In this work, we conduct a systematic analysis of Native Sparse Attention (NSA) and propose targeted improvements that enhance long-context modeling. A key insight is that alternating between local (sliding-window) and global (compression, selective) attention across layers, rather than using fixed patterns, enables more effective propagation of long-range dependencies and substantially boosts performance on long-sequence tasks. Meanwhile, we further refine NSA's branches with Latent Attention that the sliding-window branch is enhanced with Multi-head Latent Attention (MLA) while compression and selective branches adopt Group-head Latent Attention (GLA). These changes reduce KV-cache memory by 50\% versus NSA while improving the model's common-sense reasoning and long-text understanding capabilities. Experiments on models from 340M to 1.3B parameters (trained on 15B and 100B tokens) show our method matches or exceeds full attention and native sparse attention in both common-sense reasoning and long-context understanding tasks.

Vision Transformers have been tremendously successful in computer vision tasks. However, their large computational, memory, and energy demands are a challenge for edge inference on FPGAs -- a field that has seen a recent surge in demand. We recognize the benefits of recent works on logic and Look Up Table (LUT) based networks, such as LogicNets, NeuraLUT, DWN, among others, in offering models that simultaneously reduce both the memory and compute footprints. However, these models natively do not perform well on common vision tasks, such as CIFAR-10/100. In this work, we propose LL-ViT, a novel edge optimized vision transformer design that integrates layers of LUT neurons within the transformer architecture. Based on our characterization that reveals that a majority of model weights and computations are from the channel mixer (MLP layer), we design an alternate LUT-based channel mixer, and simultaneously develop an FPGA-based accelerator for LL-ViT. Contrary to some attempts to replace each multiplication with a table lookup, our architecture utilizes a neural learning approach which natively learns the LUT functions. This approach allows for reduced model sizes, and a computational and energy-efficient inference solution for vision transformer models. Evaluating on edge-suitable workloads, we achieve accuracies of 95.5% on CIFAR-10, 78.8% on CIFAR-100, and 60.9% on Tiny-ImageNet datasets, comparable to the baseline transformer. LL-ViT eliminates over 60% of the model weights and 50% of the multiplications in the model, and achieves 1.9x energy efficiency and 1.3x lower latency over an integer quantized ViT accelerator, while also offering superior throughput against prior works at a 10.9W power budget.

Predicting public transit incident duration from unstructured text alerts is a critical but challenging task. Addressing the domain sparsity of transit operations with standard Supervised Fine-Tuning (SFT) is difficult, as the task involves noisy, continuous labels and lacks reliable expert demonstrations for reasoning. While Reinforcement Learning from Verifiable Rewards (RLVR) excels at tasks with binary correctness, like mathematics, its applicability to noisy, continuous forecasting is an open question. This work, to our knowledge, is the first to bridge the gap between RLVR LLM training with the critical, real-world forecasting challenges in public transit operations. We adapt RLVR to this task by introducing a tolerance-based, shaped reward function that grants partial credit within a continuous error margin, rather than demanding a single correct answer. We systematically evaluate this framework on a curated dataset of NYC MTA service alerts. Our findings show that general-purpose, instruction-tuned LLMs significantly outperform specialized math-reasoning models, which struggle with the ambiguous, real-world text. We empirically demonstrate that the binary reward is unstable and degrades performance, whereas our shaped reward design is critical and allows our model to dominate on the most challenging metrics. While classical regressors are superior at minimizing overall MAE or MSE, our RLVR approach achieved a 35\% relative improvement in 5-minute accuracy (Acc@5) over the strongest baseline. This demonstrates that RLVR can be successfully adapted to real-world, noisy forecasting, but requires a verifier design that reflects the continuous nature of the problem.

Maximizing training throughput and cost-efficiency of RL for LLMs is essential to democratize this advanced technique. One promising but challenging approach is to deploy such a computational workflow over heterogeneous GPUs. Unlike conventional large-scale LLM pretraining, RL training generally decomposes into three coupled stages, i.e., rollout generation, reward computation, and policy/value updates, which exhibit markedly different compute intensities, memory footprints, and communication patterns. Recent research shows that fully asynchronous RL training can disaggregate these stages across disjoint hardware pools without sacrificing training stability, creating a great opportunity for real-world heterogeneous deployment. To this end, we present AReaL-Hex, a heterogeneity-aware asynchronous RL training system that effectively schedules how to execute rollout generation and policy model training over heterogeneous GPUs while enforcing data staleness bounds. Concretely, we use a two-phase scheduler: (i) a constrained search with MILP to select per-stage parallelization strategies and workload assignments given a resource budget, and (ii) a graph-partitioning step that allocates heterogeneous GPUs and interconnects to maximize end-to-end throughput. Built atop a fully asynchronous RL architecture, AReaL-Hex maps HBM-I/O-bound generation and compute-bound optimization to more cost-efficient resources and balances their producer-consumer interactions to avoid both idleness and stale rollout trajectories. On the mathematical reasoning task with various model scales (1.5B, 7B, and 14B), compared to homogeneous deployments of state-of-the-art asynchronous RL systems: (i) When maintaining the same total budgets, AReaL-Hex delivers up to 1.50x higher training throughput; (ii) When achieving the same training throughput, AReaL-Hex results in up to 1.46x reduction in training cost.

Eigenvalue problems have a distinctive forward-inverse structure and are fundamental to characterizing a system's thermal response, stability, and natural modes. Physics-Informed Neural Networks (PINNs) offer a mesh-free alternative for solving such problems but are often orders of magnitude slower than classical numerical schemes. In this paper, we introduce a reformulated PINN approach that casts the search for eigenpairs as a biconvex optimization problem, enabling fast and provably convergent alternating convex search (ACS) over eigenvalues and eigenfunctions using analytically optimal updates. Numerical experiments show that PINN-ACS attains high accuracy with convergence speeds up to 500$\times$ faster than gradient-based PINN training. We release our codes at https://github.com/NeurIPS-ML4PS-2025/PINN_ACS_CODES.

Structured electronic health records (EHR) are essential for clinical prediction. While count-based learners continue to perform strongly on such data, no benchmarking has directly compared them against more recent mixture-of-agents LLM pipelines, which have been reported to outperform single LLMs in various NLP tasks. In this study, we evaluated three categories of methodologies for EHR prediction using the EHRSHOT dataset: count-based models built from ontology roll-ups with two time bins, based on LightGBM and the tabular foundation model TabPFN; a pretrained sequential transformer (CLMBR); and a mixture-of-agents pipeline that converts tabular histories to natural-language summaries followed by a text classifier. We assessed eight outcomes using the EHRSHOT dataset. Across the eight evaluation tasks, head-to-head wins were largely split between the count-based and the mixture-of-agents methods. Given their simplicity and interpretability, count-based models remain a strong candidate for structured EHR benchmarking. The source code is available at: https://github.com/cristea-lab/Structured_EHR_Benchmark.

Time series generation focuses on modeling the underlying data distribution and resampling to produce authentic time series data. Key components, such as trend and seasonality, drive temporal fluctuations, yet many existing approaches fail to employ interpretative decomposition methods, limiting their ability to synthesize meaningful trend and seasonal patterns. To address this gap, we introduce Seasonal-Trend Diffusion (STDiffusion), a novel framework for multivariate time series generation that integrates diffusion probabilistic models with advanced learnable series decomposition techniques, enhancing the interpretability of the generation process. Our approach separates the trend and seasonal learning into distinct blocks: a Multi-Layer Perceptron (MLP) structure captures the trend, while adaptive wavelet distillation facilitates effective multi-resolution learning of seasonal components. This decomposition improves the interpretability of the model on multiple scales. In addition, we designed a comprehensive correction mechanism aimed at ensuring that the generated components exhibit a high degree of internal consistency and preserve meaningful interrelationships with one another. Our empirical studies on eight real-world datasets demonstrate that STDiffusion achieves state-of-the-art performance in time series generation tasks. Furthermore, we extend the model's application to multi-window long-sequence time series generation, which delivered reliable results and highlighted its robustness and versatility.

The increasing frequency of heavy rainfall events, which are a major cause of urban flooding, underscores the urgent need for accurate precipitation forecasting - particularly in urban areas where localized events often go undetected by ground-based sensors. In Germany, only 17.3% of hourly heavy rain events between 2001 and 2018 were recorded by rain gauges, highlighting the limitations of traditional monitoring systems. Radar data are another source that effectively tracks ongoing precipitation; however, forecasting the development of heavy rain using radar alone remains challenging due to the brief and unpredictable nature of such events. Our focus is on evaluating the effectiveness of fusing satellite and radar data for nowcasting. We develop a multimodal nowcasting model that combines both radar and satellite imagery for predicting precipitation at lead times of 5, 15, and 30 minutes. We demonstrate that this multimodal strategy significantly outperforms radar-only approaches. Experimental results show that integrating satellite data improves prediction accuracy, particularly for intense precipitation. The proposed model increases the Critical Success Index for heavy rain by 4% and for violent rain by 3% at a 5-minute lead time. Moreover, it maintains higher predictive skill at longer lead times, where radar-only performance declines. A qualitative analysis of the severe flooding event in the state of North Rhine-Westphalia, Germany in 2021 further illustrates the superior performance of the multimodal model. Unlike the radar-only model, which captures general precipitation patterns, the multimodal model yields more detailed and accurate forecasts for regions affected by heavy rain. This improved precision enables timely, reliable, life-saving warnings. Implementation available at https://github.com/RamaKassoumeh/Multimodal_heavy_rain

TRISKELION-1 is a unified descriptive-predictive-generative architecture that integrates statistical, mechanistic, and generative reasoning within a single encoder-decoder framework. The model demonstrates how descriptive representation learning, predictive inference, and generative synthesis can be jointly optimized using variational objectives. Experiments on MNIST validate that descriptive reconstruction, predictive classification, and generative sampling can coexist stably within one model. The framework provides a blueprint toward universal intelligence architectures that connect interpretability, accuracy, and creativity.

Sharing sensitive time series data in domains such as finance, healthcare, and energy consumption, such as patient records or investment accounts, is often restricted due to privacy concerns. Privacy-aware synthetic time series generation addresses this challenge by enforcing noise during training, inherently introducing a trade-off between privacy and utility. In many cases, sensitive sequences is correlated with publicly available, non-sensitive contextual metadata (e.g., household electricity consumption may be influenced by weather conditions and electricity prices). However, existing privacy-aware data generation methods often overlook this opportunity, resulting in suboptimal privacy-utility trade-offs. In this paper, we present Pub2Priv, a novel framework for generating private time series data by leveraging heterogeneous public knowledge. Our model employs a self-attention mechanism to encode public data into temporal and feature embeddings, which serve as conditional inputs for a diffusion model to generate synthetic private sequences. Additionally, we introduce a practical metric to assess privacy by evaluating the identifiability of the synthetic data. Experimental results show that Pub2Priv consistently outperforms state-of-the-art benchmarks in improving the privacy-utility trade-off across finance, energy, and commodity trading domains.