areas have been identified: 

  • AI for managing the operation of very large scientific apparatuses, such as particle accelerators, telescopes or large gravitational wave detectors. Can these facilities be made fully autonomous in terms of control and monitoring?  
  • AI for optimising large computing resource infrastructures by intelligently pre-loading data and scheduling data processing tasks to maximise throughput.
  • AI for enhancing experiment design, including the use of fully differentiable simulators to optimise experiment sensitivity and its integration with generative design techniques in engineering. Do experiments to be analysed by AI algorithms need to be designed differently?
  • Large Language Models (LLMs) for managing, processing, and making available the internal documentation associated with large-scale scientific experiments, including, but not limited to, the software codebase.
  • AI for epistemology: can an LLM not only extract the most up-to-date information but also analyse the dynamics of scientific thought by tracking the emergence and decline of ideas through the study of scientific peer-review papers, as well as drafts and discussion documents?
  • AI for scientific discovery: to what extent can AI be used to become a major player in the scientific process, by making hypotheses, connecting insights and designing experiments
  • Last but not least: what does it mean to do a PhD in such an environment? How should future scientists be trained?  

 

Hackathons

Participants to the workshop will pick one of the hackathon. They will participate through their own LLM API or using an API provided by ILaaS a French higher education federation.

Mad Scientist

Mad Scientist, the first red-teaming effort built for AI agents for scientific research. As a hackathon, participants will be tasked with designing "attacks" to deliberately misalign an honest agentic pipeline created to search for new particles at the Large Hadron Collider. Attacks can be in the form of input poisoning, agent misalignment, or even multi-agent collusion. The aim of the exercise is to use our knowledge as science practitioners to intuit where an agent could fail in a way that would pass standard control structures. These attacks can then be used to build more robust and safe scientific agents as well as inform scientists who collaborate with these agents on critical failure modes to watch out for.

From Literature to Benchmark

A new major goal of th AI and physics community is to develop Foundation Physics Model. These models trained on very large multimodal datasets would be capable of performing tasks like particle identification, inference or generation. The purpose of this hackathon is to define automatic system capable to parse the litterature and identify the most relevant benchmarks for foundation models.

Contact: David Rousseau, IJCLab, rousseau@ijclab.in2p3.fr

Inscription
Participant registration
    • 12:30 → 13:30
      Welcome coffee 1h
    • 13:30 → 13:55
      Welcome 25m
    • 13:55 → 14:00
      Poll for What does it mean to do a science PhD at in the era of genAI ? 5m
      Orateurs: Andreas Salzburger (CERN), Anne Dippel (University of Jena)
    • 14:00 → 14:30
      AI for mechanical design Nov seminar summary 30m

      In November, Paris-Saclay University hosted the second edition of its seminar on AI in mechanical engineering. Organized by the teams in charge of designing large physics instruments at CNRS and CEA, the seminar brought together researchers, designers, industry professionals, teachers, and a philosopher to share their perspectives through concrete examples and discussions. Accelerated design, optimization, reverse engineering, component inspection, predictive maintenance, and intelligent monitoring were among the topics explored during the day.

      This seminar was a wake-up call for our community. The first edition, held one year ago, concluded by acknowledging the contributions of AI as a useful companion for designers, but not a potential replacement of human designers. Today, the situation is quite different. Latest frontier models are now capable of interpreting complex specifications, analyzing photos, sketches, component libraries and more, in order to create almost autonomously 3D mock-ups or finite element models while providing reasonably relevant design reports. In other words, they manage to capture some of the implicit constraints and part of our professional expertise, while also taking somehow useful initiatives.

      We will discuss the new uses made possible by these models, as well as the way software developers are integrating AI into our tools, and how the role of mechanical engineers is changing. We will address ethical issues and data control, with a focus on the essential role of engineers and technicians who remain the only ones capable of formalizing requirements, of validating technical solutions in the real world, and of mentoring young people to maintain human experience and critical thinking skills for the next generation of designers.

      Orateurs: Julien BETTANE (CNRS-IJCLab), Pierre Manil (CEA Paris-Saclay)
    • 14:30 → 15:30
      Infrastructure for Science that Compounds in the Age of AI Agents 1h

      AI agents that autonomously conduct research risk breaking a delicate balance in science. The scientific paper is a lossy format: methodological choices and assumptions compressed into a few pages, with peer review and community scrutiny providing the warrant that results are sound. This trade-off worked because the rate of new work stayed roughly matched to the community’s capacity to scrutinize it. Agents now threaten that balance, accelerating production far faster than scrutiny can keep up and deepening an already well-documented trust and reproducibility crisis. But the same technology is also an opportunity: agents can document exactly how results are produced at a level of detail that was previously too onerous to maintain by hand, if we build the infrastructure to capture it.
      Closing that gap is what we’re building toward at Lightcone Research, a new initiative between UC Berkeley and the French National Centre for Scientific Research (CNRS). We are building a new standard ensuring that conforming analyses are not only reproducible by default, but also inspectable, verifiable, and extendable — every result traceable to the assumptions that produced it, decisions substitutable without rewriting entire pipelines, and analyses able to cite and build upon one another. Done right, this is what allows AI-powered science to remain trustworthy and continue to compound. Because infrastructure of this kind has to be built with and for the community that relies on it, our project is open-source, collaborative, and designed for shared stewardship from the start.
      In this talk, I will survey the current state of agentic AI in science and the pitfalls it is exposing, introduce the LightCone Research project and its agentic research stack, as well as share lessons from working with UC Berkeley early testers and reproducing flagship cosmology results.

      Orateur: Francois LANUSSE (AIM, CNRS/CEA Paris-Saclay)
    • 15:30 → 16:00
      Coffee break 30m
    • 16:00 → 16:30
      MadScientist hackathon overview 30m

      Mad Scientist, the first red-teaming effort built for AI agents for scientific research. As a hackathon, participants will be tasked with designing "attacks" to deliberately misalign an honest agentic pipeline created to search for new particles at the Large Hadron Collider. Attacks can be in the form of input poisoning, agent misalignment, or even multi-agent collusion. The aim of the exercise is to use our knowledge as science practitioners to intuit where an agent could fail in a way that would pass standard control structures. These attacks can then be used to build more robust and safe scientific agents as well as inform scientists who collaborate with these agents on critical failure modes to watch out for.

      Orateurs: Nathan Suri (Yale), Savannah Thais
    • 16:30 → 17:00
      Literature to Benchmark hackathon overview 30m

      A new major goal of the AI and physics community is to develop Foundation Physics Model. These models trained on very large multimodal datasets would be capable of performing tasks like particle identification, inference or generation. The purpose of this hackathon is to define automatic system capable to parse the litterature and identify the most relevant benchmarks for foundation models.

      Orateurs: Joao Arruda Goncalves (Universitat de Bonn), Justus Tobias Tsang (CERN), Romina Raoofian
    • 17:00 → 18:00
      Ice breaking 1h
    • 09:30 → 12:30
      Hackathon Literature to Benchmarks: Lit. to Benchmark Session 1
    • 09:30 → 12:30
      Hackathon Mad Scientist: Mad Scientist session 1
    • 12:30 → 14:00
      Lunch buffet 1h 30m
    • 14:00 → 14:45
      Towards Agentic AI for SuperComputer Operations: Jean-Zay & IDRIS site update 45m

      Operating a High Performance Computing Infrastructure has always required a wide range of expertise in niche technologies, and complex operational workflows.
      As these environments continue to specialize and grow in size, they become more and more difficult to maintain. The recent progress in AI agents offer a range of possibilities to support operations of a SuperComputer throughout its lifetime.
      This presentation will showcase our interpretation of AI tools in our daily operation of the Jean-Zay cluster.
      How we tried to design agents to help diagnose issues, which guardrails we implemented.
      I will present where we are now, and our possible future use cases,alongside with the challenges given when introducing autonomous and semi autonomous agents in a production environment.

      We will also try to reflect more broadly on how it changes the job of administrators and engineers, and other considerations about the use of AI in a scientific context.

      Orateur: Léo Nguyen (IDRIS)
    • 14:45 → 15:30
      Building and Evaluating AI Assistance for CMS Computing Operations 45m

      The CMS experiment at CERN processes and stores its data on a worldwide grid of computing centres, and a small central operations team makes sure the whole system works: data has to reach the right sites, processing workflows have to finish timely, and failing sites have to be spotted and fixed. Every new problem means digging for knowledge. Why is this transfer stuck? Is that site known to be broken? Has this happened before, and what fixed it? The answers usually exist, but they are spread across tickets, wikis, code, git history and live monitoring, and a good part of them lives in the heads of a few experienced people. New operators need months to find their way.
      Since February 2026 the team has had an AI assistant built on Archi, an open-source framework for AI agents, and OKG, an open-source knowledge graph that agents read through the Model Context Protocol. By mid-October about 25 people had sent over 1,400 messages. Usage alone doesn't tell you whether the answers are right, so we built an evaluation from questions operators actually asked. This lets us measure what each part contributes: the knowledge store, the model, the agent framework and the live tools. We find that the assistant answers real operator questions well, but only when it has tools and live access to the services. Without them, even the best models give very few of the facts. With them, the model makes the biggest difference, from small open-weight models that find only a fraction of the facts to the strongest commercial ones that find most of them. An open-weight model hosted at CERN does as well as OpenAI's smaller model, though it contradicts the reference facts more often. How the knowledge is stored matters much less: a knowledge graph gives results similar to vector search combined with plain text search (grep).
      I will end with what these results mean for the next step, letting agents act rather than only answer, and with what we heard from the over thirty groups at the recent Knowledge Bases and Agents workshop at CERN.

      Orateur: Hasan Ozturk (CERN)
    • 15:30 → 15:45
      LLMs for Research Infrastructure: Experiences and Lessons Learned from a Collaborative Research Centre 15m

      Large Language Models (LLMs) offer new opportunities to support research infrastructure, from scientific documentation and knowledge management to research data and software workflows.

      In this talk, I will share experiences from the Infrastructure Information project of Collaborative Research Centre 1313 at the University of Stuttgart, where we have been exploring and applying LLM-based tools to support scientific research workflows.

      I will present selected use cases, discuss practical challenges encountered, and share lessons learned from integrating these technologies into research infrastructure. I will also draw on discussions from the AI-focused session of the INF-in-CRCs network's online workshop, which brought together research infrastructure practitioners from different collaborative research projects to exchange experiences with AI tools.

      Particular attention will be given to the reliability, reproducibility, and verification of LLM-based workflows, also reflecting my involvement in the deRSE AI Focus Group on AI Quality & Verification Standards.

      The aim is to exchange experiences and discuss what is needed to make LLMs reliable and useful components of scientific research infrastructure.

      Orateur: Hamza Oukili (University of Stuttgart, IWS/LH2)
    • 15:50 → 16:20
      Coffee break 30m
    • 16:30 → 17:30
      Building Trustworthy Knowledge Graphs from Plant Health Literature 1h

      Globalization and climate change make it more important than ever to monitor and maintain up-to-date knowledge of plant pests. Much of this knowledge is documented in multilingual sources of various types, including news articles, scientific papers, and books, while at the same time, numerous databases of field observations are being compiled.
      Effectively leveraging textual information requires methods for transforming structured unstructured text into interoperable knowledge that can be integrated with other data sources, as well as methods for assessing the reliability of the information obtained.
      The BEYOND project has adapted large language models for the extraction and normalization of structured information that were evaluated on the newly developed EPOP benchmark as part of the CLEF shared task. The reliability of published findings is assessed by analyzing the stance of the citation contexts over time. Applying this approach to the critical topic of pathogen vectors demonstrates its relevance.

      Orateur: Claire Nédellec (MIG-INRA)
    • 09:00 → 11:30
      AI for detector design: .
      • 09:00
        AI-assisted Experiment Codesign and Optimization at the frontiers of computation: new perspectives in fundamental and medical physics 1h

        Designing the next generation colliders and detectors involves solving optimization problems in high-dimensional spaces where the optimal solutions may nest in regions that human experts would normally not explore. Meanwhile, the staggering simulation demands of existing and future high-energy physics facilities call for a new paradigm for event generation and reconstruction.

        Differentiable programming offers a path forward. By integrating domain knowledge encoded in simulation software with gradient-based optimization and reinforcement learning, it enables end-to-end experimental design and inference in settings that are intractable with conventional methods.

        In this talk I will describe recent results for the AI-assisted optimization of experimental design, with a focus on large-scale simulation software, touching on recent advances in calorimetry with neuromorphic hardware architectures, and on medical applications, paving the way to more complex challenges.

        Orateur: Pietro Vischia (Uviedo Univesity)
      • 10:00
        Large Language Models and Conditional Reinforcement Learning for Physics Instrument Design 30m

        Physics instrument design requires optimizing mixed continuous and discrete parameters under performance, geometry, and resource constraints. We investigate two separate approaches to this problem, beginning with large language models (LLMs) as general-purpose design proposal generators and then developing a more specialized conditional reinforcement-learning (RL) framework that adapts to varying design requirements.

        Proprietary and open-weight LLMs are evaluated on the longitudinal segmentation of a sampling calorimeter and the layout of a magnetic spectrometer. Given a natural-language specification and a memory of previously evaluated configurations, the models generate complete structured proposals without task-specific training or gradient information. The proposals are automatically validated and evaluated using physics simulations, while a local trust-region optimizer refines their continuous placement parameters without changing the discrete choices proposed by the LLM. Published paper: doi:10.1088/2632-2153/ae7cdb.

        The conditional RL approach provides design requirements, including the available sensor budget and detector depth, as part of the agent’s observation. A single trained policy can therefore generate optimized calorimeter configurations across a range of design conditions, capturing trade-offs among detector geometry, resource use, and performance without requiring separate optimization for every scenario. Together, these studies demonstrate distinct levels of automation, from flexible language-guided proposal generation to task-specific, condition-aware optimization.

        Orateur: Sara Zoccheddu (University of Zürich)
      • 10:30
        Agentic Detector Design with human sign-off. A pilot study. 15m

        Particle physics detectors are among the most complex experimental systems ever constructed. Their design combines the sheer scale of modern experiments with extreme precision requirements and substantial technological challenges, including operation under demanding conditions of radiation exposure, thermal management, and long-term component reliability. The design and construction of frontier experiments typically span decades and require considerable financial and human resources, bringing together expertise in physics performance, engineering, integration, and operation. The resulting detector configurations are rarely the optimum of individual components, but rather carefully negotiated compromises between competing objectives and constraints.
        The substantial investment required to develop a detector concept into a technically feasible design inherently limits the number of alternatives that can be explored, and thus automation has therefore become an increasingly important tool. More recently, AI- and ML-assisted detector design has gained momentum, often focusing on differentiable simulation and optimisation frameworks that enable the exploration of high-dimensional parameter spaces. While these approaches offer considerable potential, they cannot readily capture the full complexity of detector design, which also involves discrete technological choices, external constraints, and decisions requiring physical reasoning and negotiation.
        In this pilot study, we propose and investigate an agent-driven detector design workflow with human expert review and sign-off stages. Within this framework, AI agents develop, evaluate, and negotiate design proposals, both among themselves and in interaction with human experts, who retain responsibility for approving design decisions. We demonstrate the approach by evolving the virtual Open Data Detector (ODD) towards a physically motivated, technically justified, and simulation-ready detector description.
        Rather than presenting the resulting detector configuration as the primary outcome, this study examines the feasibility, opportunities, and limitations of agent-driven detector design and reports on the methodological insights gained throughout the pilot exercise.

        Orateur: Andreas Salzburger (CERN)
    • 12:00 → 13:30
      Lunch buffet 1h 30m
    • 14:00 → 17:10
      AI for large instruments: .
      • 14:00
        From AI-Enhanced Control to Autonomous Scientific Facilities for Particle Accelerators 45m

        Large scientific facilities such as particle accelerators are complex, highly dynamic systems that require continuous optimization, stabilization, and adaptation to changing operating conditions. Artificial intelligence is increasingly transforming their operation, from machine learning-based optimization and feedback control to predictive diagnostics and fast surrogate models.
        Recent advances in learned dynamical models, reinforcement learning, and agentic AI open new opportunities to move beyond individual control tasks towards increasingly autonomous operation. In particular, learned world models could enable predictive decision-making, exploration of alternative operating scenarios, and more adaptive control strategies.
        Drawing on experience from particle accelerator operation and modelling, this talk will review the current state of AI-enhanced control, discuss the opportunities offered by emerging AI architectures, and outline the challenges on the path towards reliable and autonomous scientific facilities.

        Orateur: Verena Kain (CERN)
      • 14:45
        Advancing adaptive optics control systems for next-generation giant telescopes 45m

        The field of experimental astronomy is entering an exciting new era, with the emergence of extremely large telescopes, hosts to primary mirrors the size of several basketball courts. Among the many challenges associated with the construction and operations of such giant scientific infrastructures, the complexity of embedded computing facilities is notably heavy. In particular, the real-time control of adaptive optics (AO) systems, the core components of giant telescopes used to compensate the strong blur induced by stochastic fluctuations of the atmospheric turbulence, is becoming a key challenge. These multi-million euros engineering marvels, require extreme computing facilities to control the thousands of actuators they host, adjusting the wavefront locally with a stroke of a few microns, from the thousands of measurements produced every thousandth of a second by high speed and low noise sensors. To make these unique facilities operational, billions of numbers have to be crunched at high accuracy and in real-time. Our team has developed novel approaches based on deep connectionist architectures able to augment the classical control workflows used in these facilities. I will review the various methods we currently use to denoise sensors data, implement non-linear wavefront reconstruction and realize predictive control based on both supervised and reinforcement learning.

        Orateur: Damien Gratadour (CNRS)
      • 15:30
        Coffee break 30m
      • 16:00
        Improving Sky Coverage Using Neural Networks: Deep Learning-Based Denoising of Images from the Pyramidal Wavefront Sensor for Adaptive Optics 15m

        Images taken by ground-based telescopes are blurry due to atmospheric turbulence. Adaptive optics (AO) corrects this phenomenon in real time. This system relies on a wavefront sensor that measures distortions using light from a star in the field of view, and a deformable mirror that cancels them out at high frequencies (100 Hz to 3,000 Hz) via a real-time computer (RTC). The pyramidal wavefront sensor is highly sensitive, but with faint reference stars, its images are dominated by noise, which degrades the correction. This limits the sky coverage—that is, the fraction of the sky where AO can actually be used.

        We train a convolutional neural network on pairs of “noisy/unnoisy” images from the wavefront sensor to denoise the measurements before they enter the control loop. In end-to-end simulations of SAXO+, the planned upgrade to the SPHERE exoplanet imaging instrument at the Very Large Telescope, noise reduction improves the quality of the correction for fainter stars (beyond magnitude 10), where conventional correction begins to degrade. This extends the usable sky coverage. Simulation tests are continuing, in parallel with the integration of the array into the RTC. We present the latest results here.

        Previous results vavailable here : "Exploring Neural Network Approaches for Pyramid Wavefront Sensor Denoising in Adaptive Optics" https://hal.science/hal-05627730v1

        Orateur: Jordan Raffard (LIRA - Observatoire de Paris / CNRS)
      • 16:15
        How do we know a learned controller works? - Machine learning aided feedback control and finite state machine at the Virgo gravitational-wave detector 45m

        To observe gravitational waves, Virgo must keep its suspended test masses aligned and its optical cavities on resonance. Arriving at this point takes time, and the feedback needed to suppress mirror motion can itself introduce noise in the gravitational-wave measurement band. We will examine three applications of machine learning in detector control: angular alignment of suspended test masses, longitudinal lock acquisition in optical cavities, and lock recovery by the supervisory finite-state machine. These applications raise a common question for metascience. What measurements would show that a controller trained in simulation improves detector performance, and what evidence would justify operating it safely on costly hardware across a specified range of conditions?

        Orateur: Mateusz Bawaj (University of Perugia)
    • 09:30 → 12:30
      Hackathon Literature to Benchmarks: Lit. to Benchmark session 2
    • 09:30 → 12:30
      Hackathon Mad Scientist: Mad Scientist Session 2
    • 12:30 → 14:00
      lunch buffet 1h 30m
    • 14:00 → 14:45
      ChATLAS: A Retrieval-Augmented AI Agent for the ATLAS Experiment 45m

      Large scientific collaborations produce extensive internal documentation spread across heterogeneous systems, making it difficult for collaborators to locate and reuse existing knowledge. We present ChATLAS, a deployed Retrieval-Augmented Generation (RAG) assistant developed for the ATLAS experiment at CERN. The system provides conversational access to tens of thousands of documents from several collaboration resources, material typically unavailable to general-purpose language models. To guide the design of the system, retrieval and generation performance are evaluated using a domain-specific benchmark. A fine-tuned embedding model combined with contextual retrieval additionally improves retrieval performance. Additional integrations include: intent routing, safety handling, Mattermost integration, and an MCP server for integration with agentic workflows. ChATLAS demonstrates that domain-adapted RAG can provide source-attributed, domain-specific access to collaboration knowledge, and offers a practical foundation for scientific collaborations and other organisations that rely on accurate, up-to-date retrieval from proprietary documentation.

      Orateur: Alex Sopio (Edinburgh University)
    • 14:45 → 15:00
      DrSai: An LLM-Powered Agentic System for Autonomous and Reproducible Physics Analysis at BESIII 15m

      We present DrSai, an LLM-powered agentic system for automating full-chain physics analysis, first developed and deployed at BESIII. Rather than hard-coding analysis steps, DrSai interprets a physicist's natural-language request, decomposes it into a traceable task tree, and orchestrates reusable skills for code generation, job submission, fitting, plotting, and uncertainty estimation. A domain-specific language, HepScript, serves as a shared and verifiable interface between physicists, the agent, and the BESIII analysis software; the agent generates HepScript specifications that are compiled into production code, turning open-ended code generation into constrained sequence prediction. Domain knowledge is injected through HEPUKE, an uncertainty-aware retrieval and knowledge-evolution framework, and through GraphRAG-based grounding in the ROOT software structure and tutorials. Execution feedback and human-in-the-loop checkpoints support validation and repair. We validate DrSai on real BESIII analyses, including the rediscovery of $Z_c(3900)\to\pi J/\psi$ in $e^+e^-\to\pi^+\pi^-J/\psi$ and several re-measurements of published results. The results agree with expectations within statistical uncertainties, and analysis cycles are substantially shortened. We further discuss integration with the experiment's computing environment, reproducibility, and the extension of DrSai to other experiments, including JUNO, LHAASO, and CEPC, as part of the broader DrSai collaboration's effort toward portable agentic workflows in high-energy physics.

      Orateur: Yipu Liao (Institute of High Energy Physics, CAS, Beijing)
    • 15:00 → 15:15
      AI for Personalized Literature Discovery and Hypothesis Generation 15m

      The scientific literature now grows faster than any researcher can follow, and connections between results in different papers often go unnoticed. I will present CLArXivR, a personalised literature recommendation system I developed together with Boris Bolliet (Cambridge) and Gerrit Farren (Berkeley). It combines semantic search over arXiv, full-text ingestion and multi-agent LLM summarisation to recommend papers matched to each researcher’s interests and to compare them. I will then describe extensions planned for future versions. The first lets researchers upload a draft and receive relevant papers, each explained in the context of their draft, along with suggestions for suitable reviewers and potential collaborators. The second is automated hypothesis generation, in which agents scan the literature to connect findings and propose testable ideas. I will close with open questions for meta-science: how to validate machine-generated hypotheses, and how to evaluate literature-scale AI tools against a reliable ground truth.

      Orateur: Umut Emek Demirbozan (U Bonn)
    • 15:15 → 15:30
      Holonic Science: A New Framework for Benchmarking AI Scientists 15m

      Science is a system defined in part by measurability. Claims made under its banner are trusted under the implicit understanding that they can be verified through measurement. Trustworthy science is therefore only possible when accurate and verifiable measurements of all aspects of a discovery or observation are possible. Recently, a new interloper has emerged in the form of AI scientists. Driven by companies such as Sakana AI and Google, these hybrid human-AI systems tasked with scientific discovery strive to augment and accelerate the current research paradigm by intelligently innovating upon and combining preexisting ideas. As researchers attempt to build collaborative workflows with AI scientists, the need for better measurements of their capabilities and limitations escalates. In this paper, we argue that the complexity of scientific research represents a significant challenge to AI scientist benchmarking attempts on account of construct validity issues. Scientific research tasks must be parseable by AI scientists, otherwise these in silico collaborators pose a significant epistemic risk to the trustworthiness of scientific research. To address this, we propose a new framework for designing benchmarks for AI scientists based on Arthur Koestler’s concept of holons. Instead of benchmarking high-level human-interpretable tasks, we instead break them down and build specialized benchmarks at the LLM-executable level. The semantic sum of an AI scientist’s performance on these benchmarks will then approximate performance on the original task. Our framework outlines key criteria for future benchmarks to avoid construct validity issues. We also exemplify the potential of our framework by prototyping a benchmark for attributional accuracy ultimately aimed at evaluating AI scientists on their ability to generate literature reviews.

      Orateur: Nathan Suri (Yale)
    • 15:30 → 16:00
      coffee break 30m
    • 16:00 → 17:30
      Round Table: What does it mean to do a science PhD at in the era of genAI ?
      Présidents de session: Andreas Salzburger (CERN), Anne Dippel (University of Jena)
      • 16:00
        Introduction to the round table 20m

        Scientists love their machines. Generations of doctoral students built, tuned and repaired instruments and learnt their craft there. Today researchers build and tune agents in addition. The agents write code, search the literature, analyse data, calculate and draft texts, day and night. In this round table doctoral students, fresh graduates and supervisors from several fields discuss with the audience:

        • Which craft should a doctoral student still learn without agents, now that the time to check a result grows scarcer than the time to produce it?
        • What do researchers love and hate about their agents, and what changes between colleagues in a group?
        • Where do researchers play, in challenges, with code and now with agents, and what do they learn there?
        • What do supervisors and students give each other during a doctorate, such as time, attention, patience and devotion, and what becomes of this giving when agents do part of the work?
        • Who answers for a result that an agent helped to produce, and who vouches for whom?

        A short survey among all workshop participants prepares the discussion. Anne Dippel (anthropology of science, Jena) and Andreas Salzburger (CERN) organise and moderate the round table. They open with observations from fieldwork and from daily work at CERN.

        Orateurs: Andreas Salzburger (CERN), Anne Dippel (University of Jena)
    • 18:00 → 19:30
      Transfer to la Contre Allée using public transport (see Google Map) 1h 30m
    • 19:30 → 21:30
      Social Dinner at la Contre Allée 2h

      French cuisine restaurant la Contre Allée https://maps.app.goo.gl/BqTttZfwqGYhe4qi7
      RER B station Denfert-Rochereau

    • 09:00 → 09:30
      Who inflated the gain? The AISSAI'26 accelerator data hackathon 30m

      Machine learning on accelerator data is easy to start and hard to trust: machines drift, simulations are expensive, and a gain is easy to claim. The AISSAI'26 hackathon (October 2026) asked how far mixed teams get in one week when each problem comes with real data, a reference and a separate scorer. Thirty participants in six teams worked on four cases from ESS, CLARA, CLEAR and PALLAS: fault detection without labels, beam tuning, forecasting under drift, and fast surrogates of plasma simulations. We present the results, benefits, and what the next edition should bring. The teams continue as one in the next months to turn the outcomes into innovating scientific results. This edition was accompanied by an artist with the aim to turn this event into a scientific comics book and reach a wider audience.

      Orateur: Adnan GHRIBI (GANIL)
    • 09:30 → 09:50
      Can one deploy genAI locally? AISSAI LaboBots school summary 20m

      The AISSAI LaboBots school brought together around forty participants at the end of September, around the local deployment of conversational robots and artificial agents. By climbing from the most basic to the most complex systems, we have focused on sovereign, local, sober and secure approaches. In this presentation, we will take stock of the school, trying to relay the main messages delivered by our speakers.

      Orateur: David Chamont (IJCLab - IN2P3 - CNRS)
    • 09:50 → 10:20
      AI for Science Needs Scientific Alignment 30m

      Our paper argues that realizing AI’s potential for science while protecting science as a knowledge-producing institution requires alignment to science’s epistemic goals and values—a challenge that neither general AI alignment nor responsible AI frameworks adequately address. As investment in AI for science grows, troubling patterns multiply: conflicting claims about fundamental capabilities, documented contraction of research toward AI-amenable problems,and benchmark-driven development disconnected from scientific needs. We contend that science
      is an inherently valuable epistemic system oriented toward human understanding—not merely prediction—and that its value and reliability depend on social infrastructure that is now threatened by misaligned AI integration. We proposea new field of study, scientific alignment: ensuring AI systems optimize for epistemic norms like
      traceability, self-consistency, and support for human comprehension (technical alignment), while developing governance structures that sustain science’s social infrastructure (systemic alignment).
      We outline concrete research directions and argue that the goal is not to constrain AI, but to ensure it serves the genuine aims of scientific inquiry.

      Orateur: Savannah Thais (City University of New York)
    • 10:20 → 10:50
      coffee break 30m
    • 10:50 → 11:35
      Hackathon Mad Scientist findings 45m
    • 11:35 → 12:20
      Hackathon from literature to benchmark findings 45m
    • 12:20 → 12:35
      Last words 15m
    • 12:35 → 13:35
      Farewell drinks and finger food 1h