@unpublished{21858,
  abstract     = {The recent surge in high-quality open-source Generative AI text models (colloquially: LLMs), as well as efficient finetuning techniques, have opened the possibility of creating high-quality personalized models that generate text attuned to a specific individual’s needs and are capable of credibly imitating their writing style by refining an open-source model with that person’s own data. The technology to create such models is accessible to private individuals, and training and running such models can be done cheaply on consumer-grade hardware. While these advancements are a huge gain for usability and privacy, this position paper argues that the practical feasibility of impersonating specific individuals also introduces novel safety risks. For instance, this technology enables the creation of phishing emails
or fraudulent social media accounts, based on small amounts of publicly available text, or by the individuals themselves to escape AI text detection. We further argue that these risks are complementary to—and distinct from—the much-discussed risks of other impersonation attacks such as image, voice, or video deepfakes, and are not adequately addressed by the larger research community, or the current generation of open- and closed-source models.},
  author       = {Iofinova, Eugenia B and Jovanovic, Andrej and Alistarh, Dan-Adrian},
  booktitle    = {arXiv},
  title        = {{Position: It's time to act on the risk of efficient personalized text generation}},
  doi          = {10.48550/arXiv.2502.06560},
  year         = {2025},
}

@article{21144,
  abstract     = {This paper deals with the algorithmic aspects of solving feasibility problems of semidefinite programming (SDP), aka linear matrix inequalities (LMIs). Since in some SDP instances all feasible solutions have irrational entries, numerical solvers that work with rational numbers can only find an approximate solution. We study the following question: Is it possible to certify feasibility of a given SDP using an approximate solution that is sufficiently close to some exact solution? Existing approaches make the assumption that there exist rational feasible solutions (and use techniques such as rounding and lattice reduction algorithms). We propose an alternative approach that does not need this assumption. More specifically, we show how to construct a system of polynomial equations whose set of real solutions is guaranteed to have an isolated correct solution (assuming that the target exact solution is maximum-rank). This allows, in particular, for us to use algorithms from real algebraic geometry for solving systems of polynomial equations, yielding a hybrid (or symbolic-numerical) method for SDPs. We experimentally compare it with a pure symbolic method in [D. Henrion, S. Naldi, and M. Safey El Din, SIAM J. Optim., 26 (2016), pp. 2512–2539]; the hybrid method was able to certify feasibility of many SDP instances on which the aforementioned paper failed. Our approach may have further applications, such as refining an approximate solution using methods of numerical algebraic geometry for systems of polynomial equations.},
  author       = {Kolmogorov, Vladimir and Naldi, Simone and Zapata, Jeferson},
  issn         = {1095-7189},
  journal      = {SIAM Journal on Optimization},
  number       = {3},
  pages        = {1630--1654},
  publisher    = {Society for Industrial and Applied Mathematics},
  title        = {{Certifying solutions of degenerate semidefinite programs}},
  doi          = {10.1137/24m1664691},
  volume       = {35},
  year         = {2025},
}

@phdthesis{20735,
  abstract     = {Left–right alternation is a defining feature of spinal locomotor circuits, yet the level of neuronal
detail required to generate and maintain this pattern remains unclear. This thesis investigates how
models spanning multiple levels of abstraction—from biophysically detailed Hodgkin–Huxley (HH)
neurons to adaptive integrate–and–fire (I&F) formulations and synfire-chain modules—can account
for the generation of fictive swimming in the spinal cord of the Xenopus laevis tadpole. The guiding
hypothesis is that a small set of neuronal mechanisms is sufficient to reproduce the essential features
of rhythmic alternation, and that moving between modeling scales helps distinguish core principles
from biological detail.
A minimal bilateral HH network comprising only four canonical neuron classes—excitatory
descending interneurons (dINs), inhibitory commissural interneurons (cINs), ipsilateral inhibitory
interneurons (aINs) and motoneurons—served as a biophysical proof of concept. Tuned to reproduce
experimentally observed firing modes, the model demonstrated that rebound-prone dIN excitability,
contralateral inhibition and modest electrical coupling are sufficient to generate stable alternating
activity, even in very small networks. These results motivated the transition to simpler models
capable of efficient analysis and scaling.
Adaptive exponential I&F (AdEx) neurons were calibrated to physiological recordings using
simulation-based inference, yielding tonic and phasic/rebound templates that preserved the key
dynamical signatures of the HH model. Phase-plane analysis clarified the mechanisms underlying
single-spike responses and rebound firing in dINs. At network level, the I&F models robustly
reproduced left–right alternation, while highlighting constraints on synaptic kinetics and adaptation
needed to avoid multi-spike responses.
Finally, a synfire-chain framework provided a complementary, timing-centric perspective, demonstrating how precise spike synchrony, synaptic delays and minimal inhibitory coupling can generate
alternating left–right sequences in a feedforward setting. Together, these approaches converge on a
common conclusion: rebound-prone ipsilateral excitation combined with precisely timed contralateral inhibition constitutes a sufficient substrate for alternating spinal rhythms.
By integrating bottom-up and top-down modeling strategies, this thesis provides a unified, extensible framework for studying spinal pattern generation. The results show that essential locomotor
dynamics can be captured across multiple abstraction levels, offering both mechanistic insight and
practical tools for future data-driven investigations of spinal circuit development, robustness and
modulation.},
  author       = {Wilson, Alexia C},
  issn         = {2791-4585},
  pages        = {110},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{Modelling the spinal cord of a tadpole: Exploring different ways to model the spinal cord in the Xenopus frog}},
  doi          = {10.15479/AT-ISTA-20735},
  year         = {2025},
}

@unpublished{21398,
  abstract     = {Seymour's decomposition theorem is a hallmark result in matroid theory presenting a structural characterization of the class of regular matroids. Formalization of matroid theory faces many challenges, most importantly that only a limited number of notions and results have been implemented so far. In this work, we formalize the proof of the forward (composition) direction of Seymour's theorem for regular matroids. To this end, we develop a library in Lean 4 that implements definitions and results about totally unimodular matrices, vector matroids, their standard representations, regular matroids, and 1-, 2-, and 3-sums of matrices and binary matroids given by their standard representations. Using this framework, we formally state Seymour's decomposition theorem and implement a formally verified proof of the composition direction in the setting where the matroids have finite rank and may have infinite ground sets.},
  author       = {Dvorak, Martin and Figueroa-Reid, Tristan and Hamadani, Rida and Hwang, Byung-Hak and Karunus, Evgenia and Kolmogorov, Vladimir and Meiburg, Alexander and Nelson, Alexander and Nelson, Peter and Sandey, Mark and Sergeev, Ivan},
  booktitle    = {arXiv},
  title        = {{Composition direction of Seymour's theorem for regular matroids — Formally verified}},
  doi          = {10.48550/arXiv.2509.20539},
  year         = {2025},
}

@phdthesis{20371,
  abstract     = {Quantum mechanics reveals a world that defies classical determinism, where uncertainty, superposition, and fluctuations are fundamental aspects. Engineering devices that harness these quantum features requires not only precision, but also a deep understanding of how they interact with their surrounding environment. Superconducting circuits, which exploit
macroscopic quantum coherence in low-loss superconducting materials, provide a scalable platform for implementing such systems. Among the critical elements in these circuits, superinductors—high-impedance, dissipation-free inductive components—play a central role by suppressing charge fluctuations. They allow quantum states to be delocalized in phase space, protect qubits from environmental noise, and facilitate access to phenomena such as dual Josephson physics and ultra-strong coupling regimes. 
This thesis explores two complementary implementations of high-impedance circuits: geometric superinductors, demonstrating that high impedance can be achieved beyond kinetic inductance,
and Josephson junction chains, used to investigate both microwave mode properties and DC transport across the superconductor-to-insulator transition. 
Part I addresses geometric superinductors. Contrary to the common belief that high-impedance superconducting circuits require kinetic inductance, we demonstrate that purely geometric designs can achieve characteristic impedance exceeding the resistance quantum. By exploiting mutual coupling between adjacent turns, coil-based inductors achieve enhanced self-inductance, creating a reliable platform for qubits and resonators. Modeling, simulation, fabrication, and
characterization confirm that these elements behave as superinductor. With low loss, high linearity, and minimal stray capacitance, these elements are reproducible, free of uncontrolled tunneling events, and capable of strong magnetic coupling. This establishes geometric superinductors as robust, single-wave-function superconducting devices suitable for hardware protected qubits and hybrid systems.
Part II presents classical numerical simulations of a Quantum Phase Slip circuit to study dual Shapiro steps. The circuit consists of an ideal Quantum Phase Slip element embedded in a resistive-inductive environment with a parasitic capacitance.
Part III extends the investigation of high characteristic-impedance circuit elements to one-dimensional Josephson junction chains, which act as a quantum simulator for many-body physics and the superconductor–insulator transition. Different devices are realized on both sides of the DC phase transition, showing either a supercurrent branch or Coulomb blockade at zero bias. The effect of the crossover on microwave modes, however, remains insufficiently investigated. Studying these modes provides insight into the interplay between disorder and phase-slip events. Small differences in circuit component sizes determine which side of the transition a device falls on, making these results relevant not only for fundamental understanding but also for the design of quantum devices, emphasizing the crucial role of the
electromagnetic environment in stabilizing and controlling fragile quantum states. 
Together, these results illustrate how carefully engineered high characteristic-impedance elements provide a link between macroscopic circuits and the inherently uncertain quantum world, enabling experiments that probe, control, and ultimately exploit quantum fluctuations for applications in quantum information, metrology, solid state physics and beyond.

},
  author       = {Trioni, Andrea},
  isbn         = {978-3-99078-067-1},
  issn         = {2663-337X},
  pages        = {202},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{High-impedance quantum circuits for mesoscopic physics: Geometric superinductors and insulating Josephson Chains}},
  doi          = {10.15479/AT-ISTA-20371},
  year         = {2025},
}

@misc{20641,
  abstract     = {Protein conformational energy landscapes are shaped not only by intramolecular interactions but also by their environment. In protein crystals and protein-protein complexes, intermolecular contacts alter this energy landscape, but the exact nature of this alteration is difficult to decipher. Understanding how the crystal lattice affects protein dynamics is crucial for crystallography-based studies of motion, yet its influence on collective motions remains unclear. Aromatic ring flips in the hydrophobic core represent sensitive probes of such dynamics. Here, we compare the kinetics of aromatic ring flips in the protein GB1 in crystals, in complex with its binding partner IgG, and in solution, combining advanced isotope labeling with quantitative NMR methods. We show that rings in the core flip nearly a thousand times less frequently in crystals than in solution. Enhanced-sampling molecular dynamics simulations, based on a new crystal structure, reproduce these elevated barriers and reveal how the crystal restrains motions. },
  author       = {Becker, Lea Marie and Schanda, Paul},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{Data for "Aromatic Ring Flips Reveal Reshaping of Protein Dynamics in Crystals and Complexes"}},
  doi          = {10.15479/AT-ISTA-20641},
  year         = {2025},
}

@article{20926,
  abstract     = {Most current machine learning interatomic potentials (MLIPs) rely on short-range approximations, without explicit treatment of long-range electrostatics. To address this, we recently developed the Latent Ewald Summation (LES) method, which infers electrostatic interactions, polarization, and Born effective charges (BECs), just by learning from energy and force training data. Here, we present LES as a standalone library, compatible with any short-range MLIP, and demonstrate its integration with methods such as MACE, NequIP, Allegro, CACE, CHGNet, and UMA. We benchmark LES-enhanced models on distinct systems, including bulk water, polar dipeptides, and gold dimer adsorption on defective substrates, and show that LES not only captures correct electrostatics but also improves accuracy. Additionally, we scale LES to large and chemically diverse data by training MACELES-OFF on the SPICE set containing molecules and clusters, making a universal MLIP with electrostatics for organic systems, including biomolecules. MACELES-OFF is more accurate than its short-range counterpart (MACE-OFF) trained on the same data set, predicts dipoles and BECs reliably, and has better descriptions of bulk liquids. By enabling efficient long-range electrostatics without directly training on electrical properties, LES paves the way for electrostatic foundation MLIPs.},
  author       = {Kim, Dongjin and Wang, Xiaoyu and Vargas, Santiago and Zhong, Peichen and King, Daniel S. and Inizan, Theo Jaffrelot and Cheng, Bingqing},
  issn         = {1549-9626},
  journal      = {Journal of Chemical Theory and Computation},
  number       = {24},
  pages        = {12709--12724},
  publisher    = {American Chemical Society},
  title        = {{A universal augmentation framework for long-range electrostatics in machine learning interatomic potentials}},
  doi          = {10.1021/acs.jctc.5c01400},
  volume       = {21},
  year         = {2025},
}

@phdthesis{20074,
  abstract     = {Prenatal immune challenges pose significant risks to human embryonic brain and eye development. However, we still lack knowledge about the safe usage of anti-inflammatory drugs during pregnancy. Human induced pluripotent stem cell (hIPSC)-derived brain organoid models provide a unique opportunity to investigate neuronal development and have started to explore functional consequences upon viral infection. However, brain organoids usually lack microglia, the brain-resident immune cells. They are present in the early human embryonic brain and actively participate in neuronal circuit development. At the same time, microglia are known for their immune-sensing properties and will influence viral-mediated effects. In my thesis, I was interested to study the multifunctional role of human microglia during retinal development. 
In chapter 1, I characterize the innate occurrence of IBA1+-microglia-like cells within the retinal organoid differentiation (Bartalska et al., 2022). Therefore, we differentiate hIPSC using an unguided retinal organoid differentiation protocol and observe the presence of IBA1+-microglia-like cells alongside retinal cups between week 3 and 4 in 2.5D culture. However, instead of infiltrating the neuroectodermal sides, they enrich within non-pigmented, 3D-cystic compartments that develop in low numbers parallel to 3D-retinal organoids. To enrich for IBA1+-microglia precursors (preMG), we guided the differentiation with a low-dosed BMP4 application, which prevents retinal cup development and enhances microglia and 3D-cysts formation. We characterize the differentiated preMG for their microglia-like identity and validated their functionality. In parallel, mass spectrometry identifies the 3D-cysts to express mesenchymal and epithelial markers. We confirm that comparable 3D-cysts are also the preferential environment for IBA1+-microglia-like cells within the unguided retinal organoid differentiation. 
In chapter 2, I investigate how microglia influence retinal development and whether they contribute to viral-mediated consequences (Schmied et al., 2025). Here, we assemble preMG, which we have characterized in chapter 1, into 3D-retinal organoids. Once the outer plexiform layer forms, microglia-like cells (iMG) populate them and interact with retinal cell types. However, at this developmental stage, the ganglion cell number decreases in 3D-retinal organoids. Thus, we adapted the model into 2D which promotes their survival. Integrated iMG engulf ganglion cells and control their cell number. In parallel, we apply the immunostimulant POLY(I:C) to mimic a fetal viral infection. Although POLY(I:C) stimulation affects iMG phenotype, it does not influence their interaction with ganglion cells. Furthermore, iMG presence significantly contributes to the supernatant’s inflammatory secretome and increases retinal cell proliferation. Simultaneous exposure to the non-steroidal anti-inflammatory drug (NSAID) ibuprofen dampens POLY(I:C)-mediated consequences of the iMG phenotype and ameliorates cell proliferation. Remarkably, while POLY(I:C) disrupts neuronal calcium dynamics independent of iMG presence, ibuprofen rescues this effect only in the presence of iMG. Mechanistically, ibuprofen blocks the enzymes cyclooxygenase 1 and 2 (COX1/ PTGS1 and COX2/ PTGS2) simultaneously, from which iMG predominantly express COX1. Selective inhibition of COX1 does not restore the calcium peak amplitude upon POLY(I:C) stimulation, indicating ibuprofen’s effect depends on the presence and interplay of both, COX1 and COX2. 
In summary, we characterized the 3D-retinal organoid model for the occurrence of IBA1+-microglia like cells. As the innately developing IBA1+-cells enrich in mesenchymal over retinal structures, we optimized a protocol to differentiate IBA1+-microglia precursors. By combining these two models we generate microglia-assembled retinal organoids. Our results underscore the importance of microglia during neurodevelopment, in the context of prenatal immune challenges and provide insight into the mechanisms by which ibuprofen exerts its protective effects during embryonic development.},
  author       = {Hübschmann, Verena},
  isbn         = {978-3-99078-060-2},
  issn         = {2663-337X},
  pages        = {151},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{Human microglia impact neuronal development in retinal organoids}},
  doi          = {10.15479/AT-ISTA-20074},
  year         = {2025},
}

@article{20342,
  abstract     = {Safety and liveness stand as fundamental concepts in formal languages, playing a key role in verification. The safety-liveness classification of boolean properties characterizes whether a given property can be falsified by observing a finite prefix of an infinite computation trace (always for safety, never for liveness). In the quantitative setting, properties are arbitrary functions from infinite words to partially-ordered domains. Extending this paradigm to the quantitative domain, where properties are arbitrary functions mapping infinite words to partially-ordered domains, we introduce and study the notions of quantitative safety and liveness. First, we formally define quantitative safety and liveness, and prove that our definitions induce conservative quantitative generalizations of both the safety-progress hierarchy and the safety-liveness decomposition of boolean properties. Consequently, like their boolean counterparts, quantitative properties can be min-decomposed into safety and liveness parts, or alternatively, max-decomposed into co-safety and co-liveness parts. We further establish a connection between quantitative safety and topological continuity and provide alternative characterizations of quantitative safety and liveness in terms of their boolean analogs. Second, we instantiate our framework with the specific classes of quantitative properties expressed by automata. These quantitative automata contain finitely many states and rational-valued transition weights, and their common value functions Inf, Sup, LimInf, LimSup, LimInfAvg, LimSupAvg, and DSum map infinite words into the totally-ordered domain of real numbers. For all common value functions, we provide a procedure for deciding whether a given automaton is safe or live, we show how to construct its safety closure, and we present a min-decomposition into safe and live automata.},
  author       = {Boker, Udi and Henzinger, Thomas A and Mazzocchi, Nicolas Adrien and Sarac, Naci E},
  issn         = {1860-5974},
  journal      = {Logical Methods in Computer Science},
  number       = {2},
  publisher    = {EPI Sciences},
  title        = {{Safety and liveness of quantitative properties and automata}},
  doi          = {10.46298/lmcs-21(2:2)2025},
  volume       = {21},
  year         = {2025},
}

@phdthesis{20167,
  author       = {Schön, Hanna},
  isbn         = {978-3-99078-061-9},
  issn         = {2663-337X},
  pages        = {171},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{The ER complex SUTU-7/MACO-1 regulates the fate of mRNAs encoding GPCRs}},
  doi          = {10.15479/AT-ISTA-20167},
  year         = {2025},
}

@article{20538,
  abstract     = {In this study, we describe an integrated approach for methyl group assignment comprising precursor-based selective methyl group labeling, a novel pulse sequence for methyl to backbone coherence transfer and chemical shift predictions using UCBShift 2.0. The utility of this novel α-ketoacid isotopologue is shown by the adaptation of an HMBC-HMQC pulse sequence that simultaneously connects geminal methyl groups of leucine and valine residues to each other and to the protein backbone. By additional 13C,2H-labeling of residues other than valine and leucine residues of the protein, important chemical shift information about neighboring residues (following valine and leucine residues) can be achieved. Thus, different valine and leucine residues in a protein can be characterized as a specific chemical shift vector. Frequency matching with predicted chemical shifts via UCBShift 2.0 using experimental data taken from a subset of the BMRB database revealed a correct assignment performance of about 90%. With applications to proteins of 60.2 kDa and 134 kDa (4 × 33.5 kDa) in size, we demonstrate that the approach provides valuable information even for very large proteins.},
  author       = {Knödlstorfer, Sonja and Toscano, Giorgia and Ptaszek, Aleksandra L. and Kontaxis, Georg and Napoli, Federico and Schneider, Jakob and Maier, Katharina and Kapitonova, Anna and Lichtenecker, Roman J. and Schanda, Paul and Konrat, Robert},
  issn         = {1089-8638},
  journal      = {Journal of Molecular Biology},
  number       = {23},
  publisher    = {Elsevier},
  title        = {{A novel HMBC-CC-HMQC NMR strategy for methyl assignment using triple-13C-labeled α-ketoisovalerate integrated with UCBShift 2.0}},
  doi          = {10.1016/j.jmb.2025.169465},
  volume       = {437},
  year         = {2025},
}

@unpublished{19398,
  abstract     = {Receptor-like kinases (RLKs), particularly the Transmembrane Kinase (TMK) family, play essential roles in signaling and development, with TMKs being key components of auxin perception and downstream phosphorylation events. While TMKs’ involvement in auxin canalization, a process essential for vasculature formation and regeneration, has been established, nonetheless, the additional signaling and regulatory partners remain poorly understood. In this study, we identify and characterize seven leucine-rich repeat RLKs (TINT1–TINT7) as novel interactors of TMK1, revealing their diverse evolutionary, structural, and functional characteristics. Our results show that TINTs interact with TMK1 and highlight their roles in regulating various developmental processes. Majority of TINTs contributes, together with TMK1, to auxin canalization, with TINT5 linking TMK1 to other canalization component CAMEL. Beyond canalization, we also establish the role of TINT-TMK1 interactions in processes such as stomatal movement and the hypocotyl’s gravitropic response. These findings suggest that TINTs, through their interaction with TMK1, are integral components of various signaling networks, contributing to both auxin canalization and broader plant development.},
  author       = {Monzer, Aline and Mazur, Ewa and Rodriguez Solovey, Lesia and Gallei, Michelle C and Zou, Minxia and Smejkal, Michael and Cervenova, Ema and Friml, Jiří},
  booktitle    = {bioRxiv},
  title        = {{TMK interacting network of receptor like kinases for auxin canalization and beyond}},
  doi          = {10.1101/2025.02.28.640727},
  year         = {2025},
}

@phdthesis{19684,
  abstract     = {The overarching goal of this thesis is to break down the complexity of turbulent flows in terms of enumerable, coherent structures and patterns. In a five-paper series, we adopt a variety of perspectives and techniques to relate the properties of systems of increasing complexity to their underlying coherent structures. 

Initially, we take a dynamical systems point of view, seeing turbulent flow as a chaotic trajectory bouncing between exact unstable solutions of the underlying equations of motion. Using persistent homology, the main tool of topological data analysis capturing the persistence across scales of topological features in a point cloud, we introduce a method that quantifies visits of turbulent trajectories to unstable time-periodic solutions, also called periodic orbits. We demonstrate this method first in the Rössler and Kuramoto–Sivashinsky systems. Using this method in 3D Kolmogorov flow, we extract a Markov chain from turbulent data, where each node corresponds to the neighbourhood of a periodic orbit. The invariant distribution of this Markov chain reproduces expectation values on turbulent data when it is used to weight averages on the respective periodic orbits.

In more realistic, wall-bounded settings, such as plane-Couette flow (pcf) driven by the relative motion of the walls, or plane-Poiseuille flow (ppf) driven by a pressure gradient, finding exact solutions is difficult. We use dynamic mode decomposition (DMD), a dimensionality reduction method for sequential data, to identify and approximate low-dimensional dynamics without knowing any exact solutions. Most spatially-extended systems are equivariant under translations, and in such cases spatial drifts dominate DMD, hindering its use in the search for and modelling of low-dimensional dynamics. We augment DMD with a symmetry reduction method trained on turbulent data to stop it from seeing translations as a feature, improving its ability to extract dynamical information in translation-equivariant systems. We find segments of turbulent trajectories that linearize well with their symmetry-reduced DMD spectra, akin to dynamics near exact solutions. Searching for harmonics in the spectra gives leads for periodic orbits with spatial drifts, one of which converges to a new solution.

In larger domains, turbulence can localize and coexist with surrounding laminar flow. Our preceding approaches are global, taking all of a domain into account at once, and cannot readily treat each localized patch individually. Working first in a minimal oblique domain that can host a single 1D-localized turbulent patch, we find that turbulence in ppf is connected to a stable periodic orbit at a flow velocity much lower than when turbulence is first onset. We show that, well in advance of sustained turbulence, chaos sets in explosively, and for long time horizons, time series are consistent with that of a random process.

Finally, in much larger domains, we study and compare 2D-localized turbulence that appears as large-scale inclined structures, called stripes, in ppf and pcf. While appearing similar, we find that stripes in these two settings differ significantly in terms of how they sustain themselves, and in higher velocities, how they proliferate.},
  author       = {Yalniz, Gökhan},
  issn         = {2663-337X},
  pages        = {155},
  publisher    = {Institute of Science and Technology Austria},
  title        = {{Transition to turbulence: Data-, solution-, and pattern-driven approaches}},
  doi          = {10.15479/AT-ISTA-19684},
  year         = {2025},
}

@inproceedings{22825,
  abstract     = {The empirical emergence of neural collapse—a surprising symmetry in the feature representations of the training data in the penultimate layer of deep neural
networks—has spurred a line of theoretical research aimed at its understanding.
However, existing work focuses on data-agnostic models or, when data structure is
taken into account, it remains limited to multi-layer perceptrons. Our paper fills
both these gaps by analyzing modern architectures in a data-aware regime: we
prove that global optima of deep regularized transformers and residual networks
(ResNets) with LayerNorm trained with cross entropy or mean squared error loss
are approximately collapsed, and the approximation gets tighter as the depth grows.
More generally, we formally reduce any end-to-end large-depth ResNet or transformer training into an equivalent unconstrained features model, thus justifying its
wide use in the literature even beyond data-agnostic settings. Our theoretical results
are supported by experiments on computer vision and language datasets showing
that, as the depth grows, neural collapse indeed becomes more prominent.},
  author       = {Súkeník, Peter and Lampert, Christoph and Mondelli, Marco},
  booktitle    = {39th Conference on Neural Information Processing Systems},
  isbn         = {9798331338275},
  issn         = {1049-5258},
  location     = {San Diego, CA, United States},
  pages        = {48646--48677},
  publisher    = {Neural Information Processing Systems Foundation},
  title        = {{Neural collapse is globally optimal in deep regularized ResNets and transformers}},
  doi          = {10.52202/085713-1450},
  volume       = {38},
  year         = {2025},
}

@inproceedings{22824,
  abstract     = {We present new fast-rate PAC-Bayesian generalization bounds for multi-task and
meta-learning in the unbalanced setting, i.e. when the tasks have training sets of
different sizes, as is typically the case in real-world scenarios. Previously, only
standard-rate bounds were known for this situation, while fast-rate bounds were
limited to the setting where all training sets are of equal size. Our new bounds
are numerically computable as well as interpretable, and we demonstrate their
flexibility in handling a number of cases where they give stronger guarantees
than previous bounds. Besides the bounds themselves, we also make conceptual
contributions: we demonstrate that the unbalanced multi-task setting has different
statistical properties than the balanced situation, specifically that proofs from
the balanced situation do not carry over to the unbalanced setting. Additionally,
we shed light on the fact that the unbalanced situation allows two meaningful
definitions of multi-task risk, depending on whether all tasks should be considered
equally important or if sample-rich tasks should receive more weight than samplepoor ones.},
  author       = {Zakerinia, Hossein and Lampert, Christoph},
  booktitle    = {39th Conference on Neural Information Processing Systems},
  isbn         = {9798331338275},
  issn         = {1049-5258},
  location     = {San Diego, CA, United States},
  pages        = {9062--9093},
  publisher    = {Neural Information Processing Systems Foundation},
  title        = {{Fast rate bounds for multi-task and meta-learning with different sample sizes}},
  doi          = {10.52202/085713-0278},
  volume       = {38},
  year         = {2025},
}

@inproceedings{20299,
  abstract     = {Deterministic Markov Decision Processes (DMDPs) are a mathematical framework for decision-making where the outcomes and future possible actions are deterministically determined by the current action taken. DMDPs can be viewed as a finite directed weighted graph, where in each step, the controller chooses an outgoing edge. An objective is a measurable function on runs (or infinite trajectories) of the DMDP, and the value for an objective is the maximal cumulative reward (or weight) that the controller can guarantee. We consider the classical mean-payoff (aka limit-average) objective, which is a basic and fundamental objective.

Howard's policy iteration algorithm is a popular method for solving DMDPs with mean-payoff objectives. Although Howard's algorithm performs well in practice, as experimental studies suggested, the best known upper bound is exponential and the current known lower bound is as follows: For the input size I, the algorithm requires (math formular) iterations, where (math formular) hides the poly-logarithmic factors, i.e., the current lower bound on iterations is sub-linear with respect to the input size. Our main result is an improved lower bound for this fundamental algorithm where we show that for the input size I, the algorithm requires (math formular) iterations.},
  author       = {Asadi, Ali and Chatterjee, Krishnendu and De Raaij, Jakob},
  booktitle    = {The 41st Conference on Uncertainty in Artificial Intelligence},
  issn         = {2640-3498},
  location     = {Rio de Janeiro, Brazil},
  pages        = {223--232},
  publisher    = {ML Research Press},
  title        = {{Lower bound on Howard policy iteration for deterministic Markov Decision Processes}},
  volume       = {286},
  year         = {2025},
}

@inproceedings{20297,
  abstract     = {A standard model that arises in several applications in sequential decision-making is partially observable Markov decision processes (POMDPs) where a decision-making agent interacts with an uncertain environment. A basic objective in POMDPs is the reachability objective, where given a target set of states, the goal is to eventually arrive at one of them.

The limit-sure problem asks whether reachability can be ensured with probability arbitrarily close to 1. In general, the limit-sure reachability problem for POMDPs is undecidable. However, in many practical cases, the most relevant question is the existence of policies with a small amount of memory. In this work, we study the limit-sure reachability problem for POMDPs with a fixed amount of memory. We establish that the computational complexity of the problem is NP-complete.},
  author       = {Asadi, Ali and Chatterjee, Krishnendu and Saona Urmeneta, Raimundo J and Shafiee, Ali},
  booktitle    = {The 41st Conference on Uncertainty in Artificial Intelligence},
  issn         = {2640-3498},
  location     = {Rio de Janeiro, Brazil},
  pages        = {238--247},
  publisher    = {ML Research Press},
  title        = {{Limit-sure reachability for small memory policies in POMDPs is NP-complete}},
  volume       = {286},
  year         = {2025},
}

@article{20646,
  abstract     = {Describing general quantum many-body dynamics is a challenging task due to the exponential growth of the Hilbert space with system size. The time-dependent variational principle (TDVP) provides a powerful tool to tackle this task by projecting quantum evolution onto a classical dynamical system within a variational manifold. In classical systems, periodic orbits play a crucial role in understanding the structure of the phase space and the long-term behavior of the system. However, finding periodic orbits is generally difficult, and their existence and properties in generic TDVP dynamics over matrix product states have remained largely unexplored. In this work, we develop an algorithm to systematically identify and characterize periodic orbits in TDVP dynamics. Applying our method to the periodically kicked Ising model, we uncover both stable and unstable periodic orbits. We characterize the Kolmogorov-Arnold-Moser tori in the vicinity of stable periodic orbits and track the change of the periodic orbits as we modify the Hamiltonian parameters. We observe that periodic orbits exist at any value of the coupling constant of the kicked Ising model between prethermal and fully thermalizing regimes, but their relevance to quantum dynamics and imprint on quantum eigenstates diminishes as the system leaves the prethermal regime. Our results demonstrate that periodic orbits provide valuable insights into the TDVP approximation of quantum many-body evolution and establish a closer connection between quantum and classical chaos.},
  author       = {Petrova, Elena and Ljubotina, Marko and Yalniz, Gökhan and Serbyn, Maksym},
  issn         = {2691-3399},
  journal      = {PRX Quantum},
  number       = {4},
  publisher    = {American Physical Society},
  title        = {{Finding periodic orbits in projected quantum many-body dynamics}},
  doi          = {10.1103/tldp-kvkd},
  volume       = {6},
  year         = {2025},
}

@inproceedings{22828,
  abstract     = {We propose Joint Moment Estimation (JME), a method for continually and privately estimating both the first and second moments of a data stream with reduced noise compared to naive approaches. JME supports the matrix mechanism and exploits a joint sensitivity analysis to identify a privacy regime in which the second-moment estimation incurs no additional privacy cost, thereby improving accuracy while maintaining privacy. We demonstrate JME’s effectiveness in two applications: estimating the running mean and covariance matrix for Gaussian density estimation and model training with DP-Adam.},
  author       = {Kalinin, Nikita and Upadhyay, Jalaj and Lampert, Christoph},
  booktitle    = {Advances in Neural Information Processing Systems},
  isbn         = {9798331338275},
  issn         = {1049-5258},
  location     = {San Diego, CA, United States},
  pages        = {64002--64047},
  publisher    = {Neural Information Processing Systems Foundation},
  title        = {{Continual release moment estimation with differential privacy}},
  doi          = {10.52202/085713-1924},
  volume       = {38},
  year         = {2025},
}

@inproceedings{22827,
  abstract     = {Quantized training of Large Language Models (LLMs) remains an open challenge, as maintaining accuracy while performing all matrix multiplications in low precision has proven difficult. This is particularly the case when fine-tuning pre-trained models, which can have large weight, activation, and error (output gradient) outlier values that make lower-precision optimization difficult. To address this, we present HALO, a new quantization-aware training approach for Transformers that enables accurate and efficient low-precision training by combining 1) strategic placement of Hadamard rotations in both forward and backward passes, which mitigate outliers, 2) high-performance kernel support, and 3) FSDP integration for low-precision communication. Our approach ensures that all large matrix multiplications during the forward and backward passes are executed in lower precision. Applied to LLaMa models, HALO achieves near-full-precision-equivalent results during fine-tuning on various tasks, while delivering up to 1.41x end-to-end speedup for full fine-tuning on RTX 4090 GPUs. HALO efficiently supports both standard and parameter-efficient fine-tuning (PEFT). Our results demonstrate the first practical approach to fully quantized LLM fine-tuning that maintains accuracy in INT8 and FP6 precision, while delivering performance benefits.},
  author       = {Ashkboos, Saleh and Nikdan, Mahdi and Tabesh, Soroush and Lopez Castro, Roberto and Hoefler, Torsten and Alistarh, Dan-Adrian},
  booktitle    = {39th Conference on Neural Information Processing Systems},
  isbn         = {9798331338275},
  issn         = {1049-5258},
  location     = {San Diego, CA, United States},
  pages        = {131755--131780},
  publisher    = {Neural Information Processing Systems Foundation},
  title        = {{HALO: Hadamard-assisted lower-precision optimization for LLMs}},
  doi          = {10.52202/085713-3966},
  volume       = {38},
  year         = {2025},
}