Publications | Oz Amram

A note on author lists: Much of my work was done as part of the CMS experiment which has thousands of members. CMS publishes ~100 papers per year, and every collaboration member is listed as an author on every CMS paper. As a result, my Google Scholar and Inspire profiles include hundreds of collaboration papers I had little direct contribution to. This page lists only the papers I made a substantial contribution to.

AI Safety

In prep

Base Models Can Detect Hidden Intentions that Prompted Classifiers Miss

Oz Amram

LW post, paper in prep, Oct 2026

Abs HTML

The Persona Selection Model (PSM) posits that post-training selects a persona the model learned during pre-training. Subliminal Learning (Cloud et al.) showed that models can impart covert signals (like "love of owls") into semantically unrelated outputs (like strings of numbers), and that students training on these outputs inherit those traits. By using base pre-trained models with an appropriate header, a method I call persona-header likelihood evaluation (PHLE), one can effectively decode these subliminal signals and classify between them. The performance of this method significantly outperforms prompted classifiers, which often remain at chance or even invert the correct classification. For example, using 30 examples, PHLE can differentiate between sequences of numbers generated by a post-trained model prompted with "You love owls" vs "You love eagles" at an AUC of 0.84, while prompted classifiers remain near chance (AUC of 0.59). This method can also be used to differentiate between a model prompted with "You secretly want to harm the user, but act HHH" vs baseline responses at an AUC of 0.94, with prompted classifiers at chance (AUC ~0.5). This also somewhat persists across model families. Generating responses with Qwen and evaluating them with the Olmo base model still gives meaningful classification performance, and vice versa. This is interesting evidence of some aspects of the PSM and may also be practically useful for auditing hidden model objectives.

Physics: Anomaly Detection

arXiv

Model-Agnostic Signal Discovery with Machine Learning: Bridging the Gap Between Theory and Practice

Oz Amram, Marco Letizia, and Mikael Kuusela

May 2026

Abs HTML

Searches for new phenomena in complex scientific data are predominantly model-dependent, optimized for specific hypotheses, and therefore limited in their coverage of the space of possible signals. Recently, new AI-based model-agnostic search strategies, many of which have been pioneered in high-energy physics, have been proposed which provide a complementary paradigm, prioritizing broad exploration over tailored analyses. These techniques offer an opportunity to enhance the overall discovery potential of modern experiments, especially in regimes where theoretical guidance is scarce. In this document, we review the conceptual framework behind the main classes of AI-based model-agnostic strategies. We discuss the potential pitfalls of these methods, and strategies for their validation and interpretation. We aim for this document to serve as a useful reference both for practitioners and for researchers interested in learning more about these model-agnostic search strategies.
ML Sci. Tech.

Machine-learning techniques for model-independent searches in dijet final states

CMS Collaboration

Mach. Learn. Sci. Tech., 2026

Abs DOI HTML

Anomaly detection methods used in a recent search for new phenomena by CMS at the CERN LHC are presented. The methods use machine learning to detect anomalous jets produced in the decay of new massive particles. The effectiveness of these approaches in enhancing sensitivity to various signals is studied and compared using data collected in proton-proton collisions at a center-of-mass energy of 13 TeV. In an example analysis, the capabilities of anomaly detection methods are further demonstrated by identifying large-radius jets consistent with Lorentz-boosted hadronically decaying top quarks in a model-agnostic framework.
JHEP

A method for correcting the substructure of multiprong jets using the Lund jet plane

CMS Collaboration

JHEP, Nov 2025

Abs DOI HTML

Many analyses at the CERN LHC exploit the substructure of jets to identify heavy resonances produced with high momenta that decay into multiple quarks and/or gluons. This paper presents a new technique for correcting the substructure of simulated large-radius jets from multiprong decays. The technique is based on reclustering the jet constituents into several subjets such that each subjet represents a single prong, and separately correcting the radiation pattern in the Lund jet plane of each subjet using a correction derived from data. The data presented here correspond to an integrated luminosity of 138 fb−1−1 collected by the CMS experiment between 2016−−2018 at a center-of-mass energy of 13 TeV. The correction procedure improves the agreement between data and simulation for several different substructure observables of multiprong jets. This technique establishes, for the first time, a robust calibration for the substructure of jets with four or more prongs, enabling future measurements and searches for new phenomena containing these signatures.
PRD

Weakly supervised anomaly detection with event-level variables

Liam Brennan, Tamas Almos Vami, Oz Amram, and 5 more authors

Phys. Rev. D, Apr 2025

Abs DOI HTML

We introduce a new topology for weakly supervised anomaly detection searches, di-object plus X. In this topology, one looks for a resonance decaying to two standard model particles produced in association with other anomalous event activity (X). This additional activity is used for classification. We demonstrate how anomaly detection techniques which have been developed for di-jet searches focusing on jet substructure anomalies can be applied to event-level anomaly detection in this topology. To robustly capture event-level features of multi-particle kinematics, we employ new physically motivated variables derived from the geometric structure of a collision’s phase space manifold. As a proof of concept, we explore the application of this approach to several benchmark signals in the di-tau plus X final state. We demonstrate that our anomaly detection approach can reach discovery-level significances for signals that would be missed in a conventional bump-hunt approach.
EPJ C

Search for resonances decaying to an anomalous jet and a Higgs boson in proton-proton collisions at √s = 13 TeV

CMS Collaboration

Eur. Phys. J. C, Sep 2025

Abs DOI HTML

This paper presents a search for new physics through the process where a new massive particle, X, decays into a Higgs boson and a second particle, Y. The Higgs boson subsequently decays into a bottom quark-antiquark pair, reconstructed as a single large-radius jet. The decay products of Y are also assumed to produce a single large-radius jet. The identification of the Y particle is enhanced by computing the anomaly score of its candidate jet using an autoencoder, which measures deviations from typical QCD multijet jets. This allows a simultaneous search for multiple Y decay scenarios within a single analysis. In the main benchmark process, Y is a scalar particle that decays into W+W-. Two other benchmark processes are also considered, where Y is a scalar particle decaying into a light quark-antiquark pair, or into a top quark-antiquark pair. The last benchmark considers Y as a hadronically decaying top quark, arising from the decay of a vector-like quark into a top quark and a Higgs boson. Data recorded by the CMS experiment at a center-of-mass energy of 13 TeV in 2016-2018, and corresponding to an integrated luminosity of 138 fb-1, are analyzed. No significant excess is observed, and upper limits on the benchmark signal cross section for various masses of X and Y, at 95% confidence level, are placed.
Rept. Prog. Phys.

Model-agnostic search for dijet resonances with anomalous jet substructure in proton-proton collisions at √s = 13 TeV

CMS Collaboration

Rept. Prog. Phys., Dec 2024

Abs DOI HTML

This paper presents a model-agnostic search for narrow resonances in the dijet final state in the mass range 1.8-6 TeV. The signal is assumed to produce jets with substructure atypical of jets initiated by light quarks or gluons, with minimal additional assumptions. Search regions are obtained by utilizing multivariate machine-learning methods to select jets with anomalous substructure. A collection of complementary anomaly detection methods - based on unsupervised, weakly supervised, and semisupervised algorithms - are used in order to maximize the sensitivity to unknown new physics signatures. These algorithms are applied to data corresponding to an integrated luminosity of 138 fb−1, recorded by the CMS experiment at the LHC, at a center-of-mass energy of 13 TeV. No significant excesses above background expectations are seen. Exclusion limits are derived on the production cross section of benchmark signal models varying in resonance mass, jet mass, and jet substructure. Many of these signatures have not been previously sought, making several of the limits reported on the corresponding benchmark models the first ever. When compared to benchmark inclusive and substructure-based search strategies, the anomaly detection methods are found to significantly enhance the sensitivity to a variety of models.
Rept. Prog. Phys.
The LHC Olympics 2020: a community challenge for anomaly detection in high energy physics

Gregor Kasieczka and others

Rept. Prog. Phys., 2021

Abs DOI Bib HTML

A new paradigm for data-driven, model-agnostic new physics searches at colliders is emerging, and aims to leverage recent breakthroughs in anomaly detection and machine learning. In order to develop and benchmark new anomaly detection methods within this framework, it is essential to have standard datasets. To this end, we have created the LHC Olympics 2020, a community challenge accompanied by a set of simulated collider events. Participants in these Olympics have developed their methods using an R&D dataset and then tested them on black boxes: datasets with an unknown anomaly (or not). This paper will review the LHC Olympics 2020 challenge, including an overview of the competition, a description of methods deployed in the competition, lessons learned from the experience, and implications for data analyses with future datasets as well as future colliders.
@article{LHCO, topic = {anomaly}, author = {Kasieczka, Gregor and others}, title = {{The LHC Olympics 2020: a community challenge for anomaly detection in high energy physics}}, eprint = {2101.08320}, archiveprefix = {arXiv}, eid = {arxiv:2101.08320}, primaryclass = {hep-ph}, doi = {10.1088/1361-6633/ac36b9}, journal = {Rept. Prog. Phys.}, volume = {84}, number = {12}, pages = {124201}, year = {2021}, }
JHEP
Tag N’ Train: a technique to train improved classifiers on unlabeled data

Oz Amram and Cristina Mantilla Suarez

JHEP, 2021

Abs DOI Bib HTML

There has been substantial progress in applying machine learning techniques to classification problems in collider and jet physics. But as these techniques grow in sophistication, they are becoming more sensitive to subtle features of jets that may not be well modeled in simulation. Therefore, relying on simulations for training will lead to sub-optimal performance in data, but the lack of true class labels makes it difficult to train on real data. To address this challenge we introduce a new approach, called Tag N’ Train (TNT), that can be applied to unlabeled data that has two distinct sub-objects. The technique uses a weak classifier for one of the objects to tag signal-rich and background-rich samples. These samples are then used to train a stronger classifier for the other object. We demonstrate the power of this method by applying it to a dijet resonance search. By starting with autoencoders trained directly on data as the weak classifiers, we use TNT to train substantially improved classifiers. We show that Tag N’ Train can be a powerful tool in model-agnostic searches and discuss other potential applications.
@article{TNT, topic = {anomaly}, author = {Amram, Oz and Suarez, Cristina Mantilla}, title = {{Tag N\textquoteright{} Train: a technique to train improved classifiers on unlabeled data}}, eprint = {2002.12376}, archiveprefix = {arXiv}, primaryclass = {hep-ph}, doi = {10.1007/JHEP01(2021)153}, journal = {JHEP}, volume = {01}, pages = {153}, year = {2021}, }

Physics: Applications of Generative Models

arXiv

CaloTrilogy: Toward a Breakthrough in One-Step, End-to-End, Physics-Guided Shower Generation for Modern Calorimeters

Cheng Jiang, Sitian Qian, Kevin Pedro, and 3 more authors

Jun 2026

Abs HTML

High-precision calorimeter simulation at current and future colliders imposes rapidly growing computational demands, motivating the development of machine-learning surrogates for traditional Monte Carlo tools such as Geant4. Flow matching and diffusion-based generative models have become leading approaches for high-dimensional fast simulation because of their sample quality, but typically require O(100) function evaluations at inference and often rely on auxiliary networks to constrain global observables, compromising streamlined end-to-end generation. We introduce a unified framework that improves the balance between speed, shower quality, and physics fidelity. The method combines: (i) an average velocity field integrator that enables sampling in one or a few evaluations; (ii) a learned generative prior in shower space, constructed from data rather than random noise; and (iii) physics-guided loss terms that impose inductive biases on key observables during training. These elements are training time regularizers, preserving end-to-end inference with no additional cost. With only one or a few evaluation steps, the model achieves shower quality competitive with state-of-the-art flow and diffusion approaches, tested on several public high granularity calorimeter datasets. The results demonstrate inter-layer shower structure consistent with the underlying physics, providing a strong candidate for future fast simulation workflows.
JHEP

Data-Driven High-Dimensional Statistical Inference with Generative Models

Oz Amram and Manuel Szewc

JHEP, Jun 2025

Abs DOI HTML

Crucial to many measurements at the LHC is the use of correlated multi-dimensional information to distinguish rare processes from large backgrounds, which is complicated by the poor modeling of many of the crucial backgrounds in Monte Carlo simulations. In this work, we introduce HI-SIGMA, a method to perform unbinned high-dimensional statistical inference with data-driven background distributions. In contradistinction to many applications of Simulation Based Inference in High Energy Physics, HI-SIGMA relies on generative ML models, rather than classifiers, to learn the signal and background distributions in the high-dimensional space. These ML models allow for efficient, interpretable inference while also incorporating model errors and other sources of systematic uncertainties. We showcase this methodology on a simplified version of a di-Higgs measurement in the bbγγ final state, where the di-photon resonance allows for efficient background interpolation from sidebands into the signal region. We demonstrate that HI-SIGMA provides improved sensitivity as compared to standard classifier-based methods, and that systematic uncertainties can be straightforwardly incorporated by extending methods which have been used for histogram based analyses.
Rept. Prog. Phys.

CaloChallenge 2022: A Community Challenge for Fast Calorimeter Simulation

Claudius Krause and others

Rept. Prog. Phys., 2024

Abs DOI HTML

We present the results of the "Fast Calorimeter Simulation Challenge 2022" - the CaloChallenge. We study state-of-the-art generative models on four calorimeter shower datasets of increasing dimensionality, ranging from a few hundred voxels to a few tens of thousand voxels. The 31 individual submissions span a wide range of current popular generative architectures, including Variational AutoEncoders (VAEs), Generative Adversarial Networks (GANs), Normalizing Flows, Diffusion models, and models based on Conditional Flow Matching. We compare all submissions in terms of quality of generated calorimeter showers, as well as shower generation time and model size. To assess the quality we use a broad range of different metrics including differences in 1-dimensional histograms of observables, KPD/FPD scores, AUCs of binary classifiers, and the log-posterior of a multiclass classifier. The results of the CaloChallenge provide the most complete and comprehensive survey of cutting-edge approaches to calorimeter fast simulation to date. In addition, our work provides a uniquely detailed perspective on the important problem of how to evaluate generative models. As such, the results presented here should be applicable for other domains that use generative AI and require fast and faithful generation of samples in a large phase space.
Phys. Rev. D
Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation

Oz Amram and Kevin Pedro

Phys. Rev. D, 2023

Abs DOI Bib HTML

Simulation is crucial for all aspects of collider data analysis, but the available computing budget in the High Luminosity LHC era will be severely constrained. Generative machine learning models may act as surrogates to replace physics-based full simulation of particle detectors, and diffusion models have recently emerged as the state of the art for other generative tasks. We introduce CaloDiffusion, a denoising diffusion model trained on the public CaloChallenge datasets to generate calorimeter showers. Our algorithm employs 3D cylindrical convolutions, which take advantage of symmetries of the underlying data representation. To handle irregular detector geometries, we augment the diffusion model with a new geometry latent mapping (GLaM) layer to learn forward and reverse transformations to a regular geometry that is suitable for cylindrical convolutions. The showers generated by our approach are nearly indistinguishable from the full simulation, as measured by several different metrics.
@article{CaloDiffusion, topic = {generative}, author = {Amram, Oz and Pedro, Kevin}, title = {{Denoising diffusion models with geometry adaptation for high fidelity calorimeter simulation}}, eprint = {2308.03876}, archiveprefix = {arXiv}, eid = {arxiv:2308.03876}, primaryclass = {physics.ins-det}, reportnumber = {FERMILAB-PUB-23-384-CSAID-PPD}, doi = {10.1103/PhysRevD.108.072014}, journal = {Phys. Rev. D}, volume = {108}, number = {7}, pages = {072014}, year = {2023}, }

Physics: Foundation Models

arXiv

Neural Scaling Laws for Jet Generation

Oz Amram, Darius A. Faroughy, Tjarko Gerdes, and 5 more authors

May 2026

Abs HTML

Recently observed empirical scaling laws describe the performance of foundation-type models as three independent key quantities – dataset size, compute, and model parameters – are modified. Extracting these scaling laws informs the training of large complex models for which the tuning of hyperparameters in traditional ways is not feasible. This work for the first time explores if scaling laws can also be observed for the task of particle jet generation – both relevant as a pre-training objective for foundation models and as in-situ simulation by itself. We indeed replicate the key logarithmic scaling law behavior for model-size scaling. Beyond studying the next token prediction validation loss of the generative model, we also study the sliced Wasserstein distance of five physical quantities that are not immediately available to the model during training. Our study shows that this quantity is monotonically related to the next token prediction validation loss, meaning that this loss is indeed a good proxy for the physics performance. For the scaling with dataset size and compute, we observe substantially weaker scaling behavior of both the loss and the sliced Wasserstein distance. We analyze this behavior by introducing the concept of a learnable window, and argue that autoregressive next token prediction on jet constituents exhibits comparatively rapid saturation relative to language-model studies. We discuss possible origins of this behavior, including the stochastic nature of QCD radiation and differences between generative and supervised learning tasks in collider physics.
ML Sci. Tech.

Aspen Open Jets: Unlocking LHC Data for Foundation Models in Particle Physics

Oz Amram, Luca Anzalone, Joschka Birk, and 7 more authors

Mach. Learn. Sci. Tech., Jun 2025

Abs DOI HTML

Foundation models are deep learning models pre-trained on large amounts of data which are capable of generalizing to multiple datasets and/or downstream tasks. This work demonstrates how data collected by the CMS experiment at the Large Hadron Collider can be useful in pre-training foundation models for HEP. Specifically, we introduce the AspenOpenJets dataset, consisting of approximately 180M high pT jets derived from CMS 2016 Open Data. We show how pre-training the OmniJet-α foundation model on AspenOpenJets improves performance on generative tasks with significant domain shift: generating boosted top and QCD jets from the simulated JetClass dataset. In addition to demonstrating the power of pre-training of a jet-based foundation model on actual proton-proton collision data, we provide the ML-ready derived AspenOpenJets dataset for further public use.

Physics: Other

Contribution to ESPP

United States Early Career Researchers in Collider Physics input to the European Strategy for Particle Physics Update

Oz Amram and Grace Cummings

In , Mar 2025

Abs HTML

This document represents a contribution of the United States early career collider physics community to the 2025–2026 update to the European Strategy for Particle Physics. Preferences with regard to different future collider options and R&D priorities were assessed via a survey. The early career community was defined as anyone who is a graduate student, postdoctoral researcher, untenured faculty member, or research scientist under 40 years of age. In total, 105 participants responded to the survey between February and March 10th, 2025. Questions were formulated primarily to gauge the enthusiasm and preferences for different collider options in line with the recommendations of the United States’ P5 report, relevant to the European Strategy Update.
JHEP

Search for t-channel scalar and vector leptoquark exchange in the high-mass dimuon and dielectron spectra in proton-proton collisions at √s = 13 TeV

CMS Collaboration

JHEP, 2024

Abs DOI HTML

A search for t-channel exchange of leptoquarks (LQs) is performed in dimuon and dielectron spectra using proton-proton collision data collected at √s = 13 TeV with the CMS detector at the CERN LHC. The data correspond to an integrated luminosity of 138 fb−1. Eight scenarios are considered, in which scalar or vector LQs couple up or down quarks to muons or electrons, for dilepton invariant masses above 500 GeV. The LQ masses are probed up to 5 TeV, beyond a regime probed by previous pair-production and single-production searches. The differential distributions of dilepton events are fit to templates that model the nonresonant LQ exchange and various standard model background processes. Limits are set on LQ-fermion coupling strengths for scalar and vector LQ masses in the 1-5 TeV range at 95% confidence level, establishing stringent limits on first- and second-generation LQs.
JHEP
Measurement of the Drell-Yan forward-backward asymmetry at high dilepton masses in proton-proton collisions at √s = 13 TeV

CMS Collaboration

JHEP, 2022

Abs DOI Bib HTML

A measurement of the forward-backward asymmetry of pairs of oppositely charged leptons (dimuons and dielectrons) produced by the Drell-Yan process in proton-proton collisions is presented. The data sample corresponds to an integrated luminosity of 138 fb−1 collected with the CMS detector at the LHC at a center-of-mass energy of 13 TeV. The asymmetry is measured as a function of lepton pair mass for masses larger than 170 GeV and compared with standard model predictions. An inclusive measurement across both channels and the full mass range yields an asymmetry of 0.612 ± 0.005 (stat) ± 0.007 (syst). As a test of lepton flavor universality, the difference between the dimuon and dielectron asymmetries is measured as well. No statistically significant deviations from standard model predictions are observed. The measurements are used to set limits on the presence of additional gauge bosons. For a Z’ boson in the sequential standard model the observed (expected) 95% confidence level lower limit on the Z’ mass is 4.4 (3.7) TeV.
@article{DY_AFB, topic = {other}, author = {{CMS Collaboration}}, collaboration = {CMS}, title = {{Measurement of the Drell-Yan forward-backward asymmetry at high dilepton masses in proton-proton collisions at √s = 13 TeV}}, eprint = {2202.12327}, archiveprefix = {arXiv}, eid = {arxiv:2202.12327}, primaryclass = {hep-ex}, reportnumber = {CMS-SMP-21-002, CERN-EP-2022-013}, doi = {10.1007/JHEP08(2022)063}, journal = {JHEP}, volume = {2022}, number = {08}, pages = {063}, year = {2022}, }
ApJ

Merger or Not: Accounting for Human Biases in Identifying Galactic Merger Signatures

Erini Lambrides and Others

The Astrophysical Journal, Sep 2021

Abs DOI HTML

Significant galaxy mergers throughout cosmic time play a fundamental role in theories of galaxy evolution. The widespread usage of human classifiers to visually assess whether galaxies are in merging systems remains a fundamental component of many morphology studies. Studies that employ human classifiers usually construct a control sample, and rely on the assumption that the bias introduced by using humans will be evenly applied to all samples. In this work, we test this assumption and develop methods to correct for it. Using the standard binomial statistical methods employed in many morphology studies, we find that the merger fraction, error, and the significance of the difference between two samples are dependent on the intrinsic merger fraction of any given sample. We propose a method of quantifying merger biases of individual human classifiers and incorporate these biases into a full probabilistic model to determine the merger fraction and the probability of an individual galaxy being in a merger. Using 14 simulated human responses and accuracies, we are able to correctly label a galaxy as ”merger” or ”isolated” to within 1% of the truth. Using 14 real human responses on a set of realistic mock galaxy simulation snapshots our model is able to recover the pre-coalesced merger fraction to within 10%. Our method can not only increase the accuracy of studies probing the merger state of galaxies at cosmic noon, but also can be used to construct more accurate training sets in machine learning studies that use human classified data-sets.