@inproceedings{22302, abstract = {Speculative generation has emerged as a promising technique to accelerate inference in large language models (LLMs) by leveraging parallelism to verify multiple draft tokens simultaneously. However, the fundamental limits on the achievable speedup remain poorly understood. In this work, we establish the first “tight” lower bounds on the runtime of any deterministic speculative generation algorithm. This is achieved by drawing a parallel between the token generation process and branching random walks, which allows us to analyze the optimal draft tree selection problem. We prove, under basic assumptions, that the expected number of tokens successfully predicted per speculative iteration is bounded as \mathbb{E}[X] ≤ (𝜇 + 𝜇(2))log(B )/𝜇2 + O(1), where B is the verifier’s batch size, 𝜇 is the expected entropy of the verifier’s output distribution, and 𝜇(2) is this entropy’s second moment. This result provides new insights into the limits of parallel token generation, and could guide the design of future speculative decoding systems. Empirical evaluations on Llama models validate our theoretical predictions, confirming the tightness of our bounds in practical settings.}, author = {Pankratov, Sergei and Alistarh, Dan-Adrian}, booktitle = {Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics}, location = {Rabat, Morocco}, pages = {6404–6418}, publisher = {Association for Computational Linguistics}, title = {{Speculative decoding speed-of-light: Optimal lower bounds via branching random walks}}, doi = {10.18653/v1/2026.eacl-long.301}, year = {2026}, } @unpublished{22276, abstract = {Tissue tension is a key determinant of tissue shape, and its regulation is essential for both morphogenesis and the maintenance of tissue integrity. During zebrafish embryogenesis, the enveloping layer (EVL) – an epithelial monolayer covering the blastoderm – undergoes extensive spreading that is driven by pulling forces exerted at its margin and more than doubles its surface area. Yet whether and how the EVL actively regulates its tissue tension during this process remains unclear. Here, we show that the EVL maintains constant tissue tension while spreading, and that it achieves this by reducing apical cell contractility in response to the same pulling forces that drive its spreading. We identify a mechanosensitive pathway underlying this response, mediated by the scaffold/adaptor protein Kibra regulating the activity of atypical protein kinase C (aPKC) at the apical domain of EVL cells. Under low mechanical stretch, Kibra forms condensates at the base of actin-based apical projections, where it activates Myosin II to increase apical contractility through aPKC downregulation. As mechanical stretch increases, apical projections disassemble, Kibra condensates dissolve, and aPKC activity rises. Elevated aPKC activity in turn reduces apical contractility by reducing Myosin II activity, thereby maintaining constant tissue tension despite increased mechanical stretch. Together, these findings reveal a mechanosensitive mechanism that enables robust adaptation of tissue tension to changing mechanical stretch, ensuring efficient tissue spreading and morphogenesis.}, author = {Hino, Naoya and Kapoor, Tushna and Gubbala, Uday R and Hannezo, Edouard B and Heisenberg, Carl-Philipp J}, keywords = {Epithelial spreading, tissue tension, mechanosensation, aPKC, Kibra, zebrafish}, publisher = {Institute of Science and Technology Austria}, title = {{Apical domain mechanosensation regulates tissue tension homeostasis}}, year = {2026}, } @article{22322, abstract = {We study the problem of continually releasing statistics of an evolving dataset under differential privacy. In the event-level setting, we show the first polynomial lower bounds on the additive error for insertions-only graph problems such as maximum matching, degree histogram and k-core number computation. These results represent an exponential improvement on the polylogarithmic lower bounds of Fichtenberger, Henzinger and Ost [ESA 2021] for the former two problems, and are the first lower bounds in the continual release setting for the latter problem. Our results run counter to the intuition that the difference between insertions-only vs fully dynamic updates causes the gap between polylogarithmic and polynomial additive error. Indeed, we show that for estimating the size of the maximum matching or k-core number of a vertex, allowing small multiplicative approximations is what brings the additive error down to polylogarithmic. We complement these results with improved upper bounds on the additive error when no multiplicative approximation is allowed. Beyond graphs, our techniques also show that polynomial additive error is unavoidable for the Simultaneous Norm Estimation problem in the insertions-only setting. When multiplicative approximations are allowed, we circumvent this lower bound by giving the first continual mechanism with polylogarithmic additive error under (1 + ζ) multiplicative approximations, for any ζ > 0, for estimating all monotone symmetric norms simultaneously. In the item-level setting, we show polynomial lower bounds on the product of the multiplicative and the additive error of continual mechanisms for a large range of graph problems. To the best of our knowledge, these are the first lower bounds shown for any differentially private mechanism under continual release with multiplicative error. To obtain these results, we prove a new lower bound on the product of multiplicative and additive error for the 1-Way-Marginals problem, and give reductions from 1-Way-Marginals to our desired graph problems. This generalizes the prior results of Hardt and Talwar [STOC 2010] and Bun, Ullman and Vadhan [STOC 2014, SIAM J. Comput. 2018], who gave lower bounds on the additive error for the special case of mechanisms with no multiplicative error.}, author = {Aryanfard, Bardiya and Henzinger, Monika H and Saulpic, David and Sricharan, A. R.}, issn = {2836-6573}, journal = {Proceedings of the ACM on Management of Data}, number = {2}, pages = {1--27}, publisher = {Association for Computing Machinery}, title = {{Improved lower bounds for privacy under continual release}}, doi = {10.1145/3801903}, volume = {4}, year = {2026}, } @article{22326, abstract = {In many developmental systems, cells differentiate into a tissue by reading out morphogen concentration fields, a process fundamentally limited by noise. How much can the precision of this process be improved by nonlocal information, e.g., via cell-cell communication? Using a Bayes-optimal framework, we show that positional inference depends crucially on morphogen spatial correlations and on the "structural prior" that encodes the geometry of the cellular lattice performing the readout, thereby determining what a cell can reliably assume about the position of its neighbors when interpreting nonlocal morphogen signals. We derive upper bounds on positional information gain due to nonlocal readout and identify signal processing algorithms that approximate optimal positional inference, as well as simple chemical reaction schemes which implement such algorithms. Our theory suggests that correlational information can be exploited to significantly enhance developmental precision.}, author = {Zhang, Chen Y and Mateu Hoyos, Pablo and Brückner, David and Tkačik, Gašper}, issn = { 1079-7114}, journal = {Physical Review Letters}, publisher = {American Physical Society}, title = {{Nonlocal decoding of positional and correlational information during development}}, doi = {10.1103/mbjk-v4ym}, volume = {137}, year = {2026}, } @article{21777, abstract = {The advantageous characteristics attributed to the 19F nucleus have made it a popular target for nuclear magnetic resonance (NMR) once again in recent years. Aside from solution NMR, an increasing number of studies have been conducted applying solid-state magic-angle spinning (MAS) NMR to fluorine-labelled samples. Here, the high chemical shift anisotropy and strong dipolar couplings can be utilised to get structural insights into proteins and measure long distances. Despite increasing popularity and promising benefits, the sensitivity of biomolecular 19F MAS NMR often suffers from slow longitudinal T1 relaxation and therefore long recycle delays. In this work, we expand paramagnetic doping, an approach commonly used to reduce proton T1 relaxation times, to 19F-labelled biological samples. We study the effect of Gd(DTPA) and Gd(DTPA-BMA) on 19F T1 and T2, and 13C T1 and T2 relaxation in a [5-19F13C]-tryptophan-labelled protein via 19F-detected MAS NMR experiments. The observed paramagnetic relaxation enhancement substantially reduces measurement times of 19F MAS NMR experiments without compromising resolution. Additionally, we report the chemical shift assignments of all four fluorotryptophan signals in the 12×39 kDa-large protein TET2 using a mutagenesis approach.}, author = {Becker, Lea Marie and Toscano, Giorgia and Kapitonova, Anna and Singh, Rajkumar and Guillerm, Undina and Lichtenecker, Roman J. and Schanda, Paul}, issn = {2699-0016}, journal = {Magnetic Resonance}, number = {1}, pages = {29--37}, publisher = {Copernicus Publications}, title = {{Accelerated 19F biomolecular magic-angle spinning NMR with paramagnetic dopants}}, doi = {10.5194/mr-7-29-2026}, volume = {7}, year = {2026}, } @phdthesis{22255, abstract = {This thesis studies spectral rigidity and nonrigidity phenomena in dynamical systems. The central question is whether a dynamical system can be determined, up to a natural conjugacy, from its spectrum. We consider three related spectra: the length spectrum, the action spectrum, and the Lyapunov spectrum. The first part of the thesis concerns Liouville metrics on the two-dimensional torus. It is a long-standing folklore conjecture that Liouville metrics are the only integrable metrics on the torus. We prove a length-spectral rigidity result for linear conformal deformations of Liouville metrics by exploiting the dynamical properties of the rational tori -- analogues of the resonant convex caustics in billiards. We also establish a complementary classification result showing that marked-length-isospectral Liouville metrics are characterized by rearrangements of the one-dimensional functions appearing in their conformal factors, generalizing a theorem of Abbondandolo-Mazzucchelli. In particular, the second result gives nonrigidity examples within the class of Liouville metrics. The second part of the thesis studies the standard map from the viewpoint of action and Lyapunov spectra. We construct nontrivial deformations of the standard map which preserve the symplectic actions (respectively, the Lyapunov exponents) of infinitely many periodic orbits accumulating on an invariant curve. The proof combines a resonant normal form construction with Picard iteration schemes to obtain a sequence of periodic orbits accumulating on an invariant curve with a Liouville rotation number. Within the resonant normal forms we capture the dependence of these periodic orbits on the resonant Fourier coefficients of the dynamics on the invariant curve and, using the contraction mapping principle, obtain a suitable deformation achieving the prescribed spectral data associated with this sequence of orbits. The result can be viewed as a symplectic twist-map analogue of a length-spectral nonrigidity phenomenon for Riemannian manifolds and convex billiards, and it motivates the existence problem for similar 'partially length-isospectral' deformations of strictly convex billiard tables. }, author = {Li, Yunzhe}, issn = {2663-337X}, pages = {131}, publisher = {Institute of Science and Technology Austria}, title = {{Spectral rigidity and nonrigidity of dynamical systems}}, doi = {10.15479/AT-ISTA-22255}, year = {2026}, } @article{22370, abstract = {We report a unified method for the synthesis of α-polyhalomethyl amines from alkenes and alkynes, enabled by readily available hemiaminal reagents. This operationally simple transformation allows for the introduction of not only well-established CF3 and CF2H groups, but also the synthetically (and medicinally) underexplored CF2Cl and CFClH motifs—thereby broadening access to previously inaccessible chemical space of halogenated amine scaffolds. The method displays broad substrate scope and functional-group tolerance while operating under mild conditions. Late-stage functionalization of drug-like derivatives of Oxaprozin, Erlotinib, and Ibuprofen (among others) is reported.}, author = {Hofmeister, Angela K. and Iannelli, Giulia and Angyal, Péter and Malandain, Augustin and Kaiser, Daniel and Maryasin, Boris and Kählig, Hanspeter and Barel, Matteo and Novarino, Gaia and Maulide, Nuno}, issn = {1521-3773}, journal = {Angewandte Chemie International Edition}, publisher = {Wiley}, title = {{Unified synthesis of unconventional α-polyhalogenated amines through hydroaminoalkylation}}, doi = {10.1002/anie.1233707}, year = {2026}, } @article{22323, abstract = {Arrays of Josephson junctions can be tuned through anomalous metallic, quantum-critical, and insulating regimes. We introduce an alternative experimental probe, capturing microwave radiation across all three regimes, using a two-dimensional array of superconductor-semiconductor hybrid Josephson junctions as a model system. Our approach allows calibration of the sample’s circuit parameters and provides isolation from measurement back-action effects. We measure the radiation temperature of the anomalous metal and find that it is hotter than both the quantum-critical and insulating regimes. We further show that the anomalous metallic regime is more susceptible to additional heating than other regimes, explaining its emergence in otherwise thermalized systems. Turning to the quantum-critical regime, we discover nonlinear scaling of radiative noise with applied bias, consistent with theoretical predictions of universal nonequilibrium behavior at quantum-critical points.}, author = {Galvin, Kristen W and Bubis, Anton and Mikalsen, Melissa and Schiela, William F. and Elfeky, Bassel H. and Strickland, William M. and Phan, Duc T and Shabani, Javad and Higginbotham, Andrew P}, issn = {2331-7019}, journal = {Physical Review Applied}, publisher = {American Physical Society}, title = {{Microwave radiometry of a quantum-critical hybrid Josephson array}}, doi = {10.1103/75bl-mm3b}, volume = {26}, year = {2026}, } @phdthesis{21021, abstract = {This thesis examines how geometry and topology intersect in the representation, transformation, and analysis of complex shapes. It considers how continuous manifolds relate to their discrete analogues, how topological structures evolve in persistence vineyards, and how tools from topological data analysis can illuminate problems in mathematical physics. Central to this exploration is the question of how structure, both geometric and topological, persists or changes under approximation, sampling, or deformation. The work develops new approaches to skeletal and grid-based representations of surfaces, reveals the full expressive capacity of persistence vineyards, and applies topological methods to the longstanding problem of equilibria in electrostatic fields. These threads braid together into a broader understanding of how topology and geometry inform one another across theory, computation, and application.}, author = {Fillmore, Christopher D}, issn = {2663-337X}, pages = {122}, publisher = {Institute of Science and Technology Austria}, title = {{Braiding geometry and topology to study shapes and data}}, doi = {10.15479/AT-ISTA-21021}, year = {2026}, } @inproceedings{22327, abstract = {Population protocols are a model of distributed computing where 𝑛 agents, each a simple finite-state machine, interact in pairs to solve a common task against a (adversarial) interaction scheduler. This model was intensively studied in recent years; in particular, the problem of relative majority received much attention: Each agent starts with an input opinion (or color) out of 𝑘 possibilities, and the goal is for each agent to eventually output the color with the largest support in the population. Before our work, the state complexity (the minimum number of states required per agent) was only known to be between Ω(𝑘 2 ) and𝑂(𝑘 7 ). Our main contribution is a population protocol that solves the relative majority problem with 𝑘 3 states. We achieve this result with a new protocol called Circles. While prior approaches in the literature relied on duels of agents to find the majority color — an approach that proved effective for the case with two colors — Circles partitions the agents into circular linked lists of decreasing sizes, with the property that no two agents with the same initial color lie in the same circle. We show that Circles always correctly computes the desired structure against the most adversarial of schedulers (weakly fair). We then show that a trivial extension of Circles solves the relative majority problem. We extend our protocol to handle various tie-breaking mechanisms or to support the case where the agents do not share a prior ordering of the colors. Finally, we show that a modification of Circles solves the ranking problem with 2 · 𝑘^4 states, where each agent must output the rank of its initial color in the population.}, author = {Breitkopf, Tom-Lukas and Dallot, Julien and El-Hayek, Antoine and Schmid, Stefan}, booktitle = {Proceedings of the ACM Symposium on Principles of Distributed Computing}, isbn = {9798400725128}, location = {Egham, United Kingdom}, pages = {414 -- 424}, publisher = {Association for Computing Machinery}, title = {{Ranking opinions with few states in population protocols}}, doi = {10.1145/3796701.3815913}, year = {2026}, } @article{14278, abstract = {The Birkhoff conjecture says that the boundary of a strictly convex integrable billiard table is necessarily an ellipse. In this article, we consider a stronger notion of integrability, namely, integrability close to the boundary, and prove a local version of this conjecture: a small perturbation of almost every ellipse that preserves integrability near the boundary, is itself an ellipse. We apply this result to study local spectral uniqueness of ellipses using the connection between the wave trace of the Laplacian and the dynamics near the boundary and establish local uniqueness for almost all of them.}, author = {Koval, Illya}, issn = {1432-1297}, journal = {Inventiones Mathematicae}, pages = {221--298}, publisher = {Springer Nature}, title = {{Local strong Birkhoff conjecture and local spectral rigidity of almost every ellipse}}, doi = {10.1007/s00222-025-01397-y}, volume = {244}, year = {2026}, } @inproceedings{21720, abstract = {We present an exact fully-dynamic minimum cut algorithm that runs in 𝑛𝑜⁡(1) deterministic update time when the minimum cut size is at most 2Θ⁡(log3/4−𝑐⁡𝑛) for any 𝑐 >0, improving on the previous algorithm of Jin, Sun, and Thorup (SODA 2024) whose minimum cut size limit is (log⁡𝑛)𝑜⁡(1). Combined with graph sparsification, we obtain the first (1 +𝜖)-approximate fully-dynamic minimum cut algorithm on weighted graphs, for any 𝜖 ≥2−Θ⁡(log3/4−𝑐⁡𝑛), in 𝑛𝑜⁡(1) randomized update time. Our main technical contribution is a deterministic local minimum cut algorithm, which replaces the randomized LocalKCut procedure from El-Hayek, Henzinger, and Li (SODA 2025).}, author = {El-Hayek, Antoine and Henzinger, Monika H and Li, Jason}, booktitle = {Proceedings of the Annual ACM SIAM Symposium on Discrete Algorithms}, issn = {1557-9468}, location = {Vancouver, Canada}, pages = {613--663}, publisher = {Society for Industrial and Applied Mathematics}, title = {{Deterministic and exact fully-dynamic minimum cut of superpolylogarithmic size in subpolynomial time}}, doi = {10.1137/1.9781611978971.25}, volume = {2026}, year = {2026}, } @article{22403, abstract = {Linear phase‐contrast scanning transmission electron microscopy (STEM) techniques compatible with high‐throughput 4D‐STEM acquisition are widely used to enhance phase contrast in weakly scattering and beam‐sensitive materials. In these modalities, contrast transfer is often suppressed at low spatial frequencies, resulting in a characteristic contrast gap that limits contrast. Approaches that retain low‐frequency phase contrast exist but typically require substantially increased experimental complexity, restricting routine use. Dark‐field STEM imaging captures this missing low‐frequency information through electrons scattered outside the bright‐field disk, but discards a large fraction of the scattered signal and is therefore dose‐inefficient. Fused Full‐field STEM (FF‐STEM) is introduced as a 4D‐STEM imaging modality that overcomes these limitations by combining ptychographic phase reconstruction with tilt‐corrected dark‐field imaging within a single acquisition. Bright‐field data are used to estimate probe aberrations and reconstruct a high‐resolution phase image, while dark‐field data provide complementary low‐frequency contrast. The two channels are fused in Fourier space using Wiener‐band weighting based on the spectral signal‐to‐noise ratio, yielding transfer‐gap‐free images with high contrast. FF‐STEM preserves the upsampling and depth‐sectioning capabilities of ptychography, adds robust low‐frequency contrast characteristic of dark‐field imaging, and enables dose‐efficient, near–real‐time reconstruction.}, author = {You, Shengbo and Varnavides, Georgios and Khavnekar, Sagar and Palatkin, Nikita and Shao, Sihan and Wu, Mingjian and Stroppa, Daniel and Chernikova, Darya and Zhu, Baixu and Egoavil, Ricardo and Vespucci, Stefano and Krishnan, Dileep and Ye, Xingchen and Schur, Florian KM and Spiecker, Erdmann and Pelz, Philipp}, issn = {2198-3844}, journal = {Advanced Science}, publisher = {Wiley}, title = {{Gap‐free information transfer in 4D‐STEM via fusion of complementary scattering channels}}, doi = {10.1002/advs.76620}, year = {2026}, } @article{21006, abstract = {Modern experimental methods in programmable self-assembly make it possible to precisely design particle concentrations, shapes and interactions. However, more physical insight is needed before we can take full advantage of this vast design space to assemble nanostructures with complex form and function. Here we show how a substantial part of this design space can be quickly and comprehensively understood by identifying a class of thermodynamic constraints that act on it. These thermodynamic constraints form a high-dimensional convex polyhedron that determines which nanostructures can be assembled at high equilibrium yield and reveals limitations that govern the coexistence of structures. We validate our predictions through detailed, quantitative assembly experiments of nanoscale particles synthesized using DNA origami. Our results uncover physical relationships underpinning many-component programmable self-assembly in equilibrium and form the basis for robust inverse design, applicable to various systems from biological protein complexes to synthetic nanomachines.}, author = {Hübl, Maximilian and Videbæk, Thomas E. and Hayakawa, Daichi and Rogers, W. Benjamin and Goodrich, Carl Peter}, issn = {1745-2481}, journal = {Nature Physics}, pages = {294--301}, publisher = {Springer Nature}, title = {{A polyhedral structure controls programmable self-assembly}}, doi = {10.1038/s41567-025-03120-3}, volume = {22}, year = {2026}, } @article{21295, abstract = {Depending on the type of flow, the transition to turbulence can take one of two forms: either turbulence arises from a sequence of instabilities or from the spatial proliferation of transiently chaotic domains, a process analogous to directed percolation. The former scenario is commonly referred to as a supercritical transition and frequently encountered in flows destabilized by body forces, whereas the latter subcritical transition is common in shear flows. Both cases are inherently continuous in a sense that the transformation from ordered laminar to fully turbulent fluid motion is only accomplished gradually with flow speed. Here we show that these established transition types do not account for the more general setting of shear flows subject to body forces. The combination of the two continuous scenarios leads to the attenuation of spatial coupling; with increasing forcing amplitude, the transition becomes increasingly sharp and eventually discontinuous. We argue that the suppression of laminar–turbulent coexistence and the approach towards a discontinuous phase transition potentially apply to a broad range of situations including flows subject to, for example, buoyancy, centrifugal or electromagnetic forces.}, author = {Yang, Bowen and Zhuang, Yi and Yalniz, Gökhan and Vasudevan, Mukund and Marensi, Elena and Hof, Björn}, issn = {1745-2481}, journal = {Nature Physics}, pages = {424--429}, publisher = {Springer Nature}, title = {{Discontinuous transition to shear flow turbulence}}, doi = {10.1038/s41567-025-03166-3}, volume = {22}, year = {2026}, } @article{21483, abstract = {Embryogenesis in the model plant Arabidopsis thaliana provides a framework for understanding how cell polarity and patterning coordinate with hormonal signalling to establish the plant body plan. Following fertilisation, the zygote divides asymmetrically to generate apical and basal lineages, establishing the apical–basal axis that defines future shoot and root poles. Genetic and molecular analyses of classical mutants including gnom, monopteros (mp), bodenlos (bdl) and topless revealed that localised auxin biosynthesis, directional transport and downstream transcriptional responses are central to apical–basal axis establishment and organ initiation. The main components of this regulation are polarly localised PIN auxin transporters and downstream modules involving MONOPTEROS and WUSCHEL-RELATED HOMEOBOX transcription factors. Advances in microscopy have transformed the study of Arabidopsis embryogenesis: fluorescence-compatible clearing reagents and three-dimensional reconstructions now permit quantitative analyses of cell geometry, division orientation, and cytoskeletal dynamics. Live ovule imaging setups with confocal laser scanning and multiphoton microscopes enable real-time observation of embryo development, while laser-assisted cell ablation can be used to probe cell-to-cell communication and fate plasticity. Together, these methodological breakthroughs position Arabidopsis embryos as a prime model for dissecting the chemical and biophysical cues that shape plant development.}, author = {Babic, David and Zupunski, Milan and Friml, Jiří}, issn = {1469-8137}, journal = {New Phytologist}, number = {3}, pages = {1483--1491}, publisher = {Wiley}, title = {{Imaging and genetic toolbox to study Arabidopsis embryogenesis}}, doi = {10.1111/nph.71072}, volume = {250}, year = {2026}, } @article{21486, abstract = {Sex-chromosome systems are highly variable across animals, but how they transition from one to another is not well understood. Diptera have undergone multiple sex-chromosome turnovers and expansions while maintaining their general chromosomal content, which makes them an ideal clade to study such transitions. We analyzed more than 100 dipteran whole-genome assemblies and identified 4 new lineages that underwent sex-chromosome turnover (in addition to the 5 previously reported). We find that the majority of turnovers happened in the group Schizophora, which tend to have fewer genes on Muller element F (the chromosome homologous to the ancestral insect X chromosome) than lower dipterans, a factor previously hypothesized to facilitate turnover. Most derived X chromosomes have higher GC content than autosomes, consistent with a high prevalence of male achiasmy in Diptera. In addition, an excess of gene movement out of the X is detected for most of these new X chromosomes, and many of these moved genes have high testis expression in Drosophila, suggesting that out-of-X gene movement contributes to the long-term demasculinization of X chromosomes.}, author = {Layana Franco, Lorena Alexandra and Toups, Melissa A and Vicoso, Beatriz}, issn = {2056-3744}, journal = {Evolution Letters}, number = {3}, publisher = {Oxford University Press}, title = {{Causes and consequences of sex-chromosome turnovers in Diptera}}, doi = {10.1093/evlett/qrag003}, volume = {10}, year = {2026}, } @misc{21857, abstract = {The availability of powerful open-source large language models (LLMs) opens exciting use cases, such as using personal data to fine-tune these models to imitate a user’s unique writing style. Two key requirements for this functionality are personalization–in the sense that the output should recognizably reflect the user’s own writing style—and privacy–users may justifiably be wary of uploading extremely personal data, such as their email archive, to a third-party service. In this paper, we demonstrate the feasibility of training and running such an assistant, which we call Panza, on commodity hardware, for the specific use case of email generation. Panza’s personalization features are based on a combination of parameter-efficient fine-tuning using a variant of the Reverse Instructions technique [1] and Retrieval-Augmented Generation (RAG) [2]. We demonstrate that this combination allows us to fine-tune an LLM to reflect a user’s writing style using limited data, while executing on extremely limited resources, e.g. on a free Google Colab instance. Our key methodological contribution is the first detailed study of evaluation metrics for this task, and of how different choices of system components–the use of RAG and of different fine-tuning approaches–impact the system’s performance. Additionally, we demonstrate that very little data - under 100 email samples - are sufficient to create models that convincingly imitate humans, showcasing a previously unknown attack vector in language models. We are releasing the full Panza code as well as three new email datasets licensed for research use.}, author = {Nicolicioiu, Armand and Iofinova, Eugenia B and Jovanovic, Andrej and Kurtic, Eldar and Nikdan, Mahdi and Panferov, Andrei and Markov, Ilia and Shavit, Nir and Alistarh, Dan-Adrian}, booktitle = {Third Conference on Parsimony and Learning (Proceedings Track)}, keywords = {LLMs, PEFT, LoRA, personalization, efficient ML}, location = {Tübíngen, Germany}, publisher = {OpenReview}, title = {{Panza: Investigating the feasibility of fully-local personalized text generation}}, year = {2026}, } @unpublished{21859, abstract = {As artificial neural networks, and specifically large language models, have improved rapidly in capabilities and quality, they have increasingly been deployed in real-world applications, from customer service to Google search, despite the fact that they frequently make factually incorrect or undesirable statements. This trend has inspired practical and academic interest in model editing, that is, in adjusting the weights of the model to modify its likely outputs for queries relating to a specific fact or set of facts. This may be done either to amend a fact or set of facts, for instance, to fix a frequent error in the training data, or to suppress a fact or set of facts entirely, for instance, in case of dangerous knowledge. Multiple methods have been proposed to do such edits. However, at the same time, it has been shown that such model editing can be brittle and incomplete. Moreover the effectiveness of any model editing method necessarily depends on the data on which the model is trained, and, therefore, a good understanding of the interaction of the training data distribution and the way it is stored in the network is necessary and helpful to reliably perform model editing. However, working with large language models trained on real-world data does not allow us to understand this relationship or fully measure the effects of model editing. We therefore propose Behemoth, a fully synthetic data generation framework. To demonstrate the practical insights from the framework, we explore model editing in the context of simple tabular data, demonstrating surprising findings that, in some cases, echo real-world results, for instance, that in some cases restricting the update rank results in a more effective update.}, author = {Iofinova, Eugenia B and Alistarh, Dan-Adrian}, booktitle = {arXiv}, title = {{Behemoth: Benchmarking unlearning in LLMs using fully synthetic data}}, doi = {10.48550/arXiv.2601.23153}, year = {2026}, } @phdthesis{21957, abstract = {This thesis investigates algorithmic certification and approximation methods for degenerate semidefinite programs (SDPs) and the singular roots of polynomial systems. In the first part, we present a hybrid symbolic-numeric algorithm for certifying the feasibility of weakly feasible, degenerate SDPs. By reformulating linear matrix inequalities (LMIs) into a structured polynomial system via facial reduction and incidence varieties, we guarantee the existence of an isolated exact solution. This algebraic reduction enables the certification of maximum-rank numerical approximations using methods from algebraic geometry. In the second part, we address the severe ill-conditioning and loss of quadratic convergence that plague standard path-tracking methods near isolated singular roots. To overcome this, we propose tracking algorithms that achieve superlinear convergence without the computational bloat characteristic of classical deflation techniques. By modeling the solution path as a generalized fractional Puiseux series, our approach combines an explicitly derived algebraic predictor with a localized hyperplane desingularization phase during the corrector step. Furthermore, we introduce a continuous path-limit method and an extension of the geometric sequence rule to directly extract exact fractional exponents. This bypasses traditional heuristic trial-and-error methods and explicitly accommodates sparse series expansions. Numerical experiments confirm that our method significantly reduces the cumulative number of matrix inversions while achieving high-accuracy root approximations, even for heavily degenerate systems exhibiting higher coranks.}, author = {Zapata, Jeferson}, isbn = {978-3-99078-079-4}, issn = {2663-337X}, pages = {89}, publisher = {Institute of Science and Technology Austria}, title = {{Overcoming degeneracy and singularity: Techniques for semidefinite programs and homotopy continuation endgames}}, doi = {10.15479/AT-ISTA-21957}, year = {2026}, }