Papers

For local large language model agents, active context is a scarce resource: memory capacity, prefill latency, cache growth, and service objectives all constrain how many input tokens each call can afford. We present BudgetBench, an active-budget protocol and reference harness that treats the per-call input-token budget as the independent variable when comparing memory strategies. Holding the model, task, sampler, and decoding fixed, it sweeps budgets over 2K, 4K, 8K, 16K, and 32K tokens and records quality, budget utilization, latency, and, as a first-class outcome, budget-violation rates. The core contribution is this reusable measurement surface: a swappable MemoryStrategy contract, explicit budget enforcement, deterministic or versioned graders, prompt-audit metadata, and reproducibility artifacts, released at https://github.com/aviskaar/budgetbench. We substantiate the protocol with pilot studies rather than final rankings. Across a local qwen2.5:1.5b pilot (89 items each on SWE-bench Verified and LongBench v2), a hosted 50-item Qwen3 30B-A3B LongBench replication with exact tokenization, and a 500-item LongMemEval oracle study scored by the official GPT-4o evaluator, the harness exposes budget-compliance failures, non-monotonic quality curves, and operating points that single-budget evaluation hides. The budgeted-versus-full-context direction remains unresolved: the local slice is near-null and the hosted replication favors full context in point estimate. We report results transparently, including that the early pilot's tokenizer approximation undercounts some served-model prompts, so its violation rows are tokenizer-approximation diagnostics, not claim-bearing results; all timings are operational diagnostics. The reusable contribution is the protocol, harness, and failure-reporting discipline needed to scale fixed-budget memory-strategy evaluation.

AI deployment performance is shaped not by model architecture alone, but by interactions among compression, compiler transformations, and serving policies. Published benchmarks often report latency and throughput under incomparable conditions, limiting their use for deployment decisions. This paper presents a unified analytical treatment of inference optimization across the deployment stack. We introduce a three-layer taxonomy covering model-level techniques such as quantization, pruning, and distillation; compiler transformations such as graph fusion, layout optimization, and kernel autotuning; and system policies such as dynamic batching, admission control, and memory tiering. We formulate deployment as a constrained multi-objective optimization problem over accuracy, latency, throughput, memory footprint, and energy, and analyze a deployment-ranking functional with Pareto monotonicity and scale invariance. Roofline models show how memory-bandwidth hierarchies bound performance across precision regimes, while queuing models explain how service-time changes amplify response time under load. To improve comparability, we propose an evidence protocol that separates measured, derived, and analytical claims; limits numerical comparison to within-paper results; and requires reporting of hardware, software versions, batch semantics, and thermal state. We synthesize evidence from edge platforms, including Jetson AGX Orin and five inference frameworks; data center GPUs, including A100 and H100 with three LLM serving engines; and quantization studies across the Llama-3.1 family. The synthesis shows that deployment outcomes are governed by cross-layer interactions that no single-layer analysis can predict. We conclude with a constraint-aware selection procedure and open problems in compiler-serving co-optimization, cross-hardware performance prediction, and standardized energy reporting.

Physics-informed neural networks (PINNs) provide a mesh-free framework for solving partial differential equations, yet their performance in nonlinear problems is often limited by slow convergence, gradient imbalance, and insufficient resolution to capture localized intermittent structures such as shock waves[1]. These issues arise primarily from the use of fixed weights of loss and uniform collocation point distributions, which cannot adapt to the evolving complexity of the solution field during training. To address these challenges, Gradient-Guided Gaussian Adaptive Sampling Physics-Informed Neural Networks (3GAS-PINNs) is proposed in this paper, which combines uniform probability distribution and Gaussian-smoothed probability distribution derived from the spatial gradients of solution, to maintain global constraint satisfaction as well as concentrating collocation points in regions of high gradient. Thus, intermittency structures like shock wave and solitons can be accurately captured. The method is evaluated on three benchmark nonlinear problems, including one-dimensional forced Burgers equation, Korteweg-de Vries (KdV) equation and nonlinear Schrodinger equation, all of which exhibit steep gradients or strong nonlinearity. In comparison with baseline PINNs, 3GAS-PINNs can effectively promote the physical consistency in intermittent regions. The accuracy of the numerical simulation can be improved by a factor of up to 14.

Relational reasoning requires the process of perceptual understanding, comparing, and integrating the underlying relationships between concepts. This ability consists of multiple categories, such as analogical, structural, and cause-effect, each capturing a different aspect of higher-order understanding. To examine the performance of multimodal large language models (MLLM) on these relational inference tasks, we developed SciReC, a model-adaptive multimodal academic dialog benchmark. As the relational reasoning process involves multiple representations and various factors (visual understanding, exhibiting knowledge, and memory recall), we propose DMRA, a deficit-based diagnostic framework that quantifies the contribution of these components to identify the primary cause of unsuccessful cases. Claude 4.6 achieved the best performance on the overall relational score with 73%, followed by GPT 5.4 with 68%. Performance trends indicate that open-source models achieve their lowest scores on spatial relations, while proprietary models struggle more with hierarchical and sequential relations. Across domains, model performance is lowest on Astronomy and highest on Psychology. The results of DMRA reveal that relational reasoning is the primary source of error across all models, followed by memory limitations.

Large output embedding matrices create a significant memory bandwidth bottleneck during autoregressive decoding, especially for compact LLMs with large multilingual vocabularies. We reformulate the output projection followed by top-k token selection as a maximum inner product search over token embeddings and replace the dense vocabulary projection with an HNSW-based vector index. The resulting output head retrieves only a small candidate set of high-scoring tokens and can be integrated into existing decoding pipelines by scattering retrieved logits into a sparse full-vocabulary tensor. On CPU inference with Gemma 3, Llama 3.2, and Qwen 3 models, our method substantially accelerates the output projection and improves end-to-end batch-size-one decoding throughput by up to 82% for Gemma 3 270M, while preserving generation quality under AlpacaEval evaluation. These results suggest approximate retrieval is a practical alternative to dense output projections in latency-sensitive small-batch decoding.

Subscriber attrition is a costly, persistent challenge for telecommunications providers, with monthly churn of roughly 1.9% in mature markets eroding billions in revenue annually. Predictive models can flag at-risk customers accurately, yet they are routinely excluded from frontline CRM workflows because high-performing ensemble and non-linear architectures are opaque: a retention specialist cannot design a personalised intervention from a probability score alone, without knowing why a subscriber is at risk. This paper addresses that gap. We benchmark four classifiers–Logistic Regression, Random Forest, XGBoost, and LightGBM–on the IBM Telco Customer Churn benchmark (7,043 records; 19 features; 26.5% churn, balanced to 50% via SMOTE on the training partition only). Logistic Regression attains the strongest AUC-ROC (0.8411) and LightGBM the highest accuracy (78.42%); all four fall within a 0.011 AUC band (0.831–0.841), and 5-fold cross-validation confirms the leading models are effectively tied. Explanations are delivered at two granularities: a global SHAP ranking identifying tenure, total charges, and month-to-month contract as the dominant churn signals, and instance-level SHAP and LIME decompositions that expose the drivers behind each prediction. Building on these outputs, we introduce a four-layer CRM integration architecture that converts risk scores and attribution vectors into tiered segmentation, maps top features to structured retention-action templates, and routes campaign outcomes into a retraining feedback loop. Targeting the highest-risk quintile is projected to cut overall churn by 3.3–5.3 percentage points, preserving an estimated $199K--$319K per campaign cycle.

Multi-table learning remains a major challenge in machine learning for healthcare and other complex information systems. Relational data combine several sources of complexity, including large data volume, high-dimensional variables, high-cardinality categorical features, complex inter-table dependencies, and repeated temporal observations. We introduce the Relational Hypergraph Transformer (RHT), a unified architecture that represents relational databases as hypergraphs, learns pentadimensional embeddings (PentE), and performs sparse relational attention with complexity proportional to the average relational degree rather than the square of the number of entities. We formally define the architecture, derive the complexity of its attention mechanism, and provide an open-source reference implementation. We evaluate RHT on the public Synthea synthetic electronic health record dataset using multi-label prediction of SNOMED CT condition codes per encounter, a task characterized by high categorical cardinality and long-tailed label distributions. Comparisons with tabular, relational, and temporal graph baselines show that RHT produces more semantically coherent embeddings while remaining computationally scalable. In this benchmark, the highest rare-code recall is achieved by XGBoost, whereas RHT attains the strongest embedding semantic coherence. We also report ablation studies quantifying the contribution of each architectural component. Clinical validation on MIMIC-IV is planned following PhysioNet credentialing. Source code and experimental protocols are provided in the accompanying repository.

Vision-language models enable zero-shot classification through natural language prompts, but performance is sensitive to prompt formulation, especially in specialized domains. Zero-shot Prompt Ensembling (ZPE) addresses this by weighting prompts by discriminative signal, yet its behavior under domain shift remains unexplored. We evaluate ZPE in the agrifood domain using CLIP and SigLIP across four datasets and four prompt pools, spanning in-distribution (ID) food and out-of-distribution agricultural benchmarks. ZPE provides limited benefit under ID conditions but substantially improves performance and calibration under domain shift, where domain-specific pools of 51-52 prompts consistently outperform generic pools of 247-426. Lexical analysis shows that ZPE acts as an unsupervised domain-alignment detector without label access. We further introduce PID (Prompt-based Inconsistency Detection), which repurposes prompt disagreement as epistemic uncertainty, improving failure detection under severe domain shift where standard confidence measures collapse.

We present a demonstration for generative multicasting with on-device, intent-aware semantic decomposition. At the transmitter, DNN-based segmentation extracts a semantic map from the source video, decomposing it into multiple sub-signal classes based on multi-user receiver intents. The transmitter broadcasts the semantic map to all users over shared wireless/network resources, thereby utilizing orthogonal resources only to transmit the sub-signal classes intended for each user. Users partially reconstruct and partially synthesize the signal by combining the received intended classes with non-intended classes locally synthesized by a generative model from the semantic map. We derive the rate-distortion/perception curves for reconstruction/synthesis with the generative model, to adaptively set compression rates for the semantic map and intended classes. Generative multicasting significantly reduces the wireless/network resources required for existing/emerging multimedia multicasting applications. The system is real-time on a Google Coral Edge TPU with 4 TOPS (int8). This is the first demonstration of generative multicasting representing a substantial advancement in on-device generative SemCom.

Latent reasoning has emerged as a powerful alternative to text-based Chain-of-Thought (CoT), offering significant gains in computational efficiency by compressing verbose reasoning into compact embeddings. However, compressing reasoning into the latent space renders the thinking opaque, hindering its interpretability. Current methods present a stark trade-off: they either function as unexplainable ”black boxes” (e.g., Coconut), where the latent reasoning is not human-readable, or rely on separate post-hoc decoders for explainability (e.g., Heima), introducing architectural overhead and decoupling the explanation from the actual reasoning process. In this work, we present a unified framework for Self-Explainable Latent Reasoning (SELR) that trains a single model to perform efficient and inherently explainable latent reasoning. Our core contribution is a novel multi-task training objective that optimizes for two goals simultaneously: (1) an Answer Loss that optimizes the latent reasoning trajectory to produce accurate final answers, and (2) a CoT Loss that explicitly trains the same model to decode its own latent representations back into human-understandable reasoning steps. This design ensures that generated latent representations are both task-effective and semantically interpretable, eliminating the need for external decoders. We validate the effectiveness of SELR on both Large Language Models (LLMs) and Vision-Language Models (VLMs), demonstrating that SELR achieves superior token efficiency and accuracy compared to baselines, while uniquely providing self-contained explainability without auxiliary models. Project page is available at https://jasondayuan.github.io/SELR/.

Monitoring livestock behaviour under extensive conditions would provide valuable insights to assess animal adaption to environmental perturbations in agroecological systems (e.g., heat waves, parasitism, predator attacks). Animal behaviour can be monitored using accelerometer data collected from neck-collars combined with artificial intelligence models. However, large amounts of accelerometer data aligned with annotated behaviours are necessary to develop accurate models of behaviour prediction. In particular, developing reliable models for extensive systems requires data collected across a wide range of representative conditions. The dataset includes 79 hours of tri-axial accelerometer data aligned with behaviours manually annotated from video recordings for 120 Romane ewes born between 2021 and 2024. The ewes were derived from two divergent genetic lines after three and four generations of selection started 10 years ago: low and high social attractiveness, noted S-and S+, and low and high tolerance towards humans, noted H-and H+. They were reared under the extensive system applied to the Experimental Unit of La Fage (UEF, INRAE, Saint-Jean-et-saint Paul, Aveyron) where 250 sheep were reared exclusively outdoors on 280 hectares of rangeland in southern France. First batch of data was collected on March, June and July 2024 at the UEF under a range of extensive conditions, including sloping pastures and heat-wave periods. Ewes were equipped with accelerometer neck-collars specifically designed for young sheep on pasture. They were grouped on experimental paddocks for 4 to 8 hours and provided with fresh grass and ad libitum access to water. The animals were simultaneously video-recorded using an elevated CCTV camera. Behaviour annotation was carried out using Behavioral Observation Research Interactive Software focusing on the main behaviours on pasture: Grazing, Ruminating, Resting, Moving, and ”Other”, grouping all remaining activities. Annotations and corresponding accelerometer sequences were aligned using Python language, based on a time synchronization procedure. A second batch of data was acquired on November 2025 to supplement the dataset with the moving activity. For that purpose, ewes were equipped with the accelerometer collars and moved on tracks from the housing area to the pastures, corresponding to an approximately 10 minute-walk. The start and end times of the moves for each ewe were used to align the corresponding accelerometer data with the moving activity. These data were then merged with the dataset from the first batch. The resulting dataset is ready to use for applying artificial intelligence models to classify the 5 main behaviours of sheep under extensive grazing systems from accelerometer data.

Deploying large language models (LLMs) for operations research (OR) tasks remains challenging because correctness depends on a coherent modeling process, not merely a correct final answer. Standard autoregressive generation operates on a myopic policy, which sometimes fails to anticipate whether a partial formulation can be validly extended into a globally consistent optimization model. Consequently, locally plausible steps may propagate into catastrophic downstream formulation or solver code errors. To address this, we propose an uncertainty-aware, training-free inference framework for OR mathematical modeling. Without updating model parameters, our method evaluates intermediate candidate steps using short lookahead simulations to quantify downstream predictive uncertainty or probability concentration. Candidates that demonstrate a higher likelihood of yielding coherent mathematical formulations are then dynamically selected via importance resampling. Empirical evaluations across multiple OR benchmarks (including NL4OPT, MAMO, and IndustryOR) demonstrate that our framework consistently outperforms both standard and low-temperature baselines, establishing an efficient, training-free paradigm for reliable OR formulation generation.

Cross-paper comparison of sparse autoencoder (SAE) interpretability often relies on autointerpretability scores. In this evaluation pipeline, a language model (LM) explains each feature, and another LM scores the explanation. For these comparisons to be meaningful, scores must reflect stable properties of the features rather than confounding aspects of the evaluation pipeline. Through systematic experiments across four metrics (simulation, detection, fuzzing, purity), two models (Pythia-160M, Apertus-8B), and four axes of methodological variation, we show that this assumption does not hold. Specifically, we find that R1) methodological variance collectively exceeds architectural variance across all metrics and tested models; R2) each metric exhibits a distinct instability profile, with detection being the most stable and fuzzing unreliable across all conditions; R3) top-k feature rankings do not stay consistent across corpus and draw conditions, masking per-feature instability behind stable mean scores; a failure that cannot be detected by monitoring explanation similarity alone. These findings suggest that cross-paper comparisons based on autointerpretability scores may reflect pipeline differences rather than architectural differences, with implications for the ongoing debate on SAE utility. More broadly, unreliable evaluation slows progress in interpretability research at a time when reliable tools for understanding AI systems are needed. To support evaluation, we contribute a variance decomposition approach, a Stability Check, and a Minimum Reporting Checklist.

This paper tackles the problem of stock ranking and portfolio construction under realistic investment settings by jointly modeling temporal dynamics and cross-sectional dependencies. We propose the Soft-Threshold NMI-prior Transformer Graph Attention Network (STN-TGAT), which integrates a temporal Transformer with a Graph Attention Network to capture long-horizon sequential patterns and dynamic inter-stock relationships. An NMI-based prior graph combined with a soft-threshold sparsification mechanism enhances structural robustness by mitigating noisy correlations while preserving informative connections. The portfolio formation process incorporates practical considerations, including Top-5 selection within the Top-50 $S\&P$ 500 constituents, explicit weight allocation, and transaction cost adjustment, thereby aligning the evaluation with real-world trading conditions. Empirical results on real-world data demonstrate that STN-TGAT consistently outperforms benchmark models from predictive accuracy and investment profitability measured by portfolio returns. These findings suggest that combining decision-aligned training with adaptive relational modeling provides a coherent and practically effective framework for data-driven portfolio construction.

Real-world intelligent systems often require both distributed collaboration across data-isolated clients and continual adaptation to evolving tasks. This setting naturally gives rise to Federated Class Incremental Learning (FCIL), which combines Federated Learning (FL) and Continual Learning (CL). However, their combination introduces two coupled sources of interference: spatial interference from heterogeneous clients and temporal interference from sequential tasks, jointly leading to Spatial-Temporal Catastrophic Forgetting (ST-CF). Existing approaches typically address spatial and temporal interference with separate mechanisms, often incurring additional client-side computation or communication, while leaving directional interactions among updates during aggregation unregulated. In this paper, we reinterpret FCIL as a unified multi-task learning problem, where both client and task updates are represented as adaptation vectors in a shared parameter space. Based on this view, we propose Unified Geometric Surgery on Spatio-Temporal Adaptation Vectors (SUM), a purely server-side framework that performs geometric surgery on adaptation vectors during aggregation. Spatial SUM mitigates client-level interference within each round, while causal online temporal SUM removes cross-task interference over time without additional client-side computation, communication, or memory beyond standard federated training. Empirically, SUM achieves up to 22% improvement over prior FCIL methods across diverse vision and language benchmarks while remaining robust to unreliable clients and maintaining computational efficiency.

Pretrained generative foundation models cast forecasting as conditional generation from a learned predictive distribution and forecast unseen series zero-shot. We establish three results that turn their reported success into an actionable, mechanistic understanding. First (a positive benchmark result): on a power-controlled study of 1728 rolling-origin forecasts over 36 series from 19 datasets spanning the full range of STL trend strength (F_T in [0.17, 1.00]), a zero-shot Chronos model significantly outperforms four strong classical baselines – drift, seasonal-naive, Theta, and additive Holt-Winters/ETS – with the best mean MASE (1.187 vs. Theta 1.337, ETS 1.656; Friedman chi^2 = 46.08, p = 8.75e-09; Holm-corrected Wilcoxon p <= 0.015 against every baseline; a Nemenyi critical difference separating it from the classical pack). Second (a novel, quantified mechanism): a controlled synthetic experiment with a known trend-generating process shows why – and reveals that the win does not come from better trend extrapolation. When the true trend is linear, damped, or exponential, additive ETS tracks the slope (slope-tracking ratio 1.02, 1.34, 0.98) whereas Chronos systematically under-extrapolates, behaving as a trend-shrinkage estimator (ratio 0.80, 0.49, 0.36). Third (an actionable selection rule): because the advantage is a shrinkage effect, it is predictable from trend strength alone – the generative model wins 78% of low-trend series but only 44% of high-trend ones, and its edge over ETS is significant on the low-trend stratum (0.982 vs. 1.671, p = 0.002) yet a tie on the high-trend stratum (p = 0.18). Trend strength, computable before forecasting from the training context alone, is therefore a practical a-priori indicator of when to deploy a foundation model. We additionally document a calibration shortfall (80% intervals cover 0.77).

Deep learning models have shown strong potential for time series forecasting, yet their deployment in real-world environmental monitoring remains challenging due to non-stationary dynamics and limited explainability. In this work, we investigate explainability as a central tool for understanding continual learning in adaptive time series forecasting, with Experience Replay strategies. We study neural forecasting architectures such as PatchMixer, PatchTST and DLinear, augmented with attention-based sampling mechanisms to support model adaptation over time. Explainability is leveraged through attention rollout and gradient-based attribution methods (Grad-CAM) to analyze both predictive behavior and sampling strategies within a continual learning framework. Experiments conducted on real-world piezometric time series exhibiting heterogeneous patterns and regime shifts show that analyzing model and sampling behaviors provides valuable insights into the dynamics of the continual learning framework. Beyond predictive performance, our results highlight the challenges and opportunities of using explainability to understand continual learning behaviors, revealing how attribution patterns evolve over time and how they can inform data selection and adaptation strategies in non-stationary forecasting scenarios.

Air pollution causes an estimated 7.9 million premature deaths annually, making accurate forecasting a critical public health priority. Machine learning is increasingly being applied to forecast air pollution levels, yet existing benchmarks remain narrow in both geographic scope and pollutant coverage, and fail to evaluate the latest generation of time series foundation models (TSFMs) on real world, large scale data. We present Air Quality Arena (AQA), a large scale multi-country and multi-pollutant dataset (AQA-Data) and benchmark (AQA-Bench) to address this gap. AQA covers 6 major pollutants over a three year period across 7 diverse countries and 4 continents, with more than 14,000 station-pollutant series, aiming to provide a comprehensive benchmark for air quality tasks. We benchmark this dataset across 11 leading time series foundation models and classical baselines to assess performance on short-term air quality forecasting. Our results demonstrate that TSFMs are effective zero-shot forecasters and consistently outperform classical baselines, with our top-performing model employing a cross-modal architecture that leverages a vision foundation model for time series forecasting. AQA is publicly released at AirQualityArena.github.io

Remaining useful life (RUL) prediction estimates how long an engine can continue safe operation and is central to maintenance planning. N-CMAPSS extends C-MAPSS by simulating run-to-failure aero-engine trajectories using recorded real-flight profiles and retaining complete within-flight time series rather than cycle-level snapshots. However, this added realism reduces evaluation control because full-flight records increase data volume and entangle degradation cues with operating-regime variation, complicating preprocessing choices and direct comparisons of RUL modeling performance. To mitigate this issue, this paper proposes CruiseBench, a cruise-stage RUL benchmark derived from N-CMAPSS. It introduces CPM-N-CMAPSS (Cruising-Period Mask for N-CMAPSS), a mask artifact that stores cycle-local cruising intervals identified by the common-altitude method for the nine accessible subdatasets. CruiseBench applies a fixed protocol to the masked rows, using scenario descriptors and measured sensors as inputs while excluding virtual sensors, health parameters, and auxiliary metadata from the feature tensor, preserving native-resolution windows, and applying dataset-wise RUL caps. Experiments with LSTM, GRU, TCN, and TSMixer provide baseline results for this setting. Under CruiseBench-eta5-W256-S10, TSMixer obtains the lowest average RMSE, $3.4\pm1.71$, and Saxena score, $(2.50\pm2.99)\times 10^{4}$. Ablation studies show that flight-stage selection, temporal downscaling method, and RUL-cap threshold affect reported results. With its fixed cruise-stage protocol, CruiseBench provides a reproducible sub-benchmark for controlled RUL model comparison and CPM-N-CMAPSS provides a stage-specific data foundation for future transfer-learning and domain-adaptation studies.

Prior work has shown that transformers can perform exact Bayesian filtering within a fixed hypothesis class. Can they also perform Bayesian model selection – identifying the correct hypothesis class from data? We introduce model-selection Bayesian wind tunnels: controlled environments where ground-truth posteriors over hypothesis classes are available in closed form. Using fixed-point-free involutions – whose defining property f(f(x))=x is purely relational – a 2.8M-parameter transformer achieves 0.01-bit entropy agreement with the Bayesian optimum (3 seeds), with both integer tokens and opaque symbols whose meanings change every episode. This extends to non-nested comparisons: involutions vs. 3-cycles (where neither class is a subset of the other) achieve class-posterior MAE under 0.001, demonstrating genuine model selection beyond simplicity/subset bias. We then identify a sharp perceptual access condition: when the discriminative statistic requires arithmetic – modular addition (rotations) or multiplication (f(x)=cx mod p) – model selection succeeds with integer tokens but fails completely with opaque symbols, and this boundary persists under 112x scaling (2.8M to 316M parameters). A stationarity control confirms the operative factor: opaque tokens with a fixed relabeling succeed (0.009-bit MAE), showing that stable semantics, not integer identity, enable circuit compilation. Header subtask diagnostics localize the failure to the composition of header inversion with arithmetic rather than header parsing itself. Probing frontier LLMs on the same tasks shows qualitative Bayesian behavior but a large calibration gap (~55x), measured through lossy probes and therefore directional rather than exact.

Subquadratic alternatives to attention require compromises when applied to multi-dimensional data: standard convolutions lack global receptive fields and input dependency, while recurrent models require rasterizing data such as images, volumes, and partial differential equation (PDE) into an ad-hoc $1\rm D$ scan order that violates their spatial structure. We introduce HyenaND, a subquadratic, global, input-dependent operator that acts directly on the native geometry of multidimensional data through convolutions with implicitly parametrized global, input-dependent multi-dimensional convolutional kernels. Our CUDA implementation, nSubQ, fuses the FFT-convolution path to turn HyenaND's $\mathcal{O}(L \log L)$ scaling into wall-clock speedups. Across long-context genomics, computer vision, medical imaging, and PDE modeling, pure HyenaND stacks match the accuracy of strong attention baselines, while hybrid configurations that interleave HyenaND and attention layers outperform both pure attention and strong recurrence-based hybrids.

Physics-informed neural networks (PINNs) provide a mesh-free framework for solving partial differential equations, but their training is often affected by loss imbalance, optimization stiffness, and difficulty in capturing localized or multi-mode solution structures. Hard-soft PINNs (HSPINN) alleviate part of this difficulty by embedding Dirichlet or periodic constraints directly into the trial space, but the resulting fixed admissible representation can still be poorly conditioned for sharp or heterogeneous residual fields. This paper proposes a reliability-aware hard-soft PINN (RA-HSPINN) that preserves exact embedded constraints while introducing a bounded learnable reliability field to modulate the interior representation. The method combines this reliability-aware ansatz with inverse-EMA global loss balancing and lightweight regularization, while retaining the standard mean-square residual form. The reliability field is a numerical modulation variable, not a physical parameter or calibrated probability. RA-HSPINN is evaluated on nonlinear Burgers equations, periodic convection, a mixed-boundary Poisson problem, and a mixed first-order Poisson system. Compared with HSPINN, it reduces the relative error by $98.65%$ for sharp-gradient Burgers, $72.42%$ for Burgers data with noisy and incompatible initial conditions, $61.18%$ for smooth periodic convection, $60.02%$ for localized periodic convection, $29.36%$ for mixed-boundary Poisson, and $82.17%$ for a multi-mode mixed first-order Poisson system. The results show that reliability-aware modulation is most beneficial when hard-soft trial spaces are admissible but difficult to optimize, especially in localized, unreliable-data, and multi-mode PDE regimes.

We study whether alignment schemes that reshape a base model's output distribution, combined with bounded safety filters, can drive the probability of harmful behavior to zero in modern large language models. Recent research suggests that harmful behaviors can persist under preference-based alignment and that external filtering can be computationally hard in the worst case, but it remains unclear whether practical alignment pipelines that largely preserve internal representations can eliminate harmful behavior entirely rather than merely suppressing its most visible forms. We formalize this setting using support-preserving alignment operators together with bounded filtering algorithms under black-box, white-box, and statistical-query access, and analyze their ability to approximate an ideal eliminator that removes all harmful mass. Building on this framework, we provide computational and information-theoretic arguments indicating that, under these constraints, bounded filtering may fail to eliminate all harmful outputs supported by the base model's distribution. To evaluate these limits empirically, we analyze a range of state-of-the-art open-weight and hosted LLMs accessed via OpenRouter under bounded black-box, white-box, and statistical-query filters on adversarial prompts drawn from curated cybersecurity scenarios and PKU-SafeRLHF. Across models, filter classes, and query budgets, the estimated harmful-output rate decreases with additional filtering compute but consistently plateaus above zero, suggesting a persistent empirical harm floor.

We present AV-JEPA, an elegant multimodal extension of LeJEPA to audio-visual self-supervised learning. Using an early-fusion Vision Transformer and modality dropout as masking, the model is trained to align the embeddings of global and per-modality local views, while the SIGReg objective encourages a theoretically optimal distribution. This achieves cross-modal alignment in the latent space, resulting in a remarkably clean architecture with no decoder, EMA teacher, complex multi-term losses, or contrastive negatives. The proposed AV-JEPA backbone delivers competitive classification performance on VGGSound (57.1% top-1) and AudioSet (32.7 mAP) and supports zero-shot audio-video retrieval out of the box.

TriAttention is a recent method for shrinking the KV cache of long-reasoning LLMs: it scores each cached key by how much attention it is likely to receive and evicts the lowest-scoring ones. Because a key does not know how far away its future queries will sit, the score is averaged over a ladder of 17 possible future distances. We point out that this average is free: the future distance enters the score only through the position-dependent rotation, so the whole 17-fold average collapses–exactly, by a one-line algebraic identity–into a single per-band weight that is computed once, offline. Scoring a key then costs one evaluation instead of seventeen, with no change to which keys get pruned. The saving is modest and lives entirely in TriAttention's pruning-score computation, not in the attention kernel; we present it as a small, exact complement to their method, and we confirm the identity numerically.

Vision Transformers achieve strong image classification accuracy but process all image regions with nearly the same computation, even when many regions are redundant or uninformative. Recent adaptive inference methods reduce this cost by selectively compressing tokens or terminating inference early, but combining these mechanisms often causes unstable intermediate representations and accuracy degradation. We introduce Fusion, a unified adaptive inference framework that coordinates token merging, early exiting, and token pruning through a simple staged design: tokens are merged first, confidence is evaluated next, and pruning is applied only to samples that continue inference. This ordering allows the three mechanisms to operate cooperatively rather than competitively. Fusion further includes lightweight routing modules that adapt compression strength to each input and support inference-time adjustment of the accuracy–latency trade-off without retraining. On ImageNet-1k with DeiT-S, Fusion matches or surpasses state-of-the-art adaptive ViT methods at comparable compute budgets while reducing calibration error by up to $4\times$ and inference energy by $48\%$. Experiments across ImageNet-100, CIFAR-100, and ImageNette with multiple ViT backbones demonstrate consistent transferability without dataset-specific tuning.

Work-related Musculoskeletal Disorders (WMSDs) require continuous ergonomic assessments. While Rapid Entire Body Assessment (REBA) is a gold-standard observation tool, manual monitoring is labor-intensive, and vision-based automation leads to privacy concerns. This paper proposes a novel end-to-end multi-task learning framework for privacy-preserving ergonomic assessment using millimetre-wave (mmWave) radar. A spatio-temporal backbone reconstructs 3D human skeletons, which serves as the biomechanical foundation for a subsequent regression head to generate REBA risk scores. To overcome the sparsity of radar point clouds, we utilise a multi-objective loss function incorporating biomechanical limits and temporal smoothness constraints. Furthermore, we implement an oversampling strategy to address the imbalance of high-risk postures in existing datasets. Experimental results on MMFi dataset demonstrate that our framework achieves a Categorical Accuracy of 77.78% and real-time performance with an inference latency of 5.70 ms. Our method reaches a High-risk REBA MAE of 0.93, which significantly outperforms both direct regression and two-stage pipelines in high-risk scenarios, providing a robust solution for non-invasive industrial ergonomic assessment.

Lightweight neuromorphic computing offers a promising route to efficient AI, with particular benefits for resource-constrained edge deployments. However, its scalable deployment that can reliably transfer the expected performance has long been hindered by device-to-device variations, which necessitate costly and repeated re-training on new copies and undermine the practical advantages. To address this issue, we introduce a model-free temporal-switch (TS) framework to improve the direct transfer performance, without post-training calibration or adjustment. The TS framework provides a methodology to incorporate a broader spectrum of devices in the training process. In the validation using memristor-based reservoir computing, it enables high performance on unseen devices with a directly transferred readout. It achieves improved prediction in the representative Mackey–Glass benchmark, and the accuracy of 92.4% in spoken digit classification. Its efficacy is validated across different memristor families and RC configurations. Theoretical analysis not only reveals the general computational mechanism underlying its efficacy, but also underlines its potential applicability to other physical platforms.

The stable coloring of the Weisfeiler-Leman (1-WL) test is a cornerstone of Graph Neural Networks because it provides an upper bound to the expressive power of message-passing architectures. Unfortunately, computing it presents two fundamental bottlenecks. First, classic algorithms are inherently sequential and cannot exploit modern massively parallel hardware. Second, these are global algorithms, i.e., they require availability in memory of the full graph, severely limiting applicability to real-world instances. We leverage a linear-algebraic interpretation of 1-WL stable coloring and introduce two key contributions: (i)~a randomized refinement algorithm with tight probabilistic guarantees and (ii)~a correctness-preserving batching scheme that decomposes the graph into independently processable subgraphs while provably returning a stable coloring of the original graph. This approach maps directly to GPU-efficient primitives. In numerical experiments, our CUDA implementation delivers speedups up to two orders of magnitude over classical CPU-based partition refinement and, for the first time, successfully computes stable colorings on web-scale graphs with over 30 billion edges, where CPU baselines time out or fail.

Deep learning models for medical diagnosis frequently exhibit substantial performance disparities across sensitive subgroups (e.g., race, sex), even when average accuracy is high. While generative data augmentation offers a route to mitigate this, existing strategies are suboptimal; they typically address only one or two dependency channels between sensitive attributes and image features. We formalize the medical image formation process via a structural causal model, revealing that sensitive attributes actually influence image content through four distinct pathways-a structural complexity neglected by prior works. Based on this insight, we introduce CIPHER (Causal Intervention Pathways for Healthcare Equity and Robustness), a framework designed to systematically intervene on all four causal paths. To achieve this, CIPHER utilizes a diffusion backbone equipped with classifier-free guidance and null-text inversion. This technical design enables the faithful reconstruction of patient-specific anatomy while allowing for the precise, editable synthesis of counterfactuals required to break sensitive dependency chains. We tested CIPHER using chest X-ray and dermoscopy benchmarks across both standard and shifted data distributions. By employing a multi-pathway intervention strategy, our model reduced worst-group disparities by an average of 35.8% compared to disease-conditioned synthesis baselines, while also improving total diagnostic accuracy