BEGIN:VCALENDAR VERSION:2.0 PRODID:-//CERN//INDICO//EN BEGIN:VEVENT SUMMARY:4MHEP or Domain shift or idea DTSTART;VALUE=DATE-TIME:20260522T101500Z DTEND;VALUE=DATE-TIME:20260522T110000Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12468@partphys-indico.unige.ch DESCRIPTION:Speakers: Pradyun Hebbar (University of Geneva)\nhttps://partp hys-indico.unige.ch/event/2147/contributions/12468/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12468/ END:VEVENT BEGIN:VEVENT SUMMARY:Optimal detector design for the FASERν upgrade DTSTART;VALUE=DATE-TIME:20260522T093000Z DTEND;VALUE=DATE-TIME:20260522T101500Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12471@partphys-indico.unige.ch DESCRIPTION:Speakers: Vincent Riechers (University of Geneva)\nIn this pro ject I am exploring optimal detector design for the FASERν upgrade. For d ifferent detector designs\, classifiers are trained and evaluated\, and a surrogate model learns the relation between detector parameters and classi fier performance. This surrogate is then used to suggest new detector desi gns for the next optimization step.\n\nThe main challenge is the trade-off between classifier performance\, computational cost\, and fairness betwee n detector designs. I want to make sure that the classifier does not accid entally favor one type of detector because of choices such as pretraining\ , architecture\, or training setup. This likely requires moving toward a c o-design approach\, where detector geometry and classifier strategy are de veloped together. I would like to discuss my first results and maybe the r ight next steps.\n\nhttps://partphys-indico.unige.ch/event/2147/contributi ons/12471/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12471/ END:VEVENT BEGIN:VEVENT SUMMARY:universal latent space for HEP DTSTART;VALUE=DATE-TIME:20260522T081500Z DTEND;VALUE=DATE-TIME:20260522T090000Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12470@partphys-indico.unige.ch DESCRIPTION:Speakers: Kinga Wozniak ()\nWe investigate whether HEP event r epresentations\, despite being trained on disparate tasks and architecture s\, converge to a common latent geometry — a universal event space whose structure reflects the physical content of collisions rather than the spe cifics of any single model. Building on recent work suggesting platonic co nvergence in vision and language embeddings\, we develop unsupervised alig nment techniques to map between independently trained event embeddings and characterize the shared geometric structure they expose.\n\nhttps://partp hys-indico.unige.ch/event/2147/contributions/12470/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12470/ END:VEVENT BEGIN:VEVENT SUMMARY:Toward a Community Skill Registry for HEP Agents DTSTART;VALUE=DATE-TIME:20260521T141500Z DTEND;VALUE=DATE-TIME:20260521T150000Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12469@partphys-indico.unige.ch DESCRIPTION:Speakers: Giovanni Ottaviano (Sorbonne University (FR))\nLLM-b ased agents are rapidly entering HEP workflows — automating analysis pip elines\, assisting experimental design\, and accelerating theory-experimen t comparison. Yet each new project solves the same underlying problem in i solation: teaching the agent what a HEP analysis actually *is*. What seque nce of steps takes a physics goal to a publishable result? Which control a nd validation regions need to be defined\, and why? What does a properly b linded statistical model look like? What deliverables — yield tables\, s ystematic breakdowns\, limit curves\, HepData entries — does the communi ty expect at the end?\n\nThese are not only software questions\; they enco de decades of accumulated experimental and phenomenological practice\, and LLMs have no reliable way to reconstruct them from general training. More over\, this knowledge is not universal: ATLAS\, CMS\, and LHCb each have t heir own conventions for object definitions\, working points\, systematic taxonomies\, and approval workflows — differences that are meaningful an d that an agent must respect to produce results credible within a given co llaboration. The result is agents that can write syntactically correct cod e but reproduce analyses that violate basic methodology — wrong blinding policy\, missing systematic sources\, non-standard statistical treatment — in ways that may not be obvious without domain expertise.\n\nThe commu nity has robust shared infrastructure for preserving and communicating *re sults* — HepData\, RIVET\, INSPIRE — but nothing equivalent for encodi ng *methodology* in a form agents can use. Every HEP-agent project re-impl ements its own understanding of what an analysis should look like\, and ev ery implementation drifts.\n\nWe propose addressing this gap with a commun ity-owned registry of agent skills for HEP: structured\, version-pinned kn owledge packages that encode the standard analysis workflow\, its delivera bles\, and its validation criteria. Beyond immediately reducing hallucinat ion rates\, such a registry opens a concrete research direction: how shoul d physics methodology be represented for agents\, how do we evaluate wheth er an agent has followed it correctly\, and how does the quality of that r epresentation affect downstream physics results?\n\nhttps://partphys-indic o.unige.ch/event/2147/contributions/12469/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12469/ END:VEVENT BEGIN:VEVENT SUMMARY:Toward a Universal Data-Driven Background Estimate for Scalable Re sonance Searches DTSTART;VALUE=DATE-TIME:20260521T131500Z DTEND;VALUE=DATE-TIME:20260521T140000Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12465@partphys-indico.unige.ch DESCRIPTION:Speakers: Theresa Reisch (University of Geneva)\nAs part of th e RESONEX program — a planned multi-lens narrow-resonance search across ~1\,500 invariant mass histograms on ATLAS Run-3 data — we are exploring how to build a single\, lens-agnostic background estimation framework tha t could scale to this many channels without per-channel hand-tuning.\n\nTh e current thinking is a two-tier\, fully data-driven approach: Gaussian Pr ocess Regression with a Matérn-5/2 kernel as the primary method\, providi ng both a background estimate and a built-in uncertainty band\, with slidi ng-window bin averaging as an automated fallback where GPR fails validatio n. Many open questions remain — kernel and length-scale choices\, how to prevent signal absorption at scale\, hyperparameter strategies across het erogeneous histograms\, and how to calibrate signal contamination in a len s-independent way.\n\nI would like to use this talk to share where our thi nking stands\, sketch the validation strategy we have in mind\, and open a discussion on the design choices and pitfalls.\n\nhttps://partphys-indico .unige.ch/event/2147/contributions/12465/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12465/ END:VEVENT BEGIN:VEVENT SUMMARY:Anomaly detection in HEP in the era of agents DTSTART;VALUE=DATE-TIME:20260522T073000Z DTEND;VALUE=DATE-TIME:20260522T081500Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12467@partphys-indico.unige.ch DESCRIPTION:Speakers: Ivan Oleksiyuk (University of Geneva)\nRunning a rea l AD analysis in ATLAS teaches you quickly how many things can go wrong: d ata reuse creates spurious excesses\, false positive rates never quite rea ch the required threshold\, and any pipeline imperfection gets amplified i nto a fake signal. As the field considers agentic AI for broader automated searches\, these problems do not disappear — they accelerate. I want to discuss what needs to be solved first\, and where agents can already help today.\n\nhttps://partphys-indico.unige.ch/event/2147/contributions/12467 / LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12467/ END:VEVENT BEGIN:VEVENT SUMMARY:Iterative Reconstruction of Event DTSTART;VALUE=DATE-TIME:20260521T123000Z DTEND;VALUE=DATE-TIME:20260521T131500Z DTSTAMP;VALUE=DATE-TIME:20260520T130824Z UID:indico-contribution-2852-12466@partphys-indico.unige.ch DESCRIPTION:Speakers: Andreas M. Hermansen (Universite de Geneve)\nI'm goi ng to present\, Pairton\, my latest work on reconstruction of full hadroni c ttbar events.\n\nhttps://partphys-indico.unige.ch/event/2147/contributio ns/12466/ LOCATION:Villa Boninchi URL:https://partphys-indico.unige.ch/event/2147/contributions/12466/ END:VEVENT END:VCALENDAR