arXiv subjects

Inar Timiryasov

Publications and source records attributed to Inar Timiryasov.

At least 19 recordsLinked to original sources

arXiv2026

Finding and using interpretable latents in a neutrino foundation model with sparse autoencoders

We present a first application of sparse-autoencoder-based mechanistic interpretability to particle physics. Studying a neutrino foundation model pretrained on IceCube data and fine-tuned for direction reconstruction, we identify a validated atlas of physical concepts in the model representation, using a strict validation protocol consisting of held-out tests, matched nuisance controls, and replication across independent dictionary trainings. Causal interventions show that the direction head barely draws on this atlas. Motivated by this underused information, we train an uncertainty head on the same event-level representation to predict the model's angular reconstruction error. Unlike the direction head, it depends causally on quality and brightness features from the atlas. At $20\%$ selection efficiency, this interpretable estimator improves the median angular resolution from $20.2^\circ$ to $3.2^\circ$. These results suggest that mechanistic interpretability can reveal learned latent physics encoded within a model's internal representation and help design downstream tasks that exploit it.

astro-ph.HE↗

arXiv2026

Learning Standard Model structure from LHC data with Riemannian flow matching

In this work we demonstrate that a single transformer-based generative model can capture Standard Model structure spanning five decades of invariant mass, from the sub-GeV regime to the TeV continuum, a range that no single Monte Carlo sample covers. To achieve this we design \textsc{ShellFlow}, a Riemannian conditional flow matching model that, given the recorded event composition, generates each particle on its on-shell manifold. Its only physics priors are the on-shell condition and the invariant-mass formula. The model is trained on $\sim 10^{9}$ real $pp$ collision events from the ATLAS Open Data 13~TeV release and told nothing else. From a single training run, the model learns to reproduce all of the following: intra-particle kinematics, the dilepton resonances ($J/\psi$, $\Upsilon$, $Z$) at their PDG positions, the leptonic Weinberg angle, the $W$ and top-quark masses, and inter-particle correlations that enter no training objective. A substantial fraction of the Standard Model is thus learnable directly from recorded collision data.

hep-ph↗

arXiv2025

Learning to detect continuous gravitational waves: an open data-analysis competition

We report results of a public data-analysis challenge, hosted on the open data-science platform Kaggle, to detect simulated continuous gravitational-wave signals (CWs). These are weak signals from rapidly spinning neutron stars that remain undetected despite extensive searches. The competition dataset consisted of a population of CW signals using both simulated and real LIGO detector data matching the conditions of actual CW searches. The competition attracted more than 1,000 participants to develop realistic CW search algorithms. We describe the top 10 approaches and discuss their applicability as a pre-processing step compared to standard CW-search approaches. For the competition's dataset, we find that top approaches can reduce the computing cost by 1 to 3 orders of magnitude at a false-dismissal probability comparable to standard CW searches. Additionally, the competition drove the development of new GPU-accelerated detection pipelines, which facilitated their adoption in other areas of gravitational-wave data analysis. We release the associated dataset, which constitutes the first open standardized benchmark for CW detection, to enable reproducible method comparisons and to encourage further developments toward the first detection of these elusive signals.

gr-qc↗

arXiv2024

BabyLlama-2: Ensemble-Distilled Models Consistently Outperform Teachers With Limited Data

We present BabyLlama-2, a 345 million parameter model distillation-pretrained from two teachers on a 10 million word corpus for the BabyLM competition. On BLiMP and SuperGLUE benchmarks, BabyLlama-2 outperforms baselines trained on both 10 and 100 million word datasets with the same data mix, as well as its teacher models. Through an extensive hyperparameter sweep, we demonstrate that the advantages of distillation cannot be attributed to suboptimal hyperparameter selection of the teachers. Our findings underscore the need for further investigation into distillation techniques, particularly in data-limited settings.

cs.CL↗

arXiv2024

Tree-level unitarity constraints on heavy neutral leptons

Heavy neutral leptons (HNLs) can explain the origin of neutrino masses and oscillations over a wide range of masses. Direct experimental probes of HNLs become unfeasible for masses significantly above the electroweak scale.Consequently, the strongest limits arise from the non-observation of charged lepton flavor-violating processes induced by HNLs at loop level.Counter-intuitively, these bounds tighten as the HNL mass increases, an effect that persists within the perturbative regime. This work explores the precise form of these bounds for HNLs with masses well beyond the electroweak scale by analyzing the full matrix of partial waves (tree-level unitarity). At high energies, the HNL model simplifies to a Yukawa theory, allowing unitarity constraints to be expressed in terms of the total Yukawa coupling $\left|Y_{\mathrm{tot}}\right|^2$ involving HNLs, lepton doublets, and the Higgs boson. Processes with $J=0$ and $J=1/2$ yield the well-known result $\left|Y_{\mathrm{tot}}\right|^2 \leq 8\pi$. However, the most stringent result arises from processes with $J = 1$, which is given by $\left|Y_{\mathrm{tot}}\right|^2 \leq 4\pi(\sqrt{5} -1) = 8\pi/\varphi \approx 15.533$, where $\varphi$ is the Golden ratio. These results remain valid provided that the Yukawa matrix has rank 1, a condition approximately satisfied in models with two or three HNLs, with large mixing angles, and radiactively small neutrino masses. Finally, we determine the maximum mass that an HNL can have in the type-I seesaw model while remaining the sole source of neutrino masses.

hep-ph↗

arXiv2023

Baby Llama: knowledge distillation from an ensemble of teachers trained on a small dataset with no performance penalty

We present our submission to the BabyLM challenge, whose goal was to improve the sample efficiency of language models. We trained an ensemble consisting of a GPT-2 and small LLaMA models on the developmentally-plausible, 10M-word BabyLM dataset, then distilled it into a small, 58M-parameter LLaMA model, which exceeds in performance both of its teachers as well as a similar model trained without distillation. This suggests that distillation can not only retain the full performance of the teacher model when the latter is trained on a sufficiently small dataset; it can exceed it, and lead to significantly better performance than direct training.

cs.CL↗

arXiv2023

Critical Points in Palatini Higgs Inflation with Small Non-Minimal Coupling

We investigate inflation driven by the Higgs boson in the Palatini formulation of General Relativity. Our analysis primarily focuses on a small non-minimal coupling of the Higgs field to gravity in the range $0<\xi\lesssim 1$. We incorporate the renormalization group running of the relevant parameters as computed within the Standard Model and allow for small corrections. In addition to $\xi$, our model features two tunable parameters: the low-energy value of the top Yukawa coupling and an effective jump of the Higgs self-interaction. Our results indicate that critical points leading to a large enhancement of the power spectrum can be produced. However, the observed amplitude of perturbations in the CMB cannot be matched within this setting. On the one hand, this makes it difficult to generate a sizable abundance of primordial black holes. On the other hand, our finding can be viewed as further evidence that Palatini Higgs inflation has favourable high-energy properties due to robustness against quantum corrections.

hep-ph↗

arXiv2022

Heavy Neutral Leptons -- Advancing into the PeV domain

Heavy neutral leptons (HNLs) are hypothetical particles able to explain neutrino oscillations and provide a mechanism for generating the baryon asymmetry of the Universe. Quantum corrections due to such particles give rise to flavor violating processes in the charged lepton sector. Based on the fact that these corrections grow with HNL masses, we improve existing constraints by orders of magnitude in mass and mixing angle. This allows us to probe part of the parameter space of leptogenesis with multi-TeV HNLs. We also show that one will be able to infer HNL parameters in a significant portion of the parameter space for TeV-PeV masses if charged lepton flavor violating signals are detected.

hep-ph↗

arXiv2022

Preheating in Palatini Higgs inflation on the lattice

We study preheating following Higgs inflation in the Palatini formulation of gravity. We numerically evolve perturbations of the radial mode of the Higgs field and that of three scalars modeling the gauge bosons. We compare the two non-perturbative mechanisms of growth of excitations -- parametric resonance and tachyonic instability -- and confirm that the latter plays the dominant role. Our results provide further evidence that preheating in Palatini Higgs inflation happens within a single oscillation of the Higgs field about the bottom of its potential, consistent with the approximation of an instantaneous preheating.

hep-ph↗

arXiv2022

The Present and Future Status of Heavy Neutral Leptons

The existence of non-zero neutrino masses points to the likely existence of multiple SM neutral fermions. When such states are heavy enough that they cannot be produced in oscillations, they are referred to as Heavy Neutral Leptons (HNLs). In this white paper we discuss the present experimental status of HNLs including colliders, beta decay, accelerators, as well as astrophysical and cosmological impacts. We discuss the importance of continuing to search for HNLs, and its potential impact on our understanding on key fundamental questions, and additionally we outline the future prospects for next-generation future experiments or upcoming accelerator run scenarios.

hep-ph↗

arXiv2021

Why interpretation matters for BSM searches: a case study with Heavy Neutral Leptons at ATLAS

Experiments searching for Heavy Neutral Leptons (HNLs) typically interpret their results within simplified models consisting of a single HNL coupled to a single lepton flavor. However, any model which aims to describe neutrino oscillations necessarily features more than one HNL, coupled to several flavors. As we show in this work, the reinterpretation of the results of experimental searches in terms of realistic models is a non-trivial task. We perform a detailed reinterpretation of the latest ATLAS search for prompt HNLs in W decays within a minimal low-scale seesaw with two HNLs. We show that the exclusion limits obtained using the detailed reinterpretation can differ by several orders of magnitude from the limits quoted for the simplified models. Hence naively comparing the mixing angles from a realistic model to the reported limits could lead to wrongly excluding entire regions of parameter space! To overcome this issue without requiring experiments to report constraints on all possible HNL models, we propose a simple framework that allows one to easily and accurately reinterpret exclusion limits within closely-related models. We outline a number of concrete steps that can be taken by experiments to implement this method with minimal effort, and we discuss its applicability to other models of feebly interacting particles.

hep-ph↗

arXiv2021

Reinterpreting the ATLAS bounds on heavy neutral leptons in a realistic neutrino oscillation model

Heavy neutral leptons (HNLs) are hypothetical particles, motivated in the first place by their ability to explain neutrino oscillations. Experimental searches for HNLs are typically conducted under the assumption of a single HNL mixing with a single neutrino flavor. However, the resulting exclusion limits may not directly constrain the corresponding mixing angles in realistic HNL models -- those which can explain neutrino oscillations. The reinterpretation of the results of these experimental searches turns out to be a non-trivial task, that requires significant knowledge of the details of the experiment. In this work, we perform a reinterpretation of the latest ATLAS search for HNLs decaying promptly to a tri-lepton final state. We show that in a realistic model with two HNLs, the actual limits can vary by several orders of magnitude depending on the free parameters of the model. Marginalizing over the unknown model parameters leads to an exclusion limit on the total mixing angle which can be up to 3 orders of magnitude weaker than the limits reported in arXiv:1905.09787. This demonstrates that the reinterpretation of results from experimental searches is a necessary step to obtain meaningful limits on realistic models. We detail a few steps that can be taken by experimental collaborations in order to simplify the reuse of their results.

hep-ph↗

arXiv2021

Unleashing the full power of LHCb to probe Stealth New Physics

In this paper, we describe the potential of the LHCb experiment to detect Stealth physics. This refers to dynamics beyond the Standard Model that would elude searches that focus on energetic objects or precision measurements of known processes. Stealth signatures include long-lived particles and light resonances that are produced very rarely or together with overwhelming backgrounds. We will discuss why LHCb is equipped to discover this kind of physics at the Large Hadron Collider and provide examples of well-motivated theoretical models that can be probed with great detail at the experiment.

hep-ph↗

arXiv2021

Reconciling resonant leptogenesis and baryogenesis via neutrino oscillations

Right-handed neutrinos offer an elegant solution to two well established phenomena beyond the Standard Model (SM) - masses and oscillations of neutrinos, as well as the baryon asymmetry of the Universe. It is also a minimalistic solution since it requires only singlet Majorana fermions to be added to the SM particle content. If these fermions are nearly degenerate, the mass scale of right-handed neutrinos can be very low and accessible by the present and planned experiments. There are at least two well studied mechanisms of the low-scale leptogenesis: baryogenesis via oscillations and resonant leptogenesis. These two mechanisms were often considered separate, but they can in fact be understood as two different regimes of one and the same mechanism, described by a unique set of quantum kinetic equations. In this work we show, using a unified description based on quantum kinetic equations, that the parameter space of these two regimes of low-scale leptogenesis significantly overlap. We present a comprehensive study of the parameter space of the low-scale leptogenesis with the mass scale ranging from $0.1$ GeV to $\sim 10^6$ GeV. The unified perspective of this work reveals the synergy between intensity and energy frontiers in the quest for heavy Majorana neutrinos.

hep-ph↗

arXiv2021

An allowed window for heavy neutral leptons below the kaon mass

The extension of the Standard Model with two gauge-singlet Majorana fermions can simultaneously explain two beyond-the-Standard-model phenomena: neutrino masses and oscillations, as well as the origin of the matter-antimatter asymmetry in the Universe. The parameters of such a model are constrained by the neutrino oscillation data, direct accelerator searches, big bang nucleosynthesis, and requirement of successful baryogenesis. We show that the combination of all these constraints still leaves an allowed region in the parameter space below the kaon mass. This region can be probed by the further searches of NA62, DUNE, or SHiP experiments.

hep-ph↗

arXiv2020

Freeze-in and freeze-out generation of lepton asymmetries after baryogenesis in the $\nu$MSM

The $\nu$MSM -- an extension of the Standard Model by three relatively light singlet Majorana fermions $N_{1,2,3}$ -- allows for the generation of lepton asymmetry which is several orders of magnitude larger than the observed baryon asymmetry of the Universe. The lepton asymmetry is produced in interactions of $N_{2,3}$ (with masses in the GeV region) at temperatures below the sphaleron freeze out $ T \lesssim 130$ GeV and can enhance the cosmological production of dark matter (DM) sterile neutrinos $N_1$ (with the mass of the keV scale) happening at $T \sim 200$ MeV due to active-sterile neutrino mixing. This asymmetry can be generated in freeze-in, freeze-out, or later in decays of heavy neutral leptons. In this work, we address the question of the magnitude of the late-time asymmetry (LTA) generated by the heavy neutral leptons $N_{2,3}$ during their freeze-in and freeze-out, leaving the decays for later work. We study how much of this asymmetry can survive down to the lower temperatures relevant for the sterile neutrino DM creation. We find that this LTA could result in the production of a sizeable fraction of dark matter. We also examine a role played by magnetic fields and the Abelian chiral anomaly in the generation of LTA, not accounted for in the previous studies. We argue that the production of LTA can be increased significantly and make an estimate of the influence of this effect.

hep-ph↗

arXiv2020

Uniting low-scale leptogeneses

In this work we demonstrate that what was previously considered as different mechanisms of baryon asymmetry generation involving two right-handed Majorana neutrinos with masses far below the GUT scale -- leptogenesis via neutrino oscillations and resonant leptogenesis -- are actually united. We show that the observed baryon asymmetry can be generated for all experimentally allowed values of the right-handed neutrino masses above $M_N \gtrsim 100$ MeV. Leptogenesis is effective in a broad range of the parameters, including mass splitting between two right-handed neutrinos as big as $\Delta M_N/M_N \sim 0.1$, as well as mixing angles between the heavy and light neutrinos large enough to be accessible to planned intensity experiments or future colliders.

hep-ph↗

arXiv2020

Projected NA62 sensitivity to heavy neutral lepton production in $K^+ \to \pi^0 e^+ N$ decays

Heavy neutral leptons (HNLs) appear in many extensions of the Standard Model of particle physics. In this study, we investigate to which extent the NA62 experiment at CERN could improve the existing bounds on the HNL mixing angle $|U_e|^2$ by performing a missing mass search in $K^+ \to \pi^0 e^+ N$ decays in flight. We show that the limit $|U_e|^2 \simeq 2\times 10^{-6}$ can be reached with the currently available data in the mass range 125 -- 144 MeV, which is currently not well covered by production searches. Future data, together with a dedicated trigger and/or improvements in rejection of out-of-acceptance photons, can improve this limit by another order of magnitude.

hep-ph↗