Papers

Reinforcement learning with verifiable rewards (RLVR) is deployed to make models better at reasoning tasks, but its side effect on what models will divulge is under studied. Here we show that RLVR on facts increases extraction of personally identifiable information (PII) the instruct model had already memorized. We first confirm that instruct models have already memorized PII but leave them latent, rarely surfacing one when asked. We then apply RL on benign factual data that contains no PII of any kind, and re-probe: a targeted probe over name->email pairs, and an untargeted free-recall prompt that simply asks the model to list the addresses it knows. PII extraction rises sharply under both: on DeepSeek-V3.1, verbatim recall@k increases from 0.155 to 0.370, a 2.4x gain. The effect scales with model size: across three models spanning 8B to 671B parameters, absolute leakage is largest in the biggest model. Meanwhile model's reasoning abilities and refusal rates are retained, indicating that RL selectively changes which memorized information is accessible rather than broadly altering the model. In summary, memorized private data can be made markedly more extractable by training that never touches it. This gives an adversary a route to memorized data that requires no privacy-relevant training signal and no access to the data itself – only the ability to fine-tune on something innocuous.

Reasoning-augmented text-to-image models such as GoT-R1 emit an explicit textual plan - object names, attributes, and bounding boxes - before generating image tokens. When such a model fails a compositional prompt, is the plan wrong, or is the plan right and the decoder unfaithful? Because the plan is machine-readable it can be edited before decoding, which makes the two separable. We first validate the ruler. Swapping the two bounding boxes inside the model's own chain demonstrably flips the generated layout: detector-based accuracy falls 0.75 -> 0.48 (p<1e-3), while a widely used VQA-based spatial metric rises. A five-rater human study agrees with the detector on 81% of items and with the VQA judge on 57%. All spatial results therefore use geometric scoring. Under sound measurement the decoder is a faithful executor: 94% of generated layouts realize the planned relation, and object-box binding survives reordering of the plan's object segments. The planner is the bottleneck. It writes wrong relations for phrasing-dependent reasons - 98% accuracy on "left" against 54% on "right" for semantically identical layouts, a raster-order bias we isolate with a mention-order control - and cluttered geometry that the decoder faithfully reproduces. Editing the plan therefore fixes the image without retraining: symbolic verification with resampling gives +5.0 points (p<1e-3), minimal in-place repair +6.0 (p=.02), rewriting only box geometry +10.7 (p<1e-4), and replacing the plan outright +13.3 (p=1e-4). Gains are indifferent to the plan's prose style and to its likelihood under the planner, but not to its geometry. Modular planner-decoder designs are therefore viable, provided the plan is internally consistent: box-text contradictions induce object duplication and identity fusion. We release the plan-fidelity evaluation protocol, all plans, and 12k generated images.

Diffusion models are widely used as priors for linear inverse problems, yet endpoint quality does not reveal when measurement information enters reverse denoising or how it is allocated across signal directions. We study this process through the smoothed likelihood force, the difference between exact posterior and prior scores at each noise level. For a fixed measurement, its expected squared norm gives both posterior–prior relative-entropy dissipation and reverse-path relative-entropy growth. Averaging over measurements yields an information–minimum mean-square error (I-MMSE) identity linking information gain to denoising-error reduction. Under finite second moments, the force energy and its ratio to prior-score energy decay quadratically in the noising kernel's signal coefficient at high noise. Solvable models show that conditioning removes class separation already explained by the measurement, reduces a uniform index entropy over \(n\) empirical samples from \(\log n\) to \(H(I\mid r)\), and makes assimilation depend on operator–prior alignment even for identical singular values. Experiments in models with tractable posteriors evaluate these predictions. In a separate illustration with a frozen FFHQ model, masks sharing the same spectrum yield different prior-normalized null-space trajectory statistics.

Continuous-time flow and diffusion models are widely used across many application domains, from large-scale deployment in computer vision and protein folding to emerging adoption for modeling language, time series, and quantum states. After training, inferring statistical properties from continuous-time models is costly. Wavefunction flows target this cost by recasting learned transport as unitary evolution, whose final Born distribution approximates the target distribution. This prepares a coherent amplitude encoding (a qsample) that can be post-processed by quantum algorithms offering a quadratic advantage over Monte Carlo sampling. We present the first numerical study of these flows, in which we represent time-dependent potentials and states as tensor networks. At spatial dimension $d=8$, storage falls by $\sim 10^7\times$ relative to the dense grid of $N^d$ points, and evolution wall-clock time falls by $\gtrsim 10^3\times$ against a baseline extrapolated from the measured $d\le 5$ scaling. We validate our pipeline by reproducing the $O(1/\sqrt{p_{\rm rare}})$ scaling of rare-event sampling.

Mixture-of-Experts (MoE) LLMs scale model capacity efficiently through sparse activation, but their large expert parameter footprint, routing imbalance, and long-context KV-cache growth make deployment difficult on commodity hardware. Practical deployment often requires stacking multiple compression techniques: expert pruning removes redundant experts, weight quantization lowers model memory footprint, and KV-cache compression reduces long-context memory pressure. However, these techniques are typically evaluated in isolation, leaving open how they interact when applied together in realistic deployment pipelines. In this work, we present MoEXBench, a systematic benchmark for evaluating composable MoE compression as an end-to-end deployment workflow. MoEXBench studies 10 MoE models ranging from 30B to 235B total parameters across standard-attention, hybrid linear-attention, and sliding window attention architectures. It evaluates 20%-50% expert pruning rates, 1 to 16 bit weight-quantization schemes, and multiple KV-cache precision settings, applied both individually and in combination. MoEXBench introduces an eight-module evaluation suite that jointly measures composable-compression quality, workload and architecture robustness, pruning/quantization/KV cache sensitivity, and deployment efficiency on commodity hardware. Our results reveal non-trivial interactions among compression methods: composable compression cannot be predicted from standalone techniques, compression rate alone does not reliably predict quality loss or runtime gain, expert pruning is the dominant degradation source, and average quality can hide workload and architecture-specific failures. By releasing normalized module scores, compressed artifacts, and reproducible scripts, MoEXBench enables practical accuracy-memory-latency comparison across MoE families and hardware backends.

Joint-Embedding Predictive Architectures (JEPAs) learn world models by predicting future representations rather than reconstructing observations. Variational JEPA (VJEPA) gives this prediction probabilistic semantics and permits flexible conditional predictive families. We propose MDN-VJEPA, a multimodal probabilistic JEPA that instantiates the latent predictor as a mixture density network (MDN), representing future states with a Gaussian mixture whose components encode alternative futures, their probabilities, and within-mode uncertainty. For sequential prediction, we address two additional structural requirements: multi-modal futures should remain temporally coherent across a trajectory, and predictive beliefs should remain consistent when propagated to the same horizon through different temporal path compositions, as motivated by the Chapman–Kolmogorov equation. Together, these considerations define a three-level hierarchy of multimodal prediction, coherent multimodal trajectories, and compositionally consistent multimodal dynamics. Controlled experiments show that MDN-VJEPA closely recovers known branching predictive laws, that retaining the complete multimodal belief rather than its conditional mean can lead to materially different and substantially better planning decisions under branching futures, and that trajectory-level coupling and distributional consistency improve temporal coherence and multi-horizon agreement. These results support a simple principle: a probabilistic JEPA should preserve not only how uncertain the future is, but which distinct futures constitute that uncertainty, how they form coherent trajectories, and how those beliefs compose through time.

Representation quality is a central determinant of PINNs' performance, yet standard training leaves representations to emerge implicitly while fitting the final solution. We introduce representation curriculum, an ordered process in which representations are explicitly learned, transferred independently of their predictors, and progressively refined. We realize it with Staged Depth Training (SDT), which trains a shallow prefix under a temporary physics-informed head, discards the head, and freezes the learned prefix while adding depth, without equation-specific encodings or changes to the final architecture. Across the 20 default forward problems in PINNacle with three backbones, SDT improves 40 of 59 equal-budget problem–backbone cells by at least 5% and remains within that band in the rest, with a 32.8% geometric-mean error reduction on a PirateNet-style backbone. Mechanistic ablations suggest that the gain is not explained by optimizer restarts or shallow warm-starting alone. Representation visualizations and hyperparameter-basin analyses provide diagnostic evidence on representation geometry and local sensitivity to shared hyperparameters. On Poisson–Boltzmann 2D, SDT also more than doubles the fitted depth-scaling exponent for both backbones. These results support representation curriculum as a promising training strategy for improving PINNs while preserving the deployed architecture and inference cost.

Neural combinatorial optimization solvers generate many candidate solutions per instance and report the best one found, using the same sample budget for every instance regardless of difficulty. A companion study showed that reallocating a fixed budget toward harder instances can improve solution quality, but that the standard way of measuring this improvement is biased: deciding an allocation and evaluating it on the same data can manufacture an apparent gain even when none exists. This left open what property of a workload determines whether reallocation is worth doing, and whether a policy that spends part of the budget to decide how to allocate the rest still pays once that cost is counted. This paper answers both questions through pre-registered confirmatory experiments – analysis and verdict criteria fixed before data collection – across three independently trained solvers and two ways of constructing harder workloads on the traveling salesman problem. Within the workloads we study, the deciding property is how varied the instances within a workload are in difficulty, not how difficult the workload is on average: a uniformly easy or uniformly hard workload offers little room for reallocation, while a mixed workload offers substantial room. A budget-aware policy that pays for its own information about instance difficulty recovers most, though not all, of the improvement available when that information is assumed free. Every experiment was independently recomputed from its written specification, and every correction to an earlier version – including two that weakened the paper's own claims – is reported with the direction it moved the conclusion. The paper offers a specific empirical answer and a template for verifying that answer is not an artifact of how it was measured.

We study the problem of aggregating opinions from multiple black-box experts in noisy, conflict-prone settings where expert reliability varies across inputs. Static aggregation methods, such as majority voting, fail to capture this variability and often yield unreliable outcomes under disagreement. We propose a tractable, probabilistic-circuit-based fusion framework that dynamically combines expert responses using context-specific credibility estimates, enabling principled and reliable reasoning. The framework is agnostic to the underlying experts and does not require access to their internal representations or any retraining. We empirically validate our approach on multiple-choice question answering tasks using multiple LLMs as experts, comparing against individual models and static ensemble baselines. Our method consistently improves predictive performance and produces more reliable decisions under conflict, highlighting the effectiveness of context-aware credibility modeling for robust multi-expert fusion.

Global forecasting models pool many series and learn one shared function. Gradient-boosted trees are their most common form. We measure a failure of this design that has not, to our knowledge, been documented. Train a global tree on the individual series of a hierarchy, then ask it for the hierarchical aggregate. The aggregate sits far outside the model's training range, and the forecast collapses. The model under-predicts the total by 30-50x in our production deployment, and by up to 496x in a public M5 reconstruction. The mechanism is known: beyond its training range, a tree predicts a constant. It surfaces at the aggregate because the total dwarfs every training series. The cure is not new. Per-series scaling, the preprocessing step that Montero-Manso and Hyndman (2021) recommend, prevents the collapse. So do a weighted aggregate-level training row and seasonal differencing. Our contribution is the characterization. The collapse reproduces on five panels: a production business-to-business marketplace, a synthetic hierarchy, M5, Australian Tourism, and a public business-buyer panel. It holds on three tree libraries, is invariant across training seeds, and is statistically significant. Its onset is immediate and tracks a simple support bound: a scale gap of only 1.15x already costs a third of the total. No standard configuration change prevents it: pooling every hierarchy level into training fails at scale, and the one knob that fits linear models in the leaves softens it without curing it. Rolling the forecasts forward recursively separates the cures: the aggregate-row cure re-collapses, per-series scaling degrades but stays low, and only seasonal differencing keeps its one-step accuracy unchanged. We close with a three-step procedure for diagnosing and preventing the failure in deployed systems.

Sharing a deep encoder does not, by itself, fix the central confound of task-comparison scores. We show that cross-evaluated heads on a frozen shared representation inherit the extrapolation confound of shallow exchange scores: pure input rotations with fixed labels inflate a deep exchange score from about 0 to 0.80, while representation-novelty scores are blind in the complementary direction (flat under label permutations that change the task completely). Transplanting a conditional two-discriminator discrepancy into the embedding space resolves both blind spots: the functional axis stays within +-0.001 under rotations and tracks label-permutation drift mass monotonically. Built into a mixture-of-heads lifecycle, the two-axis gate attains better decision quality with fewer heads than exchange or novelty triggers at a matched training budget. On generalized category discovery, the same chunk-level functional axis separates semantic novelty from photometric shift with AUROC 0.98-0.99 where per-input OOD scores (MSP, Energy, Mahalanobis, KNN) sit near chance for that distinction. All findings replicate across frozen ImageNet-21k ViT-B/16 and self-supervised DINOv2 backbones on CIFAR-100, and extend to residual adapter pools with recurrence, where a null-calibrated novelty trigger never fires on mechanism changes while the two-axis gate handles them with full recurrence reuse. We state explicitly the common-factoring condition under which embedding-space conclusions transfer to the original mechanism.

High-resolution surface solar radiation (SSR) is important for solar forecasting and grid operation. However, physically consistent reanalysis products are too coarse to resolve localized cloud-driven variability. In this paper, we study a multisource downscaling task that reconstructs high-resolution SolarCube SSR fields from coarse ERA5 radiative variables and co-registered satellite channels. The task is challenging because a single ERA5 grid cell may contain both sunlit and cloud-shadowed regions. As a result, the missing high-resolution correction can be spatially sharp and inherently ambiguous. One-stage predictors often oversmooth these structures. Post-hoc refinement also introduces a stage-wise mismatch: the generator is optimized independently, even though its output determines the refiner's initial state. We introduce SolarFlowRefiner, a refinement-aware flow-matching framework for SSR downscaling. A conditional FlowMatch generator first predicts a normalized correction to an upsampled ERA5 baseline. The refiner is then trained on prediction-conditioned states between the current FlowMatch output and the target residual. This exposes the refiner to the structured errors produced by the generator. The refinement objective is also backpropagated through the FlowMatch sampler, allowing generation and correction to be jointly optimized for the final reconstruction. Experiments on a day-blocked ERA5–SolarCube benchmark show consistent improvements over standalone generation and post-hoc refinement. More broadly, SolarFlowRefiner provides a general strategy for coupling generative predictors with iterative correctors.

We present a projected sensitivity study for hadronic mono-$Z$ dark-matter production using CMS Run~2015D HTMHT open data corresponding to 2.256382381~\invfb, from which 1{,}439{,}523 events satisfy the hadronic mono-$Z$ selection. Backgrounds are modelled with a conditional flow-matching continuous normalizing flow trained on the selected HTMHT events and evaluated on a held-out validation split reweighted to the full selected population. To mitigate artifacts from missing-object features and avoid in-sample scoring bias we apply sentinel imputation for undefined angular features, persist the train/validation split indices, and enforce a minimum reported background yield of 20 events when selecting the working point. A signal-side offline trigger proxy is applied to the simulated signal before scoring. Under this procedure the baseline analysis yields expected significances of 2.89$\sigma$, 7.62$\sigma$, and 7.41$\sigma$ for three simplified-model benchmarks. An ablation study that removes the detailed extra-jet kinematics reduces the expected significance by 53–71%, indicating that extra-jet topology carries substantial discriminating power in the hadronic mono-$Z$ channel. These results are projected sensitivities (no unblinding performed); the limitations and reproducibility of the study are discussed in Sections limitations and reproducibility.

A composite structural index summarises a network in one number, and for a triangle-based index it is spectrally redundant: Tr(A^3) is the third moment of the adjacency spectrum. The non-redundant content sits one level down, in diag(A^3), which depends on eigenvectors and is not spectrally determined. A corollary in the theory paper predicted that the global scalar should tie sharpened spectral baselines rather than beat them, while the node-wise attribution should do better where the number of structural epicentres is unknown. We construct Omega-N by localizing each of the four factors. The direct localization is badly conditioned; two corrections from published practice fix it, a configuration-null excess per factor and a personalized-PageRank neighbourhood at several scales, giving ten interpretable features per node, with no attributes, training or embeddings. Against a recursive feature engine at five levels of recursion, Omega-N wins on three and ties on two of the six in-domain evaluations, the sixth a declared null where every arm returns chance, with ten features against its 28 to 252 before pruning. Two statistics from the graph and labels, not from performance, partition the eight benchmarks without error, and the two they exclude are the two on which it loses. The strongest application is drug-target prioritisation on protein interaction networks: +0.032 to +0.103 AUPRC over a six-feature centrality battery and +0.084 to +0.208 over the four-feature one, across three constructions, replicated on an independent AP-MS network and label source (degree-matched: +0.0723 on STRING, +0.0560 on BioPlex, p=0.00195). Adding Omega-N to centralities plus Node2Vec changes nothing. The claim is narrow and it is the point: ten named features, computed without training, match or beat hand-crafted centralities and a recursive engine, and do not touch learned representations.

Test-time scaling (TTS) improves language model reasoning by allocating additional test-time compute, but prior work mainly studies closed-ended tasks such as math and coding. We study TTS for automated equation discovery, an open-ended setting where models search over candidate equations and rely on observed datapoints for feedback. We formulate LLM-driven equation discovery as an iterative search process that unifies Best-of-N, sequential refinement, tree search, and evolution-style methods under a common compute-allocation view. To isolate allocation effects from prompt engineering and other heuristics, we compare minimal parallel controllers under fixed budgets. On LLM-SRBench equation-discovery tasks, we find that search width is the dominant allocation parameter: the best width in our sweep generally increases with the compute budget, while the population–branching split and controller choice matter less. Appropriate width selection also improves wall-clock efficiency by increasing parallelism. These results suggest that, given an informative verifier, controlling exploration and exploitation is central to scaling LLM-based equation discovery.

Vehicular traffic is a major source of air pollution; however, the contribution of remotely acquired traffic information to local machine-learning (ML) air-pollution models remains insufficiently characterised. This study evaluates four interpretable tree-based ML models (Random Forest, Extra Trees, LightGBM, and XGBoost) under six predictor scenarios combining progressively larger predictor sets, ranging from remotely acquired traffic, meteorological, and temporal variables alone to the inclusion of measurements from one and four neighbouring monitoring stations, to estimate NO$_2$, PM$_{10}$, PM$_{2.5}$, and O$_3$ concentrations across several sites in London. ML model performance was compared with a ridge linear regression model as a baseline, with spatial interpolation methods and with a cross-site validation experiment. When modelling without data from neighbouring stations, the RMSE for NO$_2$ ranged from 9.73 to 11.66 $\mu$g/m$^3$ without traffic information, compared with 8.72 to 11.52 $\mu$g/m$^3$ when traffic information was included. Additionally, for NO$_2$, SHAP analyses indicate that traffic-related variables can contribute at levels comparable to pollutant measurements from neighbouring monitoring stations in traffic-dominated~environments.

Current work in modeling symbolic music primarily relies on representations extracted from MIDI-like data. While such formats allow for modeling symbolic music as sequences of notes, they omit the large space of symbolic annotations common in western sheet music broadly known as expression text, such as tempo or dynamics, which specify time- and velocity-dependent controls on the musical composition and performance. To alleviate this gap, we present MusPyExpress, an extension to the popular symbolic music processing library MusPy that enables the extraction of expression text along with symbolic music for downstream modeling. Utilizing this extension, we parse the PDMX dataset to illustrate the wealth of expression text available in MusicXML datasets. Additionally, we introduce multiple generative tasks, including joint expression-note generation, expression-conditioned music generation, and expression tagging, that take advantage of this additional notational information.

Recent mesh-based simulation advances have, in no small part, relied on neural surrogates of two distinct families: global models that route information through a small set of latent tokens, and local models that perform message passing across mesh edges. Consistent with both classes is the inability to perform beyond low-dimensional problems and small-scale or oversimplified meshes, the simulation regimes where industrial problems reside. Our work shows this explicitly and presents a unified formulation. In global approaches, latent-token attention acts as a spatial low-pass filter, while local message passing lacks the global reach necessary to propagate information across large mesh spaces. Viewed through the error, the two operators are the halves of a multigrid cycle: one corrects errors at the lower end of the spectrum, the other at the higher end, and neither can do the other's job. We introduce Read-Write-Relax (RWR), which interleaves latent attention with message-passing relaxation under a unified formulation. The interleaved processor lowers error across the entire spectrum, making RWR the most accurate model in nearly every comparison across our industrial and public benchmarks. It is also markedly data-efficient in the scarce-data regimes, accurate on the engineering quantities of interest, and scales full-field predictions to challenging, large-scale problems.

Longitudinal electronic health records (EHRs) document patients' sequences of clinical visits over time, preserving the temporal evolution of disease progression and care delivery. However, real longitudinal EHRs are difficult to access because they contain large amounts of fine-grained, patient-specific information. Synthetic EHR generation therefore provides a valuable approach for preserving the statistical patterns and clinical structure of patient visit trajectories, enabling broader modeling and analysis when real records are limited. Although recent generative models have made progress in producing future visit sequences, they remain limited in explicitly integrating inter-visit irregular temporal evolution and intra-visit clinical event structures in EHRs, leading to clinically inconsistent and temporally unrealistic visit sequences. In this work, we propose SynEHR, a lightweight adaptive LLM-based framework for longitudinal EHR synthesis. There are two novel designs in SynEHR, i.e., a Temporal State Conditioning Module captures irregular temporal states across visits and a Temporal-Relational Adaptation Module combines these states with patient history to dynamically construct patient-specific relational representations. SynEHR then builds on a parameter-efficient LoRA-adapted language-model generator with next-visit generation capability to train the two modules for temporally and clinically informed generation. Extensive experiments on real-world EHR datasets across fidelity, privacy, and downstream utility evaluations demonstrate that SynEHR outperforms state-of-the-art models by generating more clinically coherent and temporally faithful longitudinal EHR data.

When discrete-time optimizers operate at the edge of stability, they exhibit near-two-periodic behavior. These oscillatory dynamics are reminiscent of conservative systems, such as the dynamics generated by symplectic integrators. However, a precise formulation of the connection between discrete-time optimizers at the edge of stability and discrete mechanics remains underexplored. Recently, Litman introduced the "edge coupling": a functional on consecutive gradient descent iterates whose critical points encode the fixed points and two-point orbits of the gradient descent dynamics. Here we extend the edge coupling to heavy-ball and Nesterov momentum. We show that its critical points characterize the fixed points and two-point orbits, with its Hessian characterizing their stability. We also show that the edge coupling can be identified with the symmetric Verlet action, formalizing the connection between the edge of stability and discrete mechanics.

Multi-task learning often combines losses that span several orders of magnitude, causing homoscedastic uncertainty weighting to degrade severely. We propose Bounded Precision-Geometry Scaling (BPGS), a method that maps each task's log-variance through a bounded sigmoid parameterisation anchored to detached batch loss statistics, and decouples network optimisation from uncertainty optimisation. Its normalised task weights are provably invariant to uniform rescaling under non-degenerate loss scales. We evaluate BPGS on synthetic stress tests and three real-world benchmarks: NYUv2 dense prediction, Yeast multi-label classification, and RF1 multi-target regression. Under pure loss rescaling from $\times 1$ to $\times 1000$, its macro score changes from 0.777 to 0.778, whereas Kendall weighting drops from 0.780 to 0.637; $\ell_1$-normalising Kendall's weights does not close the gap. On NYUv2, BPGS records the lowest depth absolute relative error (0.223), depth RMSE (0.790), and total loss (1.891) among all compared methods, including Nash-MTL. Sensitivity studies on batch size and calibration show small variation across the tested ranges, and runtime overhead relative to Kendall is under 1%. BPGS posts the highest Yeast micro-F1 (0.616) and is competitive on RF1, though PCGrad leads RMSE and MAE there. These findings establish BPGS as a scale-robust alternative to homoscedastic uncertainty weighting, notably effective when loss-scale disparities dominate multi-task optimisation.

Neural-network surrogate models are increasingly used to accelerate scientific simulations, but their deployment in extrapolative and autoregressive settings requires input-dependent estimates of prediction error. In this work, we introduce GeoQ (Geometry-Aware Conditional Quantile Error Estimation), a non-intrusive calibration framework for estimating surrogate error at individual query points. GeoQ represents the error at a query point as an anchor-averaged calibration error plus a learned nonnegative correction. This correction is modeled as an upper conditional quantile of the anchor-relative error increment, using geometry-based features that encode representation-space displacement and local support density. A cross-fitting procedure generates approximately out-of-sample calibration tuples, while a feature-space k-nearest-neighbor support score identifies regions \textcolor{black}{where the learned error model is supported by calibration data}. We evaluate GeoQ on scalar regression, chaotic dynamics, medium-range weather forecasting, and Richtmyer-Meshkov instability prediction. The results demonstrate that geometry-aware conditional quantile modeling provides a practical and non-intrusive approach for validity-aware error estimation in scientific surrogate models.

TinyML systems are enabling machine learning (ML) inference at the edge. However, there is little quantitative analysis of such systems. This paper presents a systematic performance and power characterization of diverse TinyML applications on microcontrollers (MCUs), spanning neural network models, software libraries, operating systems, and hardware architectures. We focus on the impact of the multiple layers of abstraction that provide higher programmability at the expense of performance and energy efficiency. We propose a model to estimate the costs of different abstraction layers and make recommendations for minimizing those costs. Our findings can help designers with Neural Architecture Search (NAS) and CNN inference optimization on edge devices.

Motivated by equivariant neural networks, we study piecewise linear equivariant maps between finite-dimensional real representations of compact groups. We show that all genuinely non-linear piecewise linear behaviour is confined to the subspaces on which the identity component of the group acts trivially, while equivariance forces linearity on the corresponding orthogonal complements. As a consequence, we obtain a compact-group analogue of the finite-group existence criterion of Gibson–Tubbenhauer–Williamson for non-zero equivariant piecewise linear maps between irreducible representations, with the identity component giving rise to a rigidity phenomenon absent from the finite-group case.

Accurate and reliable collection of electricity consumption data through Advanced Metering Infrastructure (AMI) is of great importance for the operation of smart grids, especially for the detection of non-technical losses (NTL). However, real-world datasets frequently suffer from missing values due to communication failures. This paper presents an empirical evaluation of three advanced algorithms for large-scale data imputation: the Optimally Weighted Average (OWA) method, Low-Rank Matrix Completion via SoftImpute, and a Shape-Modeling Autoencoder. Existing studies on missing value imputation in electricity consumption data often lack validation on larger datasets. Therefore, the goal of this paper is to validate the selected algorithms on a large-scale real-world electricity consumption dataset from North Macedonia that includes 17,428 commercial smart meters over two years. The robustness of each algorithm is evaluated by simulating continuous gaps in the data ranging from 1 to 168 hours. The results indicate that OWA provides the lowest overall reconstruction error across the evaluated gap sizes and strong stability in worst-case scenarios for gaps of up to one week. In contrast, the autoencoder exhibits higher variance, while SoftImpute has stable but inferior accuracy. These findings suggest that imputation methods should be selected based on the characteristics of load curve data and highlight the potential for hybrid algorithmic architectures in future grid management systems.

Cheque fraud is a material risk in after-hours business deposit operations because funds may be released within one business day, while cheque clearing takes several days. This timing gap creates a fraud exposure window for financial institutions. Prior mitigation relies on static, deposit-level checks and therefore miss historical client behavior and evolving patterns. To address this gap, we propose a multi-view ensemble ML framework that combines: Extreme Gradient Boosting (XGBoost) for known fraud patterns, Isolation Forest for label-free anomaly detection, and Graph Sample and Aggregate (GraphSAGE) for relational patterns associated with transaction activities. We then combine the three outputs into a single client-level risk score. Under stable conditions, performance is comparable to XGBoost; under a targeted distribution shift, our framework performs best (F1: 83.77%, FPR: 0.69%) versus XGBoost (F1: 82.77%, FPR: 0.72%). These results indicate improved robustness to distribution shift while preserving interpretability through plain-language explanations grounded in behavioural, anomaly, and relational evidence.

Semantic communication shifts the objective of communication systems from accurate symbol reconstruction toward meaning preservation, task accomplishment, and efficient information exchange. However, its evaluation remains fragmented across telecommunications, natural language processing, computer vision, and machine learning, and no single metric can characterize semantic quality across modalities, tasks, and channel conditions. This article surveys key performance indicators (KPIs) for text- and image-based semantic communication systems from a unified, evaluation-centered perspective. Unlike prior surveys primarily organized around architectures, applications, or transmission strategies, this work focuses on how semantic success should be defined and measured. Existing KPIs are classified according to communication goal, source modality, receiver output, reference availability, evaluation level, and channel or resource constraints. The survey reviews reconstruction-based, task-oriented, reference-free, representation-level, perceptual, and channel-aware metrics, and presents a cross-modality comparison of their roles, strengths, and limitations. It further analyzes how unresolved semantic-KPI challenges affect monitoring, quality assurance, resource optimization, fault diagnosis, and standardization. Key open problems include the absence of universal semantic success criteria and standardized semantic ground truth, semantic drift, limited reference-free evaluation, weak integration of machine-learning metrics with communication constraints, and the lack of relation-level and multimodal KPIs. Finally, future research directions are outlined toward standardized, interpretable, adaptive, task-aware, and communication-aware evaluation frameworks.

In on-device NLP tasks, limited resources of embedded hardware, such as the Raspberry Pi 5, require efficient inference strategies. This paper introduces FrugalSOT (Frugal Search Over The Models), a resource-aware model selection architecture for on-device NLP inference. FrugalSOT estimates each request's complexity by extracting features such as prompt length, named entity density, and syntactic complexity. The request is first made to the least complex model that is likely to pass a relevance threshold. If the output of that model falls short of the threshold, the request is made to a more complex model. It is important to note that the relevance threshold undergoes continuous updates in the background. using past validation outcomes in an adaptation process using a low-pass filtering mechanism, thus imparting adaptation to changing input patterns. Experimental results achieved on a Raspberry Pi 5 show that FrugalSOT reduces average inference time and overall computational resource use to a significant extent compared to a single-model baseline approach, without compromising output relevance to the same extent as the most sophisticated model. These results confirm that adaptive model selection can enable efficient, high-quality natural language processing inference on limited devices.

We investigate when a sparse nonnegative matrix can be recovered from a real-valued matrix of much lower rank by zeroing out its negative elements. The potential for such decompositions suggests a mathematical connection between sparsity and rank; we analyze a number of sparse matrices with this latent low-rank structure and use them to illustrate the geometric origins of this connection. Previous algorithms have discovered these decompositions via an alternating minimization over the factors of a low-rank matrix, but to do so, they have also needed to compute and store another matrix, neither sparse nor low-rank, that is the size of their product. We develop a stochastic, alternating least-squares algorithm that operates on smaller blocks of this dense matrix and scales as a result to much larger problems. We also show how to further accelerate this algorithm with sparse optimizations and customized CUDA kernels. As one example, we use the algorithm to analyze the sparse matrix of synaptic weights for the recently published $\textit{Drosphilia}$ connectome. The nonzero elements of this matrix, with 139,255 rows and columns, record the number of synapses between cells in the nervous system of a female fruit fly. Despite a slowly decaying spectrum of singular values, this matrix exhibits a latent low-rank structure that is predictive of cell categories across multiple levels of specificity.

Voice has been proposed as a low-cost screening signal for chronic obstructive pulmonary disease (COPD). COPD is strongly age-associated and voice changes with age, thus such results admit a trivial alternative explanation. We re-evaluate a public sustained-phonation corpus ($1246$ recordings, $68$ participants) under a strictly participant-level protocol. We therefore evaluate on repeatedly drawn age-matched cohorts and report the discrimination achieved by the confounders themselves on those same cohorts. Where raw (unmodelled) age ($0.510$ $[0.469, 0.551]$) and raw gender ($0.479$) are both measured at chance, acoustic models excluding age retain ROC-AUC $0.717$ $[0.552, 0.859]$ and average precision $0.747$ $[0.581, 0.892]$ against a one-to-one baseline of $0.5$, whereas models containing age fall to $0.531$–$0.679$. The separation is reproduced by two further learners with fixed hyperparameters. Two findings have broader methodological implications: models trained with age transfer less effectively to an age-balanced target cohort than otherwise identical models trained without age, and fourteen classical voice-quality and perturbation measures achieve comparable discrimination to a $55$-dimensional combined representation. We conclude that a non-age acoustic signal is present, that confounding by recording conditions cannot be excluded from the released features, and that the evaluation protocol in standard use cannot distinguish these possibilities.