Papers

Foundation models like V-JEPA are increasingly deployed across a heterogeneous edge-cloud continuum spanning smart glasses, phones, robots, and servers. Yet, they are typically trained as rigid, fixed-capacity networks bound to one device, supporting neither reconfigurable compute budgets nor collaborative execution. Existing approaches address only one axis at a time: slimmable and NAS-based methods provide width elasticity but assume single-device inference, whereas split-computing methods enable cross-device collaboration only for fixed partitions and static capacities. We introduce Elastic V-JEPA (EV-JEPA), a single model trained once and deployable everywhere along the continuum. Its width is reconfigurable, and it is natively partitionable across heterogeneous devices. At its core is Transition-Aware Triplet Training (T3). At each step we pick a narrow and a wide width, and train three paths: each width alone, plus a hybrid path that runs the early layers narrow and the rest wide. The hybrid switch point is where one device hands off to another, so the same weights learn both single-device and cross-device execution. Across various benchmarks, EV-JEPA falls within 1 accuracy point of independently trained baselines while improving over previous elastic training approaches by up to 11 accuracy points. Narrow-to-wide cross-device hybrid execution recovers to within 1.1 points of the highest-capacity standalone model and supports many-to-one, any-to-any, and variable-split configurations. Finally, we show that the elastic principle extends beyond the backbone to an elastic token compression mechanism that exposes controllable trade-offs among on-device computation, latency, and accuracy.

We study the inverse resonance problem for the one-dimensional Schrödinger operator on the half-line with a compactly supported complex-valued potential and a complex Robin boundary condition \[ u'(0)-a u(0)=0,\qquad a\in\mathbb C. \] The proof follows the transformation-operator, entire-function, and Fourier analytic strategy developed by Marletta, Shterenberg and Weikard for the Dirichlet problem \cite{MSW}. The Dirichlet Jost function $\psi(z,0)$ is replaced by the Robin characteristic function \[ F_a(z;q)=\psi_x(z,0;q)-a\psi(z,0;q). \] We establish the corresponding large-energy estimates, Hadamard factorization, finite-data stability, and uniqueness results. The principal new feature is that $F_a(z;q)=iz-a+O(1)$ rather than $1+O(z^{-1})$. In the transformation-operator argument the relevant boundary quantity is $B_x(0,t)-aB(0,t)$, together with the scalar term $B(0,0)$. For fixed $a\in\mathbb C$, and uniformly for $a$ in bounded subsets of $\mathbb C$, the resulting inverse resonance estimates have the same qualitative form as those proved in \cite{MSW} for the Dirichlet problem. In particular, for a fixed Robin parameter $a$, the complete set of Robin eigenvalues and resonances, counted with multiplicity, determines the potential uniquely. The text is produced with the help of AI ChatGPT for Academy.

Artificial neural networks (ANN) implemented on crossbar arrays of non-volatile synaptic devices are known to consume much less energy compared to that implemented on conventional computing units. Solution-processed resistive random-access memory (RRAM) devices based on transition-metal oxides are suitable as such synaptic devices because of their cheap and easy fabrication method, stability, and suitable memristive properties. Here, in NiO-based RRAM devices with Ti/Pt as the bottom electrode and Ag as the top electrode (NiO layer deposited through solution processing), we experimentally report forming-free binary switching between a high resistance state (HRS) and a low resistance state (LRS) as well as forming-free pulse-driven modulation (long-term potentiation (LTP), long-term depression (LTD)) through 30 intermediate synaptic states between HRS and LRS (around 5-bit synaptic weight resolution). Many of these states are measured to be stable at least up to 1000 seconds confirming their non-volatility and their utility in storage of synaptic weights for ANN crossbar arrays. On the contrary, most prior experimental reports (including reports both on solution-processing-based and sputtering-based RRAM devices) require forming to exhibit non-volatility. Avoiding forming reduces variability, eliminates the need for high voltage pulses, and makes the crossbar array design simpler. Further, we confirm the presence of oxygen vacancies in our materials stack using X-ray photo-electron spectroscopy (XPS). Finally, we show successful data classification on MNIST and CIFAR-10 data sets using a multi-layer perceptron (MLP) model and a convolution neural network (CNN) model respectively, where the synaptic weight updates of these ANN models are calibrated against the experimentally measured LTP-LTD characteristics of the NiO RRAM devices.

We consider functions $u : \R^2 \rightarrow \R^N$ from the class $\Bloc$ which locally minimize the relaxed version of a variational integral with a strictly convex density being in addition of linear growth. It turns out that the boundedness of entire local minimizers $u$ implies their constancy.

Sharing time-to-event data can leak sensitive information about individual participants in clinical studies. To protect the privacy of data, differential privacy (DP) framework offers rigorous protection for data sharing. The local variant of differential privacy is particularly appealing because each record is perturbed before it leaves the data owner or contributor. However, the injected perturbation distorts the true distribution of the source data, yielding low utility and invalid statistical inference with the released data. In this work, we propose a nonparametric estimator for the true survivor function via the deconvolution kernel function with locally differential private (LDP) right-censored survival data in which both the event times and the event indicators are privatized. Our estimator simultaneously de-biases the perturbation applied to the times and the errors induced by flipping the event indicators due to LDP. Theoretically, we establish the uniform consistency and point-wise asymptotic normality of the de-biased distribution estimator, and derive its uniform convergence rate together with the optimal bandwidth attaining that rate. Simulation studies and the analysis of a semi-synthetic dataset generated from the METABRIC breast cancer cohort demonstrate the performance of the proposed estimator across privacy budgets, censoring levels, and sample sizes. By the post-processing property of DP, the proposed de-biased estimator inherits the privacy guarantee while restoring valid inference on the survivor function.

We provide some extensions of the classical Calabi-Bernstein theorem to the case of entire solutions of certain nonlinear elliptic equations in two variables. Using an appropriate version of the Calabi correspondence, which formulates a duality between minimal graphs in $\mathbb{R}^3$ and maximal spacelike graphs in $\mathbb{L}^3$, our results are a consequence of recent Bernstein type results.

We develop a measure-theoretic framework for conditional multivariate total positivity of order two ($\mathrm{cmMTP}_2$), formulated through lattice inequalities without requiring the existence of probability densities. This formulation extends classical density-based notions of positive dependence to general probability measures, including singular distributions. We establish fundamental preservation properties under marginalization, monotone transformations, and weak convergence, thereby providing a unified treatment of dependence structures under distributional limits. We further investigate the behavior of conditional total positivity under Markov-kernel composition and derive characterizations for important copula families. In the Archimedean setting, the defining condition is expressed through log-convexity of generator derivatives, whereas for Gaussian distributions it corresponds to nonpositive off-diagonal entries of the precision matrix, or equivalently, nonnegative partial correlations. To connect these structural results with statistical inference, we introduce empirical lattice contrasts and establish their joint asymptotic normality. The resulting simultaneous testing procedures achieve asymptotic size control and admit a local power analysis. Our results provide a general foundation for density-free characterization, specification testing, and inference concerning positive dependence in multivariate probability models, particularly when conventional density-based methods are inapplicable or insufficient.

Pregnancy can disrupt schooling before infant care begins. We use monthly histories from the National Longitudinal Survey of Youth 1997 to study school interruptions during pregnancy and after childbirth and whether students later return to school or earn a diploma or GED. We jointly model conception and sustained interruptions in enrollment, allowing persistent unobserved differences to affect both processes. Under the timing-of-events identifying assumptions, pregnancy raises the monthly interruption rate to nearly four times the rate without pregnancy. Interruption rates are elevated before delivery and remain elevated afterward. After an interruption, students already exposed to pregnancy return to school and earn a diploma or GED less often within two years. These findings show that schooling disruption is already present during pregnancy, before the additional demands of infant care begin.

Counter-rotating vortex reconnection is a topology-transforming event that leads to strong vorticity amplification and is often associated with the most irregular regions of the velocity field. Detecting such events directly in turbulent flows is a major challenge due to the ambiguity inherent in the definition of a vortex and the rapid proliferation of intense small scale vortices. We develop a novel methodology by combining two tools. The first is a local in scale, space and time regularity indicator that isolates spatio-temporal regions of interest. The second is a reconnection indicator that exploits a well-known property of vortex reconnection namely, that the enstrophy (volume integral of the square of vorticity) flows from the symmetry to the collision plane. To examine the flow of enstrophy in the correct reference frame, a new method inspired by orientational order parameters in liquid crystals is developed. We then show a proof of concept for our reconnection indicator by comparing it with diagnostic methods available in the literature. The method is validated on experimental datasets from a turbulent von Kármán flow at two different resolutions, one of which allows us to probe the flow below the Kolmogorov scale, $\eta$. Our findings suggest that counter-rotating vortex reconnection could be a potential mechanism for the creation of motions at scales below $\eta$.

Reconstructing accurate camera poses and floor-plan layouts from sparse, unposed, wide-baseline RGB panoramas remains a challenging open problem. Direct multi-room floor-plan prediction is especially difficult in this setting: pose, geometry, semantics, and global layout structure are tightly coupled, yet a complete floor-plan is hard to learn as a single global output. Our key insight is to keep the model's predictions strictly local. Multi-view Ordered Wall Instance Segmentation (MOIS) is designed around this concept. Each panorama column predicts only the wall it observes plus the next two adjacent walls in clockwise order. The global wall partition, room layouts, and shared coordinates emerge from deterministic aggregation of overlapping local predictions. Building on MOIS, we propose Multi-view floor-plan Feed-Forward Network (MvFFN), which jointly predicts coarse camera poses, dense global geometry, and per-wall semantic outputs. A downstream Layout Chaining (LC) post processing pipeline then assembles these predictions into multi-room floor-plans with shared global coordinates, consistent wall identities, and connected room structure. The dense MOIS predictions recover cross-view wall identity and local within-room connectivity more accurately than composed modular baselines, and the assembled layouts substantially outperform prior methods on wall and junction level accuracy. We also introduce the ZInD-CrossView dataset, which augments ZInD with globally unique wall instances and cross-view segmentation labels for this task.

Computing the field within a sub-unit of cascaded systems is a recurring problem in wave physics. It is central to gradient-based inverse design, in situations where only a region of the system is modified while the rest remains fixed. Standard approaches either resimulate the entire cascade at every change, or require a detailed characterization of every surrounding sub-system. Both become impractical when many design iterations are needed. Here, we introduce a hybrid method combining the $S$-matrix approach, which propagates fields through cascaded systems, with the Green's function method, which gives the field generated by a source within a characterized system. In this approach, the sub-systems surrounding the region of interest need only be characterized once, after which the local field can be recomputed rapidly for any new configuration. We validate the method on a disordered lossy waveguide, finding excellent agreement with finite-element simulations. We also demonstrate its efficiency on a disordered multimode cavity, where it achieves a fifty-fold speed-up over full-wave simulation and reduces a wideband focusing calculation from eleven hours to ten minutes. This approach opens the way to fast, iterative field calculations in complex systems coupled to their environment, with direct applications to inverse design and wideband wavefront-shaping.

We study supersymmetric vacua of three-dimensional $\mathcal N=1$ effective theories from type IIA, type IIB/type I, and heterotic compactifications on $G_2$-structure spaces. Superpotential homogeneity and the physical scalar metric yield universal vacuum relations and axion stabilization constraints without specifying the full moduli-space metric. We express AdS$_3$ scalar masses through $\Delta$, the scaling dimension of the corresponding operator in the dual two-dimensional conformal field theory. In massive type IIA AdS$_3$ vacua, we derive the bound $\Delta_{\max}\geq(20+4\sqrt{67})/7\approx 7.5$ and an exact dimension $\Delta_b=6$ for $B_2$ axions on the torsion-free branch. Type IIB type I and heterotic AdS$_3$ vacua contain a universal volume-dilaton mode with $\Delta_{\phi s}=4$. On the singlet-torsion type IIB branch, we also obtain the volume-dilaton mode with $\Delta_{\phi s}=20/7$ and $\Delta_b=16/7$. Every supersymmetric Minkowski vacuum in the retained classical closed-string theories possesses an exact volume-dilaton flat direction. We characterize physical large-flux limits after tadpole constraints and axionic identifications, and identify torsion obstructions to isotropic scale separation.

We introduce the ratio of electron capture to positron decay rates of unstable nuclei as a probe for exotic tensor currents in the electroweak sector. We calculate nuclear structure corrections and newly derived radiative corrections and obtain competitive limits on exotic tensor couplings using historically available data to find $\varepsilon_T = (1.06 \pm 0.76)\times 10^{-3}$ which is competitive with collider searches and the global $\beta$-decay database. In doing so, we resolve a 50-year-old tension in the strongly hindered $^{22}$Na transition by retaining higher-order operators, obtaining a consistent description of its different weak-decay observables. Finally, we discuss the potential of this method to significantly improve tensor constraints using modern cryogenic quantum sensors.

Active learning has been shown to be more beneficial for student learning than lecturing. Consequently, researchers have developed named active learning strategies for use in physics instruction. To increase adoption and support implementation of these named methods, it is important to understand the backgrounds and experiences of instructors who use them. In this study, we interviewed 20 introductory physics and astronomy instructors who use one of Student-Centered Active Learning Environment with Upside-Down Pedagogies (SCALE-UP), Investigative Science Learning Environment (ISLE), Tutorials in Introductory Physics and Astronomy, or Peer Instruction to gain insight into their teaching experiences, professional development, pathways to learning about their method, and supports for and difficulties in implementation. We find for our sample that all implementers are involved with the discipline-based education research community, but there is variation in the extent of engagement. While instructors identify a variety of supports for their implementation, such as from the broader education community and departmental administration, the difficulties they experience are representative of both long-standing issues associated with active learning adoption (e.g., timing constraints) and new issues related to emergent technologies (e.g., Artificial Intelligence). Finally, we uncover differences in implementers across the four named methods, particularly related to their training for the method and the supports they received for implementation. Together, these findings underscore the importance of supporting multiple dissemination avenues for named active learning methods, point to a possible need for more training opportunities for SCALE-UP and ISLE in particular, and identify nascent issues in the contemporary active learning classroom that future work may further characterize and address.

We study how the eigenspace associated with a fixed graph Laplacian eigenvalue changes under positive edge updates, allowing both new edges and increases in existing edge weights. We show that a preserved eigenvalue may have inherited eigenvectors from the original eigenspace and emergent eigenvectors outside it. We characterize the inherited eigenspace and show that it is independent of the update weights. We characterize its orthogonal complement within the preserved eigenspace through a symmetric matrix whose size is at most the number of updated edges, yielding a multiplicity formula and an explicit reconstruction of the updated eigenspace. For fixed updated edges, we characterize when emergence is possible and show that, whenever it is, equal update weights suffice. The exceptional weights form the zero set of a nonzero multi-affine polynomial in reciprocal variables and have Lebesgue measure zero. Finally, emergence cannot occur at either of the two smallest Laplacian eigenvalues, counting multiplicities.

Speech large language models (LLMs) are often built using pretrained acoustic encoders as an audio interface to an LLM decoder. However, LLMs scale poorly with input sequence length, while acoustic encoders tend to produce substantially longer sequences than typical textual inputs. This motivates learning to compress encoder outputs while preserving task-oriented information. To this end, we introduce ASSET, a reinforcement learning (RL)-based framework that learns to adaptively compress encoder outputs by first partitioning them into variable-duration segments and then pooling each segment into a continuous audio token. We use group-relative policy optimization (GRPO) to train a segmenter with task performance feedback and a token-rate penalty. In automatic speech recognition (ASR) experiments, ASSET achieves performance comparable to a transcript-assisted character-boundary oracle with approximately 25% fewer audio tokens on English LibriSpeech, while improving on a phone-boundary oracle at a similar token rate on Mandarin AISHELL. Exploratory experiments on non-ASR tasks further illustrate how learned segmentation changes when the framework is adapted to tasks beyond transcription.

We propose a coding scheme for the insertion, deletion and substitution (IDS) channel tailored to the short strands of DNA-based data storage. Instead of an inner synchronisation code, a spatially coupled LDPC (SC-LDPC) code is placed across the strands, with its coupling axis aligned to the position along the strands. Every decoded position thus becomes an anchor to synchronise the neighbouring strands. The decoding waves of the SC-LDPC code carry these anchors from both ends of the strands towards the middle. We bound the achievable rate from above by the information rate of a single strand and from below by that of a separated receiver. For strands of 100 nucleotides (the relevant regime for DNA-based data storage), our scheme outperforms a rate-matched concatenated construction.

We propose ConicFlow, which frames constrained sampling from a flow matching model at inference time as an optimal control problem that balances deviation from the pretrained trajectory against reward, subject to hard terminal constraints. We discretize this problem and adapt receding-horizon control, optimizing only a short window of the trajectory at each step for computational tractability. We address the resulting sequence of receding-horizon subproblems with a receding-horizon primal-dual algorithm. Its dual multipliers capture the difficulty of satisfying each constraint and are refined throughout sampling, steering generation toward feasible final samples. We further propose a terminal-estimate variant that avoids accumulating local truncation errors by evaluating the reward and constraints at the predicted terminal point rather than at the end of the current horizon. Under standard regularity assumptions, we establish finite-horizon guarantees on suboptimality. Under convex-affine structure, we derive a sharper bound that scales linearly with omitted horizon length, and we show that the terminal-estimate variant is provably tighter when the flow dynamics change slowly. Experiments in both low- and high-dimensional settings with a recent large-scale flow matching model show that ConicFlow achieves high reward and strong constraint satisfaction while deviating less from the pretrained flow than recent baselines.

Scientists have many tools to author plots: graphical direct manipulation apps like a spreadsheet's chart builder, text-based programming tools like Matplotlib and ggplot2, or AI chat interfaces. Textual code is the most capable of these, but writing code, even with AI assistance, is a potentially poor fit for graphical chart design. We present Plottery, a Jupyter notebook extension that leverages API type annotations to add graphical direct manipulation to the popular Matplotlib plotting library. User interactions in Plottery's in-situ graphical interface are immediately reified in the user's code, so they may seamlessly mix-and-match between GUI, code, and AI interactions. In a user study, participants seamlessly alternated modalities, with the GUI used for straightforward visual changes, code for quick edits or duplicating code, and the AI for broad support. Plottery showcases graphical interaction that retains code's benefits.

Who should give up compute for AI safety research: customers or capability research? Safety paid for by customers delays no capability. Safety paid for by research slows the frontier, which gives knowledge more time to accumulate but postpones what a finished frontier provides. In a frontier-takeoff model, the cheaper source at a given allocation is set by a threshold in demand capacity. Demand capacity is the share of compute customers can use once the frontier is finished. Research should pay too only when demand capacity lies below the threshold. Where the threshold sits depends on how fast the value of customer services diminishes. In the calibration it rises with the service budget, because each further unit taken from customers costs them more. Safety paid for by customers should come early: its ratio to customer compute falls over the takeoff unless safety productivity rises fast. In a calibration with a tight service budget, early customer-funded safety lowers the probability of catastrophe from 75 to 51 percent. That is within one point of the best rule found that may also slow research.

We prove a central limit theorem for the quadratic matching cost between the uniform measure and its empirical counterpart on the flat torus in dimensions $d\ge5$. Let $\mu$ be Haar probability measure on $\mathbb{T}^d$ and let $\mu_n$ be the empirical measure of $n$ independent samples from $\mu$. Then $$ \begin{aligned} \frac{W_2^2(\mu_n,\mu)-\mathbb{E}W_2^2(\mu_n,\mu)} {\sqrt{\operatorname{Var}(W_2^2(\mu_n,\mu))}} &\Longrightarrow N(0,1),\\ c_dn^{-1-4/d} \le \operatorname{Var}(W_2^2(\mu_n,\mu)) &\le C_dn^{-1-4/d}, \end{aligned} $$ for positive constants $c_d,C_d$ depending only on $d$ and all sufficiently large $n$. The proof introduces a random-walk approach to matching fluctuations. We construct a random walk that jumps between neighboring cells of the optimal partition and use it to describe how the matching cost responds to changes in the sample. Estimates for this walk control the effect of replacing one sample point and the interaction between two replacements. These bounds yield the variance estimates and allow us to apply a martingale central limit theorem.

A Muon Collider is one of the most attractive routes to multi-TeV lepton collisions, but it requires the accelerator and detector communities to solve a tightly coupled set of problems that have no close precedent in an operating collider. Muons must be produced, captured, cooled, accelerated, and collided before they decay. The detector must then extract clean physics from an environment dominated by decay products from the beamline. Artificial intelligence is not a substitute for accelerator physics, detector design, or engineering validation. It is, however, unusually well matched to several of the central obstacles: high-dimensional design spaces, expensive simulations, coupled subsystem trade-offs, and workflows that are currently too slow or too fragile to support broad design exploration. This report distills the AI4MuC workshop into a summary of AI use cases with the potential to change the pace of Muon Collider R&D, organized around the following physics bottlenecks: muon production and capture, ionization cooling, magnet and RF technology, detector design in beam-induced background, background simulation and generation, reconstruction, and software.

Humans and animals do not learn vision from a fixed set of data: they move, interact, and learn from the visual consequences of their actions. Motivated by this developmental view, we study whether robots can improve visual learning by actively collecting experience through curiosity-driven interaction. We introduce DySCOVER, a self-supervised framework for active robot visual learning. A robot observes RGB frames, learns a visual autoencoder that encodes and reconstructs them, and trains a recurrent world model to predict the next visual latent state conditioned on the current state and action. The agent's policy is trained only with intrinsic rewards: reconstruction error encourages visually novel observations, while the world model's error encourages transitions whose consequences are not yet predictable. This prediction error biases the policy toward more intricate trajectories that enrich the agent's set of experiences with observations beyond those encouraged by visual novelty alone. Our experiments conducted with robotic arms as the embodied agents show that adding predictive curiosity improves the vision model over using visual curiosity alone. Across variations in initial conditions, experience capacity, embodiment, and scene content, adding predictive curiosity consistently improved visual learning. These results show that predictive curiosity can improve active visual learning beyond reconstruction-based curiosity alone.

On a totally bounded metric space every non-contractive self-map is an isometry. For quasi-metric spaces, taken T_1 throughout, this fails: Zavarzina gave a hereditarily precompact counterexample and asked whether bijectivity restores rigidity; the toll spiral shows that it does not. We prove that compactness does: every compact quasi-metric space is expand-contract plastic, that is, each of its bijective non-contractive self-maps is an isometry. Countable compactness suffices, as does compactness of the conjugate topology. Bijectivity cannot be dropped: a compact, hereditarily precompact quasi-metric on the natural numbers makes the shift non-contractive and not an isometry. For arbitrary non-contractive self-maps we prove rigidity when the space is hereditarily precompact and its conjugate topology is sequentially compact, in particular when it is hereditarily precompact and Yoneda complete; neither hypothesis can be dropped. In both theorems the T_1 axiom can be weakened to R_0, with hereditary precompactness read sequentially, but not to T_0: we construct a T_0 space with all the remaining hypotheses of both theorems, carrying a bijective non-contractive self-map that is not an isometry. Under R_0 the two-sided case reduces to the metric theorem; the one-sided theorems do not, and plasticity does not pass to the symmetrization even for compact spaces. In finite dimension the unit ball of a T_1 asymmetric normed space, forward or symmetric, and its sphere are rigid for every non-contractive self-map; without T_1 the forward ball need not be; in infinite dimension the symmetric ball and sphere are not plastic in general, and the forward versions carry the quasi-metric form of the unit-ball question of Cascales, Kadets, Orihuela and Wingler, which is posed. The two theorems are machine-checked in Lean, and so are the counterexamples, apart from a few statements listed at the end.

Credential-database disclosure exposes Internet of Things gateways, edge servers and legacy verifiers to offline password guessing. Memory-hard key derivation functions (KDFs) are preferred but can exceed constrained-device budgets. ARMOR addresses this deployment gap with a schema-preserving verifier-side transform for database-only compromise, assuming an isolated Chain Registry. The registry holds secret per-account chains of injective operations; fixed terminal escape encoding preserves distinct pre-hash values under each chain while moving outputs outside printable ASCII. This disrupts evaluated stock tools' candidate coverage, but chain-aware attackers can generate the transformed bytes directly. We therefore evaluate a bounded chain-aware attack on $1{,}000$ paired salted-MD5 targets: it executes $89.35\times$ the plain arm's counted salted checks, reproduced on A6000 and A100. All target passwords are in the bounded pool; same-target ambiguity enumeration and simulated login trials recover every original password. These results describe the implemented attack, not minimum attacker cost. Bare-metal transforms take microseconds, while protected storage dominates the measured TrustZone-M components; these are component measurements, not complete authentication times. ARMOR thus provides bounded chain uncertainty under registry isolation with low transform cost, although an isolated high-entropy pepper remains cryptographically stronger. ARMOR composes as KDF(ARMOR(password)) when resources permit and the KDF accepts the complete transformed output.

Recent Speech Full-Duplex (FDX) models are capable of engaging in naturalistic conversation by listening and speaking at the same time. Building on this foundation, recent architectures introduce internal thinking streams to enhance their intelligence. To guarantee these streams process successfully in real time, existing methods enforce a fixed decoding rate, which assumes a hardware profile capable of a specific latency. We argue this static approach is severely suboptimal. Because internal thinking is intrinsically asynchronous, an ideal FDX model should sustain real-time conversational flow while allowing each stream to decode at its own hardware-adaptive rate. Thinking streams must process as fast as the hardware allows to deliver accurate answers faster. To solve this, we propose RUBATO, a novel multistream positional encoding that assigns independent clocks to entirely decouple the streams. This structural isolation empowers the model to zero-shot interpolate across any inference decoding speed. It drastically simplifies dataset construction by requiring training on only two extreme decoding speeds while smoothly adapting to any intermediate rate during deployment. RUBATO demonstrates consistent robustness across five spoken reasoning tasks. On Spoken GSM8K, a baseline model collapses to 1.7% accuracy at interpolated speeds. Conversely, RUBATO-Full sustains 48.7% to 51.1% accuracy across all rates, leveraging faster hardware to achieve a 6.5x shorter time-to-answer with a constant time-to-first-response. This advantage holds on BIG-Bench Audio and amplifies significantly during complex full-duplex interruptions. Specifically, when a user interjects while the model is actively speaking and reasoning, RUBATO-Full achieves 40.0% to 45.1% accuracy, whereas the baseline fails completely.

Gate-All-Around Field-Effect Transistors (GAAFETs) are emerging as successors to Fin Field-Effect Transistors (FinFETs) and are central to continued scaling at advanced semiconductor nodes. Despite their improved electrostatic control, recent measurements show that the drain current in GAAFETs deviates strongly from thermionic behavior at negative gate voltages, increasing as the gate voltage becomes more negative. This leakage has been attributed to direct gate-drain tunneling through the high-$k$ dielectric, enhanced by drain-induced dielectric barrier lowering (DIDBL). Here, we use predictive, charge-self-consistent Non-Equilibrium Green Function (NEGF) simulations based on the Contact Block Reduction (CBR) method to spatially resolve this leakage pathway in GAAFETs. Unlike approaches based primarily on phenomenological leakage models, our first-principles, multi-terminal quantum-transport framework treats the source, drain, and gate as current-carrying quantum-mechanical contacts and resolves the local current density throughout the device. The spatially resolved current density shows that gate-injected carriers tunnel through the high-$k$ dielectric preferentially near the drain side of the gate, where the dielectric barrier is most strongly lowered by the drain bias, and are collected by the drain. This provides a direct real-space picture of the DIDBL-driven gate-drain leakage mechanism.

This paper treats a system that performs back-to-back tasks. Each new task has a random matrix of parameters and can be processed in various processing modes. The processing mode determines the task duration and a vector of rewards (penalties are viewed as negative rewards). The reward vector can include quality, profit, or $-1$ times a penalty such as cost or energy. In wireless applications, tasks correspond to data transmissions and rewards/penalties correspond to throughput and distortion. Prior work treats the problem of maximizing the time average of a reward process subject to constraints. This paper seeks to maximize a concave function of the time average reward vector. This is difficult because it involves a multi-variable function of ratios of averages. We develop a single algorithm that does not know the probability distribution of the random task matrices. The algorithm is adaptive, meaning it responds efficiently (with optimal adaptation time) when probabilities change.

High benchmark scores in learning-based vulnerability detection mask severe degradation under distribution shifts. This paper demonstrates that calibrated embedding geometry exposes these hidden generalization boundaries on syntactically similar vulnerable and patched code pairs. An evaluation of CodeBERT and Qwen2.5-Coder across pattern, project, and temporal shifts on a strictly paired dataset reveals three main findings. First, whitened intermediate-layer embeddings with local nearest-neighbour scoring separate unfamiliar samples with a mean out-of-distribution AUROC of 0.739 (RQ1). Second, distance scores reveal a shift-specific reliability gradient where accuracy drops by over 35 percentage points and cross-project shifts cause the sharpest degradation, even though temporal shifts drift geometrically farthest (RQ2). Third, false positives on patched code systematically correlate with geometric proximity to vulnerable training neighbourhoods, increasing from 9.4% in the lowest-risk decile to 58.8% in the highest (RQ3). These calibrated distances establish a diagnostic signal that maps representation geometry directly to model degradation without requiring auxiliary models.

Optimization can be expensive when evaluating the objective and its gradient requires a large dataset or a costly simulation. A cheaper approximation (i.e., a surrogate) can provide a useful initial point, but the work spent on that surrogate must be justified by the saving on the original problem. We study how many iterations to perform on each surrogate before switching to a more accurate objective. For a given sequence of smooth, strongly convex objectives, we use convergence bounds and distances between minimizers to construct a model of the total iteration cost. We derive closed-form two-level budgets for Nesterov's accelerated gradient method and a convex reformulation for multiple levels. We also develop an online switching rule that updates the cost comparison using the gradient at the current point. We test the techniques on a logistic-regression problem using HIGGS data. The online rule reduces weighted iteration cost in all 27 tested configurations, with a median cost of 52.8% of direct optimization. The results illustrate why updating the switching decision can be more effective than fixing the budget from an initial convergence bound.