Papers

Modern networks must support changing topologies, configurations, and performance objectives, motivating fast and reliable performance estimation. Network digital twins (NDTs) enable what-if analysis for performance estimation in such network scenarios, however, existing machine learning-based NDT approaches often rely on entire topology representations, which are inherently monolithic and lack reusability under topological or traffic changes in the network. This paper introduces a composable NDT approach that decomposes networks into subgraphs represented by reusable unit twins that capture subgraph structure, configuration and traffic behaviours. A lightweight composer aggregates unit twin combinations to create NDTs that predict per-route end-to-end latency through an overall topology. Evaluation across controlled synthetic topologies and diverse traffic scenarios, real-world Topology Zoo topologies, and a public NDT challenge dataset demonstrates that the composable NDTs achieve high in-distribution accuracy while remaining stable under out-of-distribution scenarios. Comparison with monolithic full topology NDTs demonstrates that our composable approach achieves reusability, while achieving comparable or superior accuracy.

Preparing high quality training data for foundation models requires scalable pipelines that transform heterogeneous documents and videos into structured records. Such pipelines expand each parent item into an ordered and input dependent sequence of children, whose counts may be long tailed. GPUs should batch children across parents while preserving parent relationships, child order, completion status, and result routing. Existing systems either hide parallelism behind coarse grained jobs or expose flat records that force applications to manage lineage and regrouping. We present RayOrch, a programming model and distributed execution engine that preserves parent child relations throughout execution. Programs declare ordered variable cardinality expansions and matching gathers. The compiler validates each pair, while the runtime records child membership, immediate parents, immutable ordinals, and terminal states. Per Call FIFO Ready Queues batch ready children across parents. Gathers reconstruct results from declared membership and ordinals rather than batch boundaries or completion order. Parents can advance as soon as all required children become terminal. Typed parent scoped failures suppress undispatched siblings of the failed parent while allowing unrelated parents to continue. On NVIDIA H20 GPUs, RayOrch achieves 15.14 times speedup when scaling MinerU from 4 to 64 GPUs and 7.82 times speedup when scaling a video pipeline from 8 to 64 GPUs. It reduces end to end time by 13.1 percent versus Ray Data and 29.0 percent versus Daft on MinerU, and by 16.0 percent versus Ray Data on Docling. Code available at https://github.com/OpenDCAI/RayOrch .

Real-time applications often rely on optimization approaches that can find high-quality solutions to hard problems on the order of milliseconds. Metaheuristic optimization frameworks (MOFs) are useful tools for such tasks, as they provide large sets of general-purpose search mechanisms that can return solutions under different computational budgets. However, existing work largely overlooks the available computation time as an explicit dimension of analysis. In this work, we introduce STILO, a MOF specifically designed for optimization under strict time limits. STILO integrates fine-grained configuration spaces for ant colony optimization (ACO), genetic algorithm (GA), and simulated annealing (SA), combining existing and novel operators. We performed experiments using both synthetic and benchmark instances of various discrete optimization problems. The results indicate that the proposed discrete distance calculation mechanism for SA is useful for optimization under strict time limits. They also show that the relative effectiveness of the proposed problem-independent graph structures for ACO can vary across time limits, even for the same instance characteristics. More generally, the results demonstrate that the effectiveness of algorithm families and operators depends not only on the problem type, but also on the characteristics of the instance and the available computational budget.

We analyze incomplete-information games where an oracle publicly shares information with players. One oracle dominates another if, in every game, it can match the set of equilibrium outcomes induced by the latter. Characterizations are provided for deterministic signaling functions, based on simultaneous posterior matching, a constructive partition criterion, and common knowledge components. This study elaborates on the work of Blackwell (1951) in games with incomplete information, using the common knowledge components of Aumann (1976).

This paper investigates multi-material 3D printing as a method to fabricate non-planar structures with 3D-printed electrode patterns for electroadhesion. We printed flat electroadhesion pads as a planar benchmark and cylindrical pads as a non-planar demonstration, with conductive interdigitated electrodes 3D printed as part of the structure. Normal-force measurements showed voltage-controlled modulation in both geometries. At 3 kV, the flat pads generated approximately 0.11-0.13 N, while the cylindrical pads reached approximately 0.16 N at 5 N preload. These results demonstrate a step toward printed parts with built-in, electrically controlled adhesion and friction, beyond conventional planar electroadhesion pads.

Based on synthetic polarization observations, we qualitatively investigate the intermittency of the magnetized interstellar medium (ISM) through synchrotron polarization intensity (SPI) and synchrotron polarization gradient (SPG). Three principal findings emerge from our analysis: (1) The intermittency of the magnetized ISM is strongly modulated by magnetization and compressibility, with pronounced modulation effects predominantly observed in regimes characterized by strong magnetization and high compressibility. (2) The characterized intermittency of the magnetized ISM demonstrates a dependence on the electron spectral index under conditions where the random magnetic field dominates. (3) SPG enables the identification of finer three-dimensional fractal structures in the case of strong Faraday depolarization. The synergistic application of SPI and SPG substantially enhances the reliability of probing magnetized ISM intermittency from real observations.

We study the stationary focusing nonlinear Schrödinger equation on the product $\mathbb{R}^N \times \mathcal{G}$ of the Euclidean space with a compact metric graph, in the mass-constrained variational setting. Such a product is a hybrid structure of a new type: all its faces are $(N+1)$-dimensional and are glued along interfaces of codimension one, so that the energy space is a genuine Sobolev space, while both the metric and the topology of the graph enter the variational problem. We first develop the functional framework, giving two equivalent descriptions of $H^1(\mathbb{R}^N \times \mathcal{G})$, introducing partial rearrangements in each of the two variables together with the corresponding Pólya–Szegő inequalities, and proving Gagliardo–Nirenberg inequalities in a localized form, with a comparison of the optimal constants with those of $\mathbb{R}^{N+1}$ and of the half-space. We then study the mass-constrained problem. Ground states exist for every mass when $2<p<2_*:=2+4/(N+1)$. At the critical exponent $p=2_*$, they exist below a graph-dependent threshold lying between one half of the Euclidean critical mass and the full Euclidean critical mass. The latter value is attained when the graph admits a cycle covering, whereas the threshold is exactly halved in the presence of a terminal edge. In both cases, the threshold is sharp. For $2_*<p<2+4/N$ global minimizers do not exist, but we prove the existence of local minimizers below a further mass threshold, for which we give an explicit lower bound. Finally, we describe the dimensional crossover: below a critical mass the minimizers do not depend on the graph variable, and we characterize the threshold below which the semi-trivial solution is a local minimizer in terms of the first nonzero eigenvalue of the Kirchhoff Laplacian on $\mathcal{G}$.

Historians have reconstructed the twentieth-century transformation of general relativity and gravitation (GRG) at the field level and through individual careers, but connecting these scales requires a way to compare researchers with the changing field over time. We develop such a comparison, setting a researcher's publications and references against GRG field literature from the same, earlier, and later two-year periods. Building on Own Vocabulary and Embedding Density Estimation from our earlier two-case study (arXiv:2501.00391), we extend the analysis to the fifty most-published authors in a NASA/ADS corpus of about 180,000 GRG records (1911 to 2000) and add two citation-based measures, Referenced Vocabulary and Citation Identity. The four measures compare an author's written language, cited literature, semantic neighbourhood, and cited-authority configuration with the surrounding field. The earlier cases suggested that closer field-vocabulary alignment accompanies a denser semantic neighbourhood. Across the fifty authors this holds only partially. Written and cited vocabularies tend to move together, usually resembling later GRG literature as the field turned towards astrophysical and cosmological research. Semantic neighbourhoods more often lie where the field's publications were concentrated in earlier periods, while co-citation patterns follow no single temporal direction, and the two citation measures frequently place the same researcher differently despite drawing on identical reference lists. Individual trajectories can thus combine vocabulary tied to later field states with older semantic or citation structures, and these divergent cases mark patterns for closer historical investigation. The approach transfers to other fields with defensible corpus boundaries and adequate coverage of texts, references, and disambiguated author identities.

The sensitivity of the $^{229m}$Th isomer transition frequency to the possible temporal variation of the fine-structure constant $\alpha$ is investigated theoretically. We evaluate both the Coulomb as well as the isomer energies in a selfconsistent Hartree-Fock plus Bardeen-Cooper-Schrieffer (BCS) approach with Skyrme energy density functional, taking a detailed account of the nuclear shape deformation and pairing correlations. Our results show that by a fine tuning of the pairing strengths and considering axial octupole deformation, the model can predict a very low isomer energy below the keV limit, which is not accessible in the presence of imposed reflection symmetry. Furthermore, in the model solution with octupole deformation, the difference between the Coulomb energy in the isomeric and ground states is one order of magnitude larger than for the case with imposed reflection symmetry. These results allow for a fully theoretical prediction of the sensitivity of the isomer transition frequency to $\alpha$ in the framework of a microscopic nuclear model, yielding values of $K\sim 10^3-10^4$, and for a deeper understanding of its underlying physical conditions.

The increasing demand for user privacy and compliance with regulations such as GDPR has made machine unlearning a fundamental requirement for modern recommender systems. However, Sequential Recommender Systems (SRS) pose unique challenges for unlearning due to their reliance on temporal interaction patterns. Existing approaches either require computationally prohibitive full retraining or fail to account for the sequential nature of user behavior. We propose SURF (Subtractive Updates for Recommender Forgetting), a lightweight framework for approximate machine unlearning in SRS. SURF operates in three stages: (i) identifying the neighborhood of the item to forget in the embedding space, (ii) training an auxiliary model on this compact local subset, and (iii) subtracting the auxiliary model's scores from the original model at inference time. Experiments against five baselines on 7 datasets show that SURF achieves unlearning effectiveness comparable to full retraining while substantially reducing computational cost, yielding up to a 32% improvement in NDCG@20 while requiring just 2% of the original retraining baseline time budget. We share our code at https://github.com/FilippoBetello/SURF.

I present the Spitzer Data Fusion, a multi-wavelength photometric database providing far-ultraviolet to far-infrared flux measurements, together with photometric and spectroscopic redshifts, for 4.4 million IRAC-selected sources over 65\,deg$^2$ in eight of the most widely studied extragalactic survey fields. A companion product, the SERVS Data Fusion, provides 2.8 million sources over 18\,deg$^2$ selected from deeper Spitzer warm-mission imaging. Catalogs are band-merged with wavelength-dependent matching radii, astrometrically registered against 2MASS, and distributed as FITS binary tables, one per field, as \href{https://doi.org/10.5281/zenodo.6120913}{DOI: 10.5281/zenodo.6120913} and as CDS/VizieR \href{https://cdsarc.cds.unistra.fr/viz-bin/cat/II/377}{catalog II/377}. The database is intended as a community resource for photometric redshift calibration, spectral energy distribution fitting, sample selection and multi-wavelength cross-identification, and is a natural bridge between the Spitzer legacy fields and ongoing Euclid, Rubin and SKA precursor surveys.

Baryon-to-meson ratios inside charged jets are studied in the Lund jet plane using PYTHIA simulations of pp collisions at $\sqrt{s} = 14$ TeV, with 15 < $p_{\rm T}^{\rm ch,jet}$ < 30 GeV/$c$ in an acceptance compatible with the future ALICE 3 detector. Jets are reclustered with the Cambridge–Aachen algorithm and identified hadrons are tagged at the splitting where they first separate from the leading branch. Three model configurations are compared: Monash, color reconnection beyond leading color (CR-BLC), and CR-BLC combined with thermodynamical string fragmentation (Thermal+CR-BLC). The $\Lambda_c^+/{\rm D}^0$ ratio rises with $\log(1/z)$ in the two configurations that include color-string junctions and is approximately flat under Monash. The same trend is present at the first Cambridge–Aachen branching, indicating that the effect is set at the level of the color topology rather than through the cumulative action of many soft splittings. Light- and strange-sector ratios show no $z$-dependent signature of either mechanism, but the angular projection reveals non-trivial structure at the first branching that is washed out when emissions are summed. The Cambridge–Aachen declustering and the Lund plane together separate the kinematic signatures of the two baryon-enhancement mechanisms in PYTHIA and provide observables accessible to future ALICE 3 measurements.

In a seminal work, Buhrman et al.\ (STOC 2014) introduced catalytic computation and proved that uniform $TC^1$ circuits are computable in catalytic logspace, the class of problems solvable in space $s$ with an additional catalytic tape of size $c$, a tape whose initial content must be restored at the end of the computation. A central ingredient of their proof is the register program model. Namely, they constructed a uniform family of register programs that computes $x^n$ using $n$ registers and four accesses to $x$. Since then, determining the number of registers and input accesses required to compute a polynomial of a given degree has become a central question in the study of catalytic computation. On one hand, we prove that the four-access bound of Buhrman et al.\ is optimal: every passive-output register program computing a polynomial of degree greater than three requires at least four input accesses, independently of the number of registers. On the other hand, we show that their register bound is not optimal. For every $t\geq2$ and every field $K$ of characteristic $0$ or greater than $2t-1$, we construct a register program for $x^{2t-1}$ with four input accesses and $t$ registers. Our proofs rely on derivations and their exponential operators. This approach represents a register program as a series of exponential derivation operators, reducing register restoration to an operator identity. Finally, we use the uniform family of register programs to improve known trade-offs for catalytic streaming algorithms and register programs for matrix powering. The generalization of the lower-bound methods and the construction of the uniform family of register programs were developed with assistance from ChatGPT 5.6.

We show that, starting from a product state, local unitary quantum evolutions generate intermediate states which exhibit a multiscale entanglement-spectrum structure required for universal entanglement embezzlement. This facilitates entanglement extraction from a catalyst many-body state while leaving it asymptotically unchanged. Remarkably, these atypical structures emerge generically at intermediate stages, well before reaching maximum entropy where thermalization has flattened out the spectral hierarchy. The resulting state is accompanied by nonlocal nonstabilizerness that diverges with the system size, consistent with a recently established equivalence between universal embezzlement and divergent nonlocal magic. Thus, without any fine tuning, a chaotic quantum evolution generates intermediate states which form a universal family of catalytic reservoirs.

Graphical User Interface (GUI) grounding is a fundamental perception task for multimodal agents, enabling them to interpret natural language instructions and interact with digital interfaces. Existing methods face a fundamental trade-off between accuracy and efficiency: direct full-image inference often fails to capture small or visually similar UI elements, while multi-crop strategies improve localization at the cost of multiple expensive Vision-Language Model (VLM) calls per query. To address this challenge, we propose RankGround, a two-stage framework that achieves accurate GUI grounding with a single VLM call per query. Central to our approach is GroundRanker, a lightweight multimodal reranker that identifies the most promising crop from a dense candidate set. Because no off-the-shelf ranking dataset is available, we construct ranking supervision data from existing grounding datasets. A strict containment criterion and boundary-aware positive augmentation improve alignment and spatial coverage in cluttered layouts. GroundRanker is then trained with a two-stage curriculum: a pointwise objective first learns coarse containment, and a listwise objective refines subtle semantic and spatial distinctions among visually similar crops. Experimental results show that RankGround consistently outperforms strong baselines while reducing computational cost. It achieves 1.4 times faster inference and improves localization accuracy by 5.5% on average over the second-best method across all backbones and screen scales, establishing a new state of the art in both efficiency and precision for GUI grounding.

We study the expected number of zeros of the Short-Time Fourier Transform (STFT) with a Gaussian window for signals degraded by complex Gaussian colored noise. We provide an exact formula for this quantity and investigate its asymptotic behavior. From a computational perspective, we propose an algorithmic method to recover a smoothed power spectral density (PSD) directly from the spatial distribution of the transform's zeros. Our results formalize the commonly-held heuristic that detection algorithms based on spectrogram zeros, though designed for white noise, also perform adequately under moderately colored noise.

AI models for multimodal medical imaging must balance modality-specific specialization with cross-modal shared representations, a trade-off that pure Mixture-of-Experts (MoE) architectures currently fail to satisfy. Expert-based routing improves in-domain learning but may sacrifice cross-modal signals, which appear particularly important for rare (low-prevalence) pathologies in our experiments. To resolve this, we introduce Generalist-Specialist-MoE (GS-MoE), a two-branch (MoE) architecture that couples a cross-modal generalist model with distinct modality-specific specialists (experts) via domain-constrained feature fusion. On RadImageNet (1.35M images, 165 pathologies, three modalities), GS-MoE recovers detection of six low-prevalence pathologies on which every baseline scores F1 $=$ 0, with per-class gains up to +0.60 F1. It attains this while even slightly exceeding dense and specialist-only MoE aggregate baselines (MCC 0.770), while using ${\sim}53\%$ fewer active parameters at inference than the strongest investigated dense model.

We show that a compact mean curvature flow whose singularities have multiplicity-one cylindrical tangent flows admits a smooth Morse resolution. The resolution agrees with the spacetime track outside any prescribed neighborhood of its singular set and all its critical points have the expected index. No nondegeneracy or isolatedness assumption is required. In the mean-convex case the result follows by smoothing the global arrival function.

Self-supervised learning (SSL) has improved speech representations, yet performance degrades in pathological domains such as electrolaryngeal (EL) speech, and the computational footprint of SSL models limits their applicability in real-time, on-device deployment. We propose a multi-teacher knowledge distillation framework to train a lightweight, streaming content encoder that generalizes across healthy (HE) and EL speech. Two teachers are distilled progressively: a frozen SSL model providing discrete phonetic cluster targets from HE speech, and an EL-fine-tuned speech recognition model supplying continuous bottleneck feature targets. Evaluated via downstream speech recognition, our approach reduces the EL word error rate to 21.2%, compared to 39.3% for the strongest zero-shot SSL baseline. Among causal convolutional, Transformer, Conformer, and Mamba-based student architectures, a Mel-Conformer achieves the best combination of EL accuracy and computational efficiency. The final encoder contains 21.9,M parameters and runs at a real-time factor of 0.30 under ONNX Runtime on a single CPU core.

Reinforcement learning allows robots to acquire complex skills, but producing policies for geometrically complex manipulation remains difficult. A promising approach is to learn on top of collision-avoidant controllers, such as geometric fabrics. However, these approaches have relied on static, hand-specified representations of the scene. Integrating active, online 3D perception into massively parallel RL training has so far been inaccessible. We introduce a GPU-accelerated method that reconstructs the scene as a collection of surfels across thousands of parallel simulation instances during active rollouts. This lets policies operate over sensor-derived, rather than hand-specified, geometry. On a suite of collision-dense manipulation tasks, our surfel fabrics enable policies to tackle geometrically complex scenes where primitive-based baselines fail, while maintaining sim-to-real transfer. Furthermore, policies learned with a scene-aware fabric are more robust to the introduction of novel geometry at test time, improving collision-free task completion under unseen obstacles from 35% to 61%. We release our reconstruction system, training code and test dataset to spur research in this direction.

Crises and population responses can form coupled dynamical systems, with crisis conditions shaping protective behaviours and collective responses altering the crisis trajectory. Existing models rarely capture both evolving crisis conditions and the reinforcement-dependent spread of competing behaviours. We propose a complex contagion with feedback (CCF) model that couples competing complex contagion with crisis dynamics through bidirectional feedback. Agents stochastically switch between competing states according to social reinforcement and adoption complexity, which is influenced by crisis conditions and information campaigns, while population-level behavioural change feeds back into the crisis. We formulate and analyse the model for a well-mixed population, characterising its equilibria and associated stability conditions. As a case study for model validation, we integrate CCF model with a census-calibrated agent-based model of COVID-19 transmission in Australia to represent social distancing adoption and discontinuation. We compare the CCF model with an existing opinion dynamics model and find that, despite using fewer parameters, it achieves comparable performance in reproducing recurrent infection waves. The resulting framework is parsimonious, analytically tractable, and can be adapted to different types of crises.

We assemble a finite-dimensional \(\mathcal{VU}\)-calculus for composite nonsmooth functions whose outer function is convex: the chain rule, separable and convex sums, a strictly differentiable perturbation of a lower semicontinuous (lsc) term, the model \(\delta_X+f_0+\theta\circ F\), and a finite maximum of \(C^1\) functions. The same algebra yields an \(\varepsilon\)-\(\mathcal{VU}\) chain rule for a proper outer approximation of the subdifferential. On the set \(\mathcal{R}_{h,F}\) of points at which the convex chain rule holds, the subspaces \(\mathcal{V}f\) and \(\mathcal{U}f\), an orthonormal frame of \(\mathcal{U}f\), and the \(\mathcal{U}\)-gradient \(\bar g_u\) are written in terms of the factors. If \(f\) is \(C^1\)-partly smooth at \(\bar{x}\), that gradient is \(\bar g_u=U_f^\top\nabla_{\mathcal{M}}f(\bar{x})\). If \(f\) is \(C^2\)-partly smooth and \(0\in\ri\partial f(\bar{x})\), the convex \(\mathcal{U}\)-Hessian \(H_U\) (when \(f\) is convex) or the local matrix \(H_\varepsilon\) (when \(f\) is prox-regular at \(\bar{x}\) for \(0\)) equals the Gram matrix \(U_f^\top\nabla^2_{\mathcal{M}}f(\bar{x})\,U_f\); the same matrix, written along the active manifold in a continuous frame, depends continuously on the base point. If in addition \(f\) is \(C^2\)-partly smooth at \(\bar{x}\), then under prox-regularity and subdifferential continuity at \(\bar{x}\) for \(0\), \(H_\varepsilon\succ 0\) is equivalent to tilt stability of \(\bar{x}\) and to strong metric regularity of \(\partial f\) at \((\bar{x},0)\), with \(\lip\bigl((\partial f)^{-1}\bigr)(0\mid\bar{x})=\|H_\varepsilon^{-1}\|\). The calculus and the test are illustrated on a hinge composite and two elementary tests of \(H_U\succ 0\).

This paper provides a self-contained proof of the rank of the pathlifting Jacobian of a DAG ReLU network by performing an induction on the network's number of hidden nodes. In fact, the induction is elementary, and the key recipe is to consider the skeleton matrix of the network, a sparse matrix encoding the network paths, and transform the representation of one of its hidden neurons into an output node. The proof relies on intermediate propositions which link the pathlifting, its Jacobian, the network parameters, and its skeleton matrix, which, on top of permitting to conclude on the rank of the pathlifting Jacobian, also provide a way to compute it without backpropagation and whose computation cost is super efficient in practice compare to usual backpropagation. The paper is provided with a Python module that implements the different propositions of the paper for feed forward networks and is used to experimentally quantifies the computational gain of computing the pathlifting Jacobian with the proposed theory.

Fu, Guil Asensio, Herzog and Torrecillas asked whether a complete ideal cotorsion pair of object ideals in an exact category always induces a complete cotorsion pair of objects. We show that it need not, even in a Hom-finite, weakly idempotent complete Frobenius exact category. Our example consists of bounded complexes of finite-dimensional vector spaces with even total cohomology dimension. We characterize special object approximations by parity and identify the obstruction as a non-split idempotent in the stable category. For cohereditary pairs, a criterion of Saorín and Šťovíček gives an affirmative answer when the stable category is idempotent complete.

Contextual ASR can benefit from semantic cues or from target words explicitly provided in the context. We introduce HearInContext, a Mandarin-English benchmark that pairs shared synthetic speech with assistant replies supporting different interpretations. The benchmark comprises 3,764 semantic test cases built around homophones. Implicit contexts exclude candidate words; explicit contexts name the target. No-context and unrelated-context controls measure the benefit of relevant history and sensitivity to irrelevant history. Context-capable models benefit from implicit cues but achieve higher target recall with explicit hints. Fine-tuning Qwen3-ASR-1.7B improves implicit-context target recall by 11.4 percentage points in both Mandarin and English, while absolute CER/WER changes on AISHELL-1 and LibriSpeech remain below 0.1 percentage points. Gains extend to explicit conditions excluded from fine-tuning and to Mandarin hotword recognition on real recordings. Code and data are available at https://github.com/OPPO-Mente-Lab/HearInContext

In interactions with uncertain opponents, dual model predictive control (MPC) can improve performance through information-seeking actions that reduce uncertainty about opponents' behavior. Its recent applications to autonomous driving, however, are limited to scenarios involving a single opponent on a single lane. This paper presents a mixed-integer dual MPC for multiple reactive opponents on multi-lane roads, jointly optimizing integer-valued maneuver decisions (lane changes and safe-region selections), and continuous motion over a scenario tree that samples plausible interactions with the opponents. As interaction complexity increases, solving the resulting mixed-integer nonlinear program becomes increasingly expensive. To reduce this computational burden, a graph neural network (GNN) predicts the optimal maneuver decisions, and high-confidence predictions are fixed before the reduced problem is solved. Simulations show that active probing behavior emerges in complex interactive scenarios, and that GNN guidance fixes $76.3\%$ of the integer decisions and reduces the solve time by $2.5\times$ on average, with negligible degradation of optimality.

We consider the problem of noisy gradient-free minimization of the k-th order partial derivative of a $\beta$-Hölder function supported on a d-dimensional cube. We show that T ^{($\beta$+d+k)/(2$\beta$+d)} log(T )^{(\beta-k)/(2\beta+d)} is a non-asymptotic minimax rate of the T step cumulative regret for all $\beta$ \ge 0. In the special case k = 0, our results cover the problem of noisy gradient-free minimization of $\beta$-Hölder functions, closing the existing gap between the known upper and lower bounds. We show that a minimizer of a suitably chosen local polynomial estimator is rate-optimal. The minimax optimal upper bound is achieved under the passive design, that is, when the query points are i.i.d. Thus, there is no advantage in considering sequential designs when it is only known that f is a $\beta$-Hölder function with no additional property. We propose an algorithm feasible in polynomial time that constructs a proxy of the minimizer of the local polynomial estimator. The procedure requires computing the estimator on auxiliary random points. The resulting polynomial time algorithm matches the lower bound.

Speech language models enable richer spoken interactions between humans and machines than cascaded systems, allowing access to paralinguistic information and lower latency. However, their accuracy on mathematical reasoning benchmarks has lagged behind those of text models. Reinforcement learning (RL) with verifiable rewards has been instrumental in extending text models' capabilities for solving complex problems and limiting hallucinations. In this work, we explore applying RL to the GLM-4-Voice speech model (Zeng et al., 2024) to bridge the gap between textual and spoken mathematical problem solving. We first adapt the model to the domain using supervised fine-tuning on synthesized spoken question-answering data. We then show that, even without extra reasoning tokens, RL improves the accuracy on GSM8K beyond levels previously achieved for speech models only with supplementary reasoning traces. When combined with existing streaming reasoning techniques, we show further gains to 74.8% free-form accuracy. This establishes a new state-of-the-art for mathematical spoken abilities with speech-native models.

Generative artificial intelligence (GenAI) is changing how students learn, yet the roles of course context, cognitive reliance, evaluation literacy, and early reliance remain underexplored. Using survey responses from 118 students across 12 AI-related courses at our institution, we examined differences in GenAI use and perceived learning experiences. We identified four user clusters: high-use students reporting many benefits, light users reporting less reliance and fewer benefits, and two moderate-use groups reporting different levels of benefit. We also found significant differences between free- and premium-version users, single- and multiple-tool users, and students experiencing different instructor policies. In multivariable regression models, academic benefit was associated with early reliance and academic task support; positive impact was associated with cognitive reliance, academic task support, confidence in GenAI reliability, and instructor policy; and negative impact was associated with early reliance and attitudinal change. The association between early reliance and negative impact became stronger as evaluation literacy increased. Finally, perceptions of GenAI-enhanced learning appear to reflect cognitive, performance, and self-efficacy benefits, while concerns about stress and diminished critical thinking are associated with lower perceived learning benefits. These findings suggest that institutions need better policies to address such inequities so that institutions can enable students to benefit from increasingly capable AI systems.

Large language models (LLMs) require increasing memory capacity to accommodate growing model weights and KV caches. High-Bandwidth Flash (HBF) offers high memory density and aggregate read bandwidth through massive plane-level parallelism, making it an attractive option for LLM serving. However, serving LLMs entirely from HBF introduces three challenges: fine-grained KV reads create placement and access imbalance, incremental writes interfere with foreground reads, and mixed KV lifetimes amplify garbage collection. Hybrid HBM/HBF designs retain HBM to support dynamic KV management, but this allocation reduces the HBF resources available under a fixed packaging budget, limiting aggregate HBF bandwidth. We present HBFlex, a full-HBF memory system with coordinated optimizations for KV reads, writes, and reclamation. HBFlex balances KV placement and attention accesses to improve plane utilization. It aggregates incremental updates and schedules writeback within sufficiently long compute windows to reduce write–read interference. It also combines lifetime-guided block packing with deferred reclamation to reduce valid-page migration. We evaluate HBFlex through trace-driven simulation across different configurations. HBFlex achieves average throughput speedups of up to 1.58$\times$ over FlashAccel and 3.30$\times$ over H3, benefiting from higher HBF bandwidth and more efficient management of dynamic KV-cache reads, writes, and erases.