2026-11-04

6 repos · 26 papers · 57 scholar

Showing 89 of 89 items

GitHubsynthesis-pnr

eext-run-api-gateway

嘉立创EDA专业版扩展:Run API Gateway — 为 AI 编程工具(OpenCode、QwenCode、KimiCode 等)提供 WebSocket API 网关桥接服务的工具。EasyEDA Pro Extension: Run API Gateway — a tool that provides WebSocket API gateway bridging services for AI programming tools (OpenCode, QwenCode, KimiCode, etc.).

TypeScript10011
GitHubai-hardware

Neuromorphic-Bearing-Telemetry

Asynchronous dynamic-threshold delta modulation front-end and LIF Spiking Neural Network (SNN) for energy-sparse industrial bearing fault diagnostics on CWRU accelerometer data.

Python10
GitHubai-hardware

POSIT-CIM

Undergraduate thesis archive: POSIT arithmetic for SRAM digital computing-in-memory

Verilog00
GitHubai-hardware

riscv-cim-accelerator

RISC-V-controlled digital execution for low-bit compute-in-memory macros, targeting INT8 pointwise convolution on FPGA. Work in progress.

00
GitHubai-hardware

tile-aware-tensor-train-compute-in-memory

Tile-budgeted Tensor-Train rank allocation for PIM / compute-in-memory crossbars: 203 tiles vs 254 (param budget) and 601 (uniform eps) to stay within 1 pt of dense ResNet-18 on CIFAR-10 (91.49%, 698 tiles at T=128). Also DS-CNN keyword spotting: 62 vs 70 dense tiles. TT-SVD, tile cost model, pareto sweeps.

Python00
arXivaccelerators

A RISC-V Simulator for the Tenstorrent Tensix Architecture

RISC-V based AI accelerators have become popular and not only for AI and ML, but also as vehicles for exploring HPC workloads more generally. Tenstorrent are one example, and whilst they document their Tensix architecture in detail and develop their software stack openly, running code has historically required a physical card, and their recently released simulator targets hardware equivalence rather than introspection. In this paper we present Wolfpine, a community developed RISC-V simulator for the Tensix architecture, supporting both the Wormhole and Blackhole generations and integrating seamlessly with the TT-Metal software stack. Constructed entirely from publicly available documentation, open source software and measurements on physical hardware, Wolfpine is written in pure Python for readability and ease of modification. An opt-in, provenance-tracked cost model adds timing fidelity incrementally, and typed event streams, per-component tracing and timeline export enable programmers to identify and diagnose performance bottlenecks that the hardware itself hides.

Nick Brown +1PDF
arXivaccelerators

LoopVQ: Sub 1-bit KV Cache Quantization for Looped Transformers

Looped transformers are a parameter-efficient alternative to standard transformers for language modeling. By reapplying the same layers across multiple loop steps, recent looped language models match the performance of non-looped models several times their size. While this architecture effectively reduces parameter memory, each loop step still stores its own keys and values, so the key-value (KV) cache grows linearly with the number of loop steps and becomes the main memory bottleneck. We observe that the KV cache of looped models is highly redundant along the loop axis and that the loop steps differ substantially in their sensitivity to quantization. Based on these observations, we propose $\textbf{LoopVQ}$, a KV cache vector quantization method for looped transformers that codes channels of multiple loop steps jointly, and fine-tunes the codebooks to match the output of the full-precision model. At 0.5 bits per element, LoopVQ outperforms existing KV cache vector quantization methods for standard LLMs on general, long chain-of-thought reasoning, and long-context retrieval tasks with Ouro and Huginn models. Its design is orthogonal to these methods, and combined with TaSQ, it retains much of the BF16 accuracy even at 0.25 bits per element. We also implement LoopVQ in SGLang, where it nearly halves the decode latency at a 32K-token context and raises the peak serving throughput by up to $4.3\times$ and $6.1\times$ on the 1.4B and 2.6B Ouro models compared to the BF16 baseline.

Minsoo Cheong +1PDF
arXivaccelerators

Provably Learning Mixtures of Linear Experts

Mixture-of-Experts (MoE) models are widely deployed in modern machine learning. However, few convergence guarantees exist for gradient-based training of MoEs: the objective is nonconvex, and even simple mixtures trap first-order methods in spurious local optima. We analyze a canonical gated mixture featuring linear experts and a gate induced by Gaussian clusters in the input space, and give, to our knowledge, the first global convergence guarantee for jointly training the gate and the experts of such a model by gradient EM on its likelihood. Given a logarithmic overparameterization relative to the true number of experts $m$, gradient EM converges from randomly initialized gates and least-squares-initialized experts to any prescribed accuracy in polynomially many iterations, under a separation condition on the gate centers. Our analysis establishes a gradient-dominance inequality for the coupled dynamics by decoupling the loss into per-cluster subproblems. Empirical results validate the predicted $m\log m$ phase transition and suggest that the least-squares warm start required by our analysis is not needed in practice: randomly initialized experts perform nearly as well. We also find that leaving the mixing weights unconstrained concentrates routing onto the true number of experts, whereas load-balancing penalties spread it, so that the fitted model can be pruned to a fraction of its width without loss of accuracy.

Adrian Candocia +2PDF
arXivaccelerators

Shuffled OPSD: When Semantic Consistency Hurts On-Policy Self-Distillation?

Knowledge distillation improves a student model's reasoning ability through supervision from a teacher. On-policy self-distillation (OPSD) uses the same model as both teacher and student under different contexts: the student generates reasoning trajectories from the question alone, while the teacher additionally receives a complete worked example and provides token-level supervision along those trajectories. Because the problem statement is shared, its content and ordering shape both student generation and teacher feedback. Preserving its semantic structure therefore seems natural, but does it always improve learning? We introduce Shuffled OPSD, which shuffles the words of the shared problem statement while keeping the teacher-only worked example intact. Across three mathematical reasoning benchmarks and three Qwen3 model scales, the method improves mean accuracy over both an intact-target control and Standard OPSD under matched training configurations. The largest gain over Standard OPSD is 3.52 percentage points on Qwen3-1.7B. Further controls show that preserving larger word blocks offers no advantage over word-level shuffling at the evaluated scales, whereas shuffling the teacher-only worked example reduces accuracy at every scale. Cross-problem mixing and target-removal experiments further distinguish the roles of problem content, word order, and teacher context. These results show that disrupting the shared problem statement can improve learning, while disrupting the teacher-only reference hurts performance. They challenge semantic coherence as a universally beneficial principle for OPSD and motivate closer study of how input structure shapes student generation and teacher supervision.

Qingyuan Zheng +1PDF
arXivaccelerators

Interpretability Analysis of End-to-End Learned CSI Feedback for Precoding-Oriented Massive MIMO

Recent end-to-end learning frameworks jointly optimize pilot design, channel state information (CSI) feedback, and downlink precoding, achieving strong performance under limited feedback overhead in massive multiple-input multiple-output (MIMO) networks. However, the internal representations and decision-making strategies learned by these architectures remain largely unexplored. This paper presents an interpretability analysis of a precoding-oriented end-to-end CSI feedback framework using explainable artificial intelligence (xAI) techniques. Building on a validated learned architecture, we investigate the structure of the learned pilots, latent feedback representations, and precoding vectors through low-dimensional visualization, density-based clustering, and cross-block representation analysis. The results reveal that the learned system undergoes a transition as the feedback constraint is relaxed. Under stringent feedback budgets, the encoder develops clustered latent representations and quantized precoding behavior. As additional feedback resources become available, the learned representations evolve into smooth geometry-preserving manifolds that retain physically meaningful channel characteristics such as dominant angular information, while enabling more adaptive precoding decisions. Furthermore, the learned pilot signals naturally form complementary sensing patterns aligned with the underlying channel statistics. The analysis also reveals a shift from gain-driven beamforming strategies under severe compression to increasingly interference-aware precoding as the feedback budget increases. These findings provide new insight into how end-to-end learned CSI feedback systems balance compression efficiency and communication performance and establish a general framework for interpreting learned communication architectures.

Roberta Bucchignani +3PDF
arXivmodel-compression

FourSight: Stable FP4 RL Post-Training via Sequence Filtering and Sparse Channel Compensation

Reinforcement learning (RL) post-training increasingly targets low-bit deployment, with NVFP4 weight-and activation quantization (W4A4) becoming the state-of-the-practice format. However, integrating NVFP4 into RL degrades rollout quality and amplifies cross-engine training-inference mismatch (TIM), destabilizing training. We show that these failures are concentrated rather than uniform - severely mismatched tail tokens concentrate in a few outlier sequences, while policy weights remain largely intact and quantization error concentrates in a few high-variance activation channels. Guided by this, we present FourSight, an algorithm-system co-design with two targeted solutions: TIM-aware sequence filtering that excludes drifted trajectories from policy updates, and sparse precision compensation that keeps only top-variance channels in BF16 while retaining native NVFP4 elsewhere. We further design a fused SpliceGEMM kernel with specialized prefill and decode datapaths to minimize overhead. FourSight sustains near-BF16 accuracy and strong FP4 deployment quality across models and engines, while improving step-to-accuracy by 1.76-2.60x and staying within 2.3% of FP4 step time up to 120B-scale models.

Yining Wei +6PDF
arXivaccelerators

Training Specialist LLMs on Your Phone: Forward-Only Optimization via Block-Spectral Coordinates

Large language models (LLMs) on personal devices keep user data local, but limited memory restricts their general capability. Since most user-specific tasks do not require the breadth of frontier models, a practical solution is to adapt a shared on-device model into lightweight specialists. However, backpropagation is prohibitive in memory, and users rarely provide labeled data. We present MobileSpec, a system that trains specialists with forward passes alone and learns from automatically checked task outcomes, as in reinforcement learning with verifiable rewards (RLVR). Because forward-only updates are built from random perturbations, training is efficient only when these perturbations resemble the weight changes of RLVR, which largely preserve the singular values of pretrained weights and rotate their singular vectors. The core algorithm of MobileSpec, FuRA-S, produces such perturbations while training only the block-spectral coordinates of pretrained weights, namely the singular values of their column blocks. As we show theoretically and empirically, rescaling these blocks independently shifts their relative contributions to each weight matrix and thereby rotates its singular vectors. Since this update can be rewritten as a transformation of the layer input, FuRA-S also adapts INT4 models without modifying their weights. With specialists of only 1-5 MB that update 0.015-0.05% of the parameters, MobileSpec raises the success rate of a phone-use agent from 6-23% up to 89-100% and trains a personalized 8B model on an iPhone 16 Pro overnight. On RLVR benchmarks, FuRA-S is competitive with full-parameter evolution strategies (ES) and outperforms ES over LoRA adapters and quantized ES on INT4 models.

Liyan Tan +5PDF
arXivmodel-compression

TriMedium: A Cross-Medium Simulator for Embodied Networking Across Air-Ground-Sea

Embodied intelligent applications have expanded from air and ground into maritime environments, where missions such as coastal reporting require coordination among underwater sensing, surface gateways, and aerial relays. Yet field trials remain costly, and existing simulation tools make it difficult to study how cross-medium communication affects cooperation among these platforms. We present TriMedium, an integrated Air-Ground-Sea simulator for cross-medium embodied networking. TriMedium has three design elements. First, it provides a shared three-dimensional world where aerial, terrestrial, and underwater entities coexist, so underwater agent motion and visual observations stay geometrically consistent with air and ground nodes. Second, it uses a common simulation clock that preserves observation timestamps while scheduling delayed communication and computation, clarifying when the shore receiver obtains each observation. Third, modular quantization and compression interfaces accommodate heterogeneous telemetry and image pipelines, while integrity checks separate intentional lossy distortion from incomplete or corrupted delivery. Three case studies validate TriMedium. Multimodal data delivery tests telemetry and image pipelines, scene construction examines cross-domain spatial consistency, and positioning assesses geometric accuracy under delayed communication. Together, they show TriMedium enables embodied networking by quantifying cross-medium delivery dynamics and fidelity across heterogeneous payloads.

Ruizhe Dang +4PDF
arXivaccelerators

Exploiting Biases in Attention for Transformers

Attention is known to be sparse in vision transformers. In particular, we analyze the unequal relevance of attention-matrix columns to the final computation and show that this phenomenon becomes increasingly pronounced with network depth. This can be interpreted as saying that tokens have an absolute importance for the attention process, independently of queries. Inspired by this, we propose a reformulation of the attention formula that makes this phenomenon explicit by adding an input-dependent bias to the query-key dot product. We show that this new term gives each token a scalar value indicating its anticipated influence on other tokens during the attention process, and can thus serve as a proxy for pruning attention maps. Across multiple transformer model sizes, this allows us to substantially reduce the computational cost of attention, in terms of both FLOPs and inference latency, while retaining most of the original classification accuracy. We publicly released the code and models at https://github.com/EIDOSLAB/AttentionWithBias.

Charles Herr +2PDF
arXivmodel-compression

BtsPruner: Role-Aware Visual Token Pruning for Large Vision-Language Models

Large Vision-Language Models (LVLMs) often process hundreds or thousands of visual tokens, resulting in substantial computational and memory overhead during inference. Existing training-free pruning methods typically estimate token importance using language-model attention, visual attention, or feature similarity. However, these methods treat visual tokens with uniform criteria and overlook that tokens may serve different roles during visual encoding. In this work, we investigate the evolution of visual token roles in LVLMs and observe that tokens progressively transition from local detail carriers to globally referenced representatives as encoding proceeds. Based on this observation, we propose BtsPruner, a training-free role-aware visual token pruning framework that preserves complementary token roles under a fixed budget. Specifically, BtsPruner first identifies the token-role transition point based on the evolution of attention concentration. It then discovers globally referenced representatives from later encoding stages and recursively retrieves local detail tokens from earlier stages to recover complementary fine-grained evidence. Extensive experiments across LLaVA-1.5, LLaVA-NeXT, Qwen-VL, InternVL, and CogVLM demonstrate that BtsPruner achieves superior performance retention under aggressive token reduction. On LLaVA-1.5-7B, BtsPruner preserves 91.6% of the dense model's performance using only 32 of 576 visual tokens (5.6%). In the high-resolution LLaVA-NeXT-7B setting, it achieves up to a 5.95x CUDA-latency speedup.

Zihao Qi +4PDF
arXivother

EvoCG: Self-Evolving Precision Control for Faster Conjugate Gradient

Modern GPUs offer far higher throughput in low precision than in FP64, but exploiting this in conjugate gradient (CG) solvers is difficult because tolerance to reduced precision evolves with the Krylov state. We present EvoCG, a mixed-precision CG solver that treats tile-wise precision as an online control action rather than a fixed setting: a reinforcement-learning policy in the iteration loop selects per-tile precision from the evolving solver state, made profitable by co-design with a Tile-CSC unified mixed-precision SpMV kernel and a low-overhead in-loop runtime. Across 47 SPD matrices from eight scientific domains, EvoCG is the fastest solver overall, with geometric-mean speedups of 2.52x, 3.06x, 2.48x, 1.93x, and 6.99x over cuSPARSE, PETSc, Ginkgo, Mille-FP64, and Mille-Mixed. On the same substrate, it improves time-to-solution over EvoCG-FP64 by 1.28x at only 11.4% more iterations and 1.09% overhead, with FP64-recomputed final true-residual norms nearly identical to those of FP64, and extends to Jacobi-preconditioned CG without retraining, yielding a further 1.15x geometric-mean speedup over EvoCG-FP64.

Jiaqi Zhai +10PDF
arXivmodel-compression

SWAP-INR: Implicit Neural Representation with Switchable Attention Paths for Time-varying Multivariate Scientific Data Compression

Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.

Hyunsoo Son +3PDF
arXivmodel-compression

GC-MoE: Overlapping Expert-Parallel Token Dispatch with FP8/FP4 Grouped GEMM via CUDA Green Contexts

Under expert parallelism, mixture-of-experts (MoE) layers pay all-to-all token dispatch, grouped GEMMs, and a combine step in every layer. In decode each rank holds only 1-256 tokens: dispatch latency and weight traffic dominate while tensor-core utilization is inherently low. In fused operators such as DeepGEMM's Mega MoE, dispatch and computation overlap only by sharing every SM: dispatch warps compete with TMA/MMA and epilogue warps for registers, shared memory, and issue slots; every phase shares one CTA resource budget, synchronization spans the whole SM array, and every replay ends with a workspace-clean pass and a cross-rank barrier. We present GC-MoE, an open-source FP8/FP4 MoE inference operator for NVIDIA Hopper and Blackwell GPUs that uses CUDA Green Contexts to split the device into a small communication partition and a large compute partition. Each role gets its own kernel, resource budget, and synchronization domain, so dispatch and GEMM tune independently. A three-node CUDA Graph (four on Hopper with fused shared experts) launches the dispatch and grouped-GEMM kernels as concurrent roots; a token-ring handshake starts computation on the first filled ring blocks, and double-buffered receive counters remove the trailing cross-rank barrier between replays. The compute partition runs operand-swapped tcgen05 block-scaled MMA on Blackwell and register-A WGMMA with in-register FP4 decoding on Hopper, with fused SwiGLU, requantization, and optional shared experts. On 8 H20 GPUs GC-MoE accelerates DeepSeek-V4.1-Flash decode by 1.188x on average on both the routed and shared-expert suites (peaks 1.296x and 1.353x); on 8 B300 GPUs it accelerates GLM-5.2 and DeepSeek-V4.1-Flash by 1.106x (1.123x with fused shared experts, peak 1.209x), beating Mega MoE at all 54 measured points at matched accuracy. GC-MoE reserves only 20 of 148 SMs on B300 for dispatch.

Yong Li +2PDF
arXivaccelerators

LoopKV: Loop-Aware KV Cache Compression for Serving Looped Transformers

Looped transformers apply a shared block several times per token to improve accuracy without adding parameters. For models such as Ouro that keep a separate key-value (KV) cache per loop, cache memory limits the batch size that amortizes repeated weight reads. Uniform-bit quantization and fixed rules per loop do not allocate precision according to each layer's sensitivity in each loop. Measurements on Ouro show that quantization sensitivity varies across (loop, layer) cells, and the sum of individual cell damages approximates the overall output damage. Sharing damage also varies by cell, with later loops generally tolerating sharing better. Skipping low-attention values to save reads produces errors that compound across loops. LoopKV uses a one-time offline profile to choose 4-bit storage, 2-bit storage or reuse of the final loop's KV for each cell. The planner minimizes reads under budgets for predicted damage and storage. During decoding, each loop reads the high 2 bits of each 4-bit value and fetches the low bits only where its own attention weights require 4-bit precision. A single fused kernel in vLLM executes the plan. Serving Ouro-1.4B on one RTX PRO 6000, the memory-optimized LoopKV-M plan stores 35% less KV than uniform 4-bit quantization (KIVI-4) on the same fused kernel and CUDA graph. It sustains 1.7-2.2x KIVI-4's goodput under a latency service-level objective with a 16K-token maximum length. None of its averages on six public benchmarks falls significantly below BF16's.

Tianle Zhou +5PDF
arXivoptimization

ModelBlaster: Closing the Performance Gap through Learning-Augmented Agents Across Models, Kernels and Hardware Backends

Optimizing accelerator code across hardware architectures is difficult. ML compilers search predefined spaces of schedules and micro-kernels, while LLM-guided optimization can explore alternative algorithms, operator decompositions, and hardware-specific techniques. However, existing LLM agents typically target a single programming interface, toolchain, or accelerator family, focus on kernel-level optimization without model-level transforms, and offer limited reuse of experience across hardware targets. We present ModelBlaster, a Memory-Augmented In-Context Reinforcement Learning (MAIC-RL) framework for optimizing data-parallel kernels and whole models across heterogeneous SIMD and SIMT accelerators. Its mutable policy is a persistent knowledge base updated at inference time by Actor-Critic Textual Gradient Descent (AC-TGD). The knowledge base maps performance states to actions, outcomes, and constraints, enabling experience reuse across kernels, models, and hardware targets. A model-level optimizer uses kernel-level knowledge to inform whole-model decisions. The same optimization loop and knowledge schema support CUDA on NVIDIA GPUs, OpenCL on Qualcomm GPUs, and RVV intrinsics on RISC-V vector cores, with only lowering and measurement layers specific to each target. Compared with PyTorch, ModelBlaster achieves geometric-mean speedups of 1.85x, 5.58x, and 5.18x on KernelBench Levels 1, 2, and 3, and 1.69x on SOL-ExecBench Levels 1 and 2 on a B200 GPU. On Qualcomm Adreno GPUs, it achieves 5.8x over the vendor SDK and 1.7x over Apache TVM; on RISC-V vector cores, it achieves 2.7x over ExecuTorch with XNNPACK across 70 KernelBench problems.

Shengjun Kris Dong +6PDF
arXivoptimization

Online Hamilton-Jacobi Reachability Onboard the Robot Control Loop: A Real-Time GPU Solver for Safety-Critical Autonomy

Hamilton-Jacobi (HJ) reachability provides certified backward reachable tubes (BRTs) for safety-critical robotic systems, but the level-set computation scales exponentially in the state dimension and has historically been performed offline. As autonomy stacks (ADAS/AV, humanoid robotics, and vision-language-action models) increasingly rely on black-box, probabilistic policies, a deterministic safety layer with formal guarantees is an essential complement. We present a CUDA implementation of the canonical WENO5 + TVD-RK3 + Lax-Friedrichs level-set Hamilton-Jacobi-Isaacs solver for 2D, 3D, and 4D dynamics that is fast enough to run inside a robot's control loop. In 32-bit floating point (FP32) on a single workstation GPU (NVIDIA RTX 4000 Ada), it computes the 3D Air3D BRT at 112 Hz, a 4D pursuit-evasion BRT at 61 Hz, and a 2D wind-affected drifter BRT at 74-147 Hz, 5.9-17.5x faster than the JAX reference hj_reachability at matched precision and 1.7-16x faster than a torch.compile reimplementation of the same scheme. In 64-bit floating point (FP64) it agrees with hj_reachability on eight benchmark cases, reproduces Mitchell's Air3D BRT to within 0.04% of the reference cell count, and runs 2.0-4.6x faster. On an embedded NVIDIA Jetson AGX Thor it computes Air3D at 66 Hz (N=41) and the 2D drifter at 54 Hz, 3.8-8.9x faster than hj_reachability on the same device. As a multi-agent application, an in-kernel label-propagation extension recovers the K-agent capture-basin Voronoi tessellation from a single combined HJ solve, independent of fleet size; it runs at 4.0 ms per tick in a closed-loop wildfire-response simulation of six quadcopters under time-varying, non-uniform wind.

Kevin Weekly +1PDF
arXivaccelerators

Anomaly Detection on Resource-Constrained Devices with Reservoir Computing

Anomaly detection at the edge requires the deployment of models that achieve competitive performance, while minimizing computational cost and memory footprint. In this article, we investigate the suitability of the Mahalanobis Distance of Reservoir States (MDRS), an unsupervised time-series anomaly detection method, for resource-constrained devices and propose two optimizations to reduce its computational requirements. Specifically, we replace random reservoir connectivity (MDRS-Rand) with a circular shift (MDRS-Circ) and show that the anomaly score for an entire time series can be computed from the final reservoir state alone. Furthermore, we show that exploiting the reservoir's temporal memory enables more efficient input preprocessing, further reducing computational cost. We evaluate the MDRS variants on the MLPerf Tiny anomaly detection benchmark against the reference autoencoder and other anomaly detection methods, and deploy their inference kernels on a NUCLEO-F401RE microcontroller to assess latency, estimate energy consumption, and measure memory footprint. Among the evaluated methods, only MDRS and the autoencoder meet the benchmark's performance target, with MDRS requiring fewer operations and smaller models. With optimized preprocessing, MDRS achieves an ROC AUC of 0.869, a model size of 24 KB, an estimated energy consumption of 24 uJ per inference, and a latency of 0.32 ms, corresponding to a 75.9x speedup over the autoencoder with standard preprocessing.

Marco Angioli +7PDF
arXivedge-ai

Replica-Aware Split Inference for Edge IoT Systems

Split Inference (SI) addresses the hardware resource limitations of Internet of Things (IoT) devices by distributing tiny Machine Learning (TinyML) model segments across multiple IoT devices. However, a device failure can interrupt the inference pipeline. This paper proposes a replica-aware recovery approach that deploys replicas of TinyML model segments originally hosted on failed devices onto surviving devices. We formulate the joint selection of replica hosts and communication protocols as a mixed-integer linear programming (MILP) problem that minimizes post-reconfiguration inference latency, accounting for computation and transmission latencies, under flash memory and runtime memory constraints. Simulations using MobileNetV2 on heterogeneous IoT networks compare the proposed approach with four baselines. We implement the model in a network of 15 IoT devices with a maximum link-rate reduction of 75%, and show that MILP approach achieves an average modeled end-to-end inference latency that is 6.8% lower than that of Computation Greedy and 18.3% lower than that of Random. Index Terms TinyML, split inference, Internet of Things, fault tolerance, replica placement.

Zied Jenhani +2PDF
arXivother

Positional Encoding is All You Need for Scalable Equivariance Constraint Relaxation

Equivariant neural networks incorporate known symmetries of the data, such as permutations, rotations, or translations, directly into their architecture, often improving sample efficiency and generalization. However, strict equivariance can make optimization difficult, for example, when the data only approximately satisfy the assumed symmetry, or when the equivariance constraints induce a challenging loss landscape. Recent studies have shown that relaxing equivariance during training by introducing non-equivariant dense linear layers can ease optimization in such cases and improve the performance of the resulting strictly equivariant model, after removing the relaxation layers at test time. While effective, this approach incurs a significant computational overhead, limiting its applicability in scalable architectures. In this work, we propose a more efficient constraint-relaxation method based on positional encoding (PE). Our framework injects a positional encoding with learnable scale that breaks the model's original symmetry. This positional signal is gradually annealed to zero, recovering the original equivariant model by the end of training. Empirically, we demonstrate significant improvements for translational and rotational equivariant models with only marginal compute and memory overhead.

Hagay Michaeli +2PDF
arXivaccelerators

Asymptotic Optimality of $cμ$-Rule for LLM Inference Scheduling under Memory Constraint

Large language model (LLM) inference, the process by which a pre-trained model receives an input prompt and generates output tokens sequentially, has become increasingly important with the widespread deployment of LLM-based applications. As inference workloads continue to scale, GPU memory has emerged as a critical bottleneck: the long input prompts and generation sequences common in agentic and reasoning workloads lead to substantial key-value (KV) cache usage. In this paper, we study the resulting online scheduling problem from a queueing perspective. Requests arrive stochastically with heterogeneous prompt and output lengths, undergo prefill and autoregressive decode, and compete for a finite KV-cache capacity. Unlike classical queueing models, the duration of a processing step depends on both prefill computation and KV-cache I/O, so service capacity is jointly determined by computation, memory access, and memory occupancy. We develop a fluid model that reduces the system to a multiclass single-server queue with effective service rates capturing computation, memory access, and memory capacity. This fluid analysis yields a $cμ$ priority rule, where the effective service rate $μ$ is jointly determined by computational speed, memory bandwidth, and memory capacity. We prove that this rule minimizes cumulative holding cost in the fluid model and, for a finite workload, also minimizes completion time and maximizes full-run output-token throughput. We then construct a feasible scheduling policy for the original finite system and prove that its scaled dynamics converge to the optimal fluid policy. Finally, experiments on a real LLM serving engine using an NVIDIA H800 show that practical implementations of the proposed priority rule substantially reduce average latency while maintaining high throughput relative to existing scheduling policies.

Zhenghao Li +2PDF
arXivaccelerators

Spectral Quantum Memory for Implicit Neural Representations

Implicit neural representations (INRs) model signals as continuous functions from coordinates to values, but their performance depends strongly on how spectral capacity is allocated. Yet most Quantum INR (QINR) architectures treat spectral components as independent outputs of separate quantum blocks. We introduce the Spectral Quantum Memory Model (SQMM), a memory coupled quantum INR in which sequential hybrid data reuploading layers (bands) communicate through a shared coherent quantum memory register. Each band reads the memory, computes on its own coordinate upload, and writes part of its state back, so later bands can build on the computation of earlier ones. When memory read and write layers are input independent, we prove that the memory state forms an operator valued Fourier series whose support is contained in the Minkowski sum of the initial memory spectrum and the local band spectra. Downstream bands then condition their Fourier coefficients on the spectral content stored by earlier bands, and the memory coupling creates no new primitive frequencies. This theory motivates a spectral memory alignment loss that encourages memory to encode the spectrum expressed by the deployed prefix of the model. Across audio and image reconstruction benchmarks, SQMM improves over compact QINR and INR baselines.

Daniel A. Moroz-Broitman +1PDF
arXivsynthesis-pnr

qLevelPlace: A Level-Based Placement Framework for RSFQ Circuits

Rapid Single Flux Quantum (RSFQ) logic can operate at ultra-high clock frequencies; however, its physical design is complicated by explicit splitter trees and deep gate-level pipelines in which every logic gate is clocked. To address this challenge, we present qLevelPlace, a level-based placement framework for RSFQ circuits. We separate gate and splitter placement into two stages to better support levelguided gate-to-row assignment and use a timing-aware quadratic cost function to penalize long interconnects. Experiments on the benchmark circuits show that, on average, qLevelPlace achieves a $2.65\times$ higher clock frequency and a 51% reduction in layout area compared with the state-of-the-art RSFQ placer qPlace.

Ziyu Liu +5PDF
arXivaccelerators

GTAC: A Graph-Transformer Actor-Critic for Carrier-Aware MU-MIMO User Grouping

Achieving higher spectral efficiency (SE) in 5G and future 6G networks requires efficient massive multi-user multiple-input multiple-output (MU-MIMO) grouping under coupled physical-layer constraints and strict scheduling deadlines. Motivated by industry efforts in AI-native RAN, we propose Graph-Transformer Actor-Critic (GTAC), a deep reinforcement learning (DRL) architecture for carrier-aware MU-MIMO user grouping. The objective is to maximize the long-term proportional-fair (PF) utility of the cell subject to constraints from the practical stack protocol. GTAC comprises a shared actor--critic encoder with spatial-edge-biased attention mechanism, a pointer decoder actor that generates grouping decisions, and a multi-stage transformer critic that estimates the carrier-wide state-value during training. Training begins by cloning a teacher policy that cannot meet real-time deployment deadlines, followed by critic warmup and two-stage reinforcement learning using proximal policy optimization (PPO), first on a narrow, power-matched carrier and then on the full carrier with a Kullback--Leibler (KL) policy anchor. After two-stage PPO, the best GTAC policy outperforms the non-real-time teacher by 11.6% in carrier SE, while exceeding its PF utility and preserving the cell-edge throughput. A CUDA implementation of GTAC achieves a 69.6 us inference latency on an RTX PRO 4500 GPU, below the targeted 100 us inference budget in practical systems.

Yan Huang +2PDF
arXivoptimization

RoboScope Kit: Unified Interfaces for Fast and Scalable Embodied-AI Evaluation

Simulation benchmarks are the standard way to compare robot policies, but each benchmark brings its own evaluator, observation layout, and action convention, and evaluation is slow: 500 RoboTwin Easy episodes with one simulator and one model on one GPU take 359.7 minutes. Profiling the official StarVLA evaluator shows that simulation occupies CPU cores while the GPU averages 14.3% utilization. RoboScope Kit wraps each model once in a Python class called a Policy, and each benchmark once in an adapter around the benchmark's own evaluator, so the benchmark's rules and scoring stay unchanged. Scheduling runs simulator workers and Policy replicas as separate pools, sizes both from measured probes, and shards work deterministically across GPU instances; runtime optimization renders only the observations an action-chunk Policy reads. Adapters cover six benchmark families. On LIBERO-Long with a StarVLA checkpoint on one 48 GiB RTX 4090, the concurrency search selects twelve simulator workers sharing three Policy replicas, which evaluate 500 episodes 4.45$\times$ faster than the official StarVLA evaluation with equivalent success rates, 2.78$\times$ faster than RLinf, and 1.52$\times$ faster than vla-eval at its tuned configuration. With CUDA MPS, which lets the replicas run their kernels concurrently, the search selects sixteen workers sharing four replicas, 6.16$\times$ faster than StarVLA, 3.85$\times$ faster than RLinf, and 2.10$\times$ faster than vla-eval. On RoboTwin Easy, four GPUs with chunk-boundary observation complete the 500 episodes in 14.4 minutes, 25.04$\times$ faster than the single-GPU baseline, with identical success outcomes and terminal steps.

Deng-Ke Wu +10PDF
arXivaccelerators

EpiLoRA: Fusing Many LoRA Adapters into Low-Precision GEMMs for Serving and RL

Low-rank adapters (LoRA) let one base model serve many fine-tuned variants: multi-tenant engines keep a single copy of the base and batch the requests of many tenants, each with its own adapter. As base models move to 8-bit (FP8) and 4-bit formats, their general matrix multiplications (GEMMs) become cheaper, but the adapters do not: engines run them as separate small GPU kernels beside the GEMM, whose latency takes a growing share of every decoding step. An adapter projects its input down to a low rank and back up, and our key observation is that the GEMM needs the down-projection only in its epilogue, the final stage in which it writes its output. EpiLoRA, a plugin for the vLLM engine, therefore launches the down-projection first and the low-precision GEMM as a dependent kernel that streams its weights meanwhile and adds every tenant's up-projection in its epilogue. On NVIDIA Hopper GPUs, EpiLoRA decodes for a single adapter with up to 24% lower time per token than vLLM's fastest configuration, and faster than SGLang and TensorRT-LLM; with many tenants, its adapter kernels deliver more throughput than vLLM's on most GPU-model pairs we tested. The same low-precision base also trains: a LoRA reinforcement-learning trainer that computes on the engine's FP8 weights needs no 16-bit copy, so a 27B model trains on one 48 GB GPU, and an engine that then keeps activations in 16 bits removes most of the remaining mismatch between the model generating training data and the one trained.

Hieu Tran +1PDF
arXivaccelerators

HyVox: Rethinking Low-Power Voice Intelligence towards Neuromorphic-von Neumann Edge Computing

We present HyVox, a hybrid neuromorphic--von Neumann system for low-energy open-set speaker identification under tight energy budgets. HyVox splits the workload at the speaker embedding. A neuromorphic processor embeds the speech with sparse, event-driven computation, and a von Neumann microcontroller holds the speaker gallery and makes every decision: it scores queries by cosine similarity, rejects unknown speakers, and enrolls new users by writing a prototype, with no weight update. On the neuromorphic processor, a hierarchical temporal abstraction (HTA) SNN backbone learns a synaptic time constant per leaky integrate-and-fire (LIF) neuron, allocating 832 neurons across fast, medium, and slow timescales. A 32-neuron spiking embedding extractor (SEE), pretrained on episodes that each enroll new speakers and reject outsiders, turns the backbone output into an embedding for speakers unseen in training; its 32 spike counts are the only interface between the two processors. To fit the chip's synapse and integer budgets, we compress the backbone 6.5x to 29.0k synapses, train it for integer bit-shift dynamics, and fine-tune it with chip-accurate spiking training (CAST). CAST trains through a differentiable replica of the chip's integer neuron update, derived from its cycle-accurate simulator, so that gradient descent optimizes the exact weights and thresholds the chip runs; it doubles the top-1 accuracy of the mapped backbone on unseen speakers. On 20 Common Voice households of unseen speakers, HyVox reaches 99.6\% top-1 and 5.5\% rejection EER with 2 members and 94.9\% and 13.2\% with 20, and the network deployed on the neuromorphic chip matches its trained model. Each inference on a 3-s window takes 0.96 s and 1.63 mJ of dynamic energy, 42\% lower latency than the fastest embedded ANN baseline and 13.5x less dynamic and 5.8x less total energy than the most energy-efficient one.

Snehasis Addy +3PDF
Scholarsynthesis-pnr

Strategi Public Relations RRI Bogor dalam Membangun Citra Positif di Era Digital Analisis Quality of Service Jaringan Starlink dan Jaringan Seluler untuk Kelayakan Komunikasi Darurat Bencana di Wilayah Terdampak Erupsi Gunung Semeru

Stable internet availability is crucial for disaster coordination, especially in volcanic hazard areas where terrestrial communication infrastructure is vulnerable to physical disruptions. This study evaluates the Quality of Service (QoS) performance of Low Earth Orbit (LEO) satellite internet based on Starlink and compares it with the best available cellular network in regions affected by the Mount Semeru eruption. Using a quantitative experimental approach with a descriptive-comparative design, field tests were conducted at multiple locations. QoS parameters analyzed include throughput, latency, jitter, and packet loss. Results show that Starlink's throughput is higher in several locations, with lower latency (30.80–53.00 ms), and zero packet loss on wired connections. These findings suggest that Starlink is a viable alternative for emergency communication, provided that terminal placement and environmental conditions are carefully managed.

Muhammad Miftahul Karim +20Jurnal JTIK (Jurnal Teknologi Informasi dan Komunikasi)
Scholarsynthesis-pnr

Sistem Informasi Geografis Visualisasi Data Stunting pada Balita Berbasis Web di Desa Sumberwaru

Stunting is a chronic nutritional problem in toddlers that affects physical growth and cognitive development. In Sumberwaru Village, stunting data management is still done manually through Posyandu notebooks and written reports, which causes reporting delays and the risk of recording errors. This study aims to develop a web-based Geographic Information System (GIS) to visualize stunting data in toddlers to support more effective data management, monitoring, and spatial analysis. This study uses a Research and Development (R&D) approach with a Waterfall development model, consisting of needs analysis, system design, implementation, testing, and maintenance. The system was developed using JavaScript and MySQL and tested by 50 users consisting of Posyandu cadres, KPM officers, and village staff. Black-box testing showed that all main functions of the system operated successfully, while usability evaluation using a Likert-scale questionnaire produced an average score of 97.12%, indicating that the system is easy to use and useful for operational activities. The developed system enables faster and more structured management of toddler and stunting data and provides digital map visualization to identify areas with high stunting prevalence.

M. Rofi +20Jurnal JTIK (Jurnal Teknologi Informasi dan Komunikasi)
Scholarsynthesis-pnr

Hybrid CNN-LSTM Model with Distance-Based Loss Function for Leak Localization in Water Distribution Networks

Water loss in distribution networks represents a critical challenge in urban water management, with significant environmental, economic, and public health implications. Although previous studies developed both leak detection and localization approaches, detection has been prioritized. This paper presents a hybrid deep-learning approach for leakage localization with a proposed loss function that minimizes the distance between predicted and actual leak locations in water distribution networks. A hybrid CNN-LSTM model was developed, combining a Convolutional Neural Network (CNN) for spatial feature extraction across the sensor array and a Long Short-Term Memory (LSTM) network for sequence modeling. In addition, a distance-based loss function was designed to optimize both classification accuracy and the physical distance between predicted and actual leak localizations. The algorithm was validated using the L-Town water distribution network benchmark. Results demonstrate that incorporating network distance into training improves localization accuracy compared to a standard cross-entropy loss. Average localization errors of 181 m for background leakages and 58 m for bursts were achieved with the distance-based loss function approach. These results represent an improvement of about 32% for both leak types relative to the standard loss function. The contribution of this paper is two-fold: first, the accuracy of leakage localization is improved through a hybrid CNN-LSTM model trained with a distance-based loss function; second, the search area for leak localization is reduced, as incorporating network distance into training minimizes the physical distance between predicted and actual leak locations. This reduction in search area translates directly to decreased inspection time and costs for water utilities.

Morad Nosrati Habibi +20Journal of water resources planning and management
Scholarsynthesis-pnr

Dynamic Network Provisioning via Reactive Synthesis

Operators of virtualized 5G networks must dynamically assign processing capacity to network flows to meet quality-of-service requirements and use infrastructure capacity efficiently. Provisioning managed by manually tuned configurations and per-deployment heuristics lose accuracy as traffic patterns shift and offer no behavior guarantees under new and unseen workloads. We propose deriving a resource allocator automatically using reactive synthesis techniques drawing from parallels in controller design for cyber-physical systems. We fix a small set of allocation policies and synthesize the rule that selects among them, using a counter-example guided inductive approach, so that the resulting allocator satisfies temporal logic specifications of QoS and utilization as long as given assumptions about traffic and infrastructure hold.

Pooja Sahu +10
Scholarsynthesis-pnr

Web-Based Household Waste Management Information System at Rumah Pilah Sawitri

Household inorganic waste management at the community-based unit Rumah Pilah Sawitri currently relies on conventional paper-based ledgers for daily transactions. This manual mechanism presents four major operational constraints: lengthy transaction handling times reaching 10 minutes per member, vulnerability to mathematical calculation errors (human error), risks of physical record deterioration or loss, and recurring delays in preparing monthly financial and operational reports. To resolve these issues, this study designs and implements a web-based household waste management information system. The software engineering methodology adopts the System Development Life Cycle (SDLC) Waterfall approach, encompassing requirements analysis, system architecture and database design using Unified Modeling Language (UML) and Entity Relationship Diagrams, program implementation using PHP and MySQL, and system verification. Functional validity was evaluated through black-box testing, while usability and acceptance were measured via User Acceptance Testing (UAT) using a 5-point Likert scale across administrators and members. Empirical implementation results confirm that the system reduces average transaction duration by 80%, decreasing from 10 minutes to 2 minutes per member by automating weight-to-price balance conversions. Furthermore, the web platform eliminates calculation discrepancies, enables instant administrative report compilation, and provides members with real-time balance tracking dashboards. UAT evaluation yielded high acceptance scores ranging from 4.7 to 4.9 across ease of use, interface clarity, and functional relevance, validating the system as viable for full operational deployment.

William David Ravelino +20International Journal Software Engineering and Computer Science (IJSECS)
Scholarsynthesis-pnr

Implementation of a Web-Based Inventory System to Support Operations at Hanbai Photocopy

Hanbai Photocopy is a micro-enterprise providing photocopying and office stationery retail services that requires structured inventory management to maintain seamless daily operations. Previously, inventory tracking relied entirely on paper-based ledger records, which frequently resulted in recording discrepancies between physical stock and written logs, slow search retrieval, and time-consuming report compilation. This study aims to design and implement a web-based inventory management system to support operations at Hanbai Photocopy through centralized and automated data administration. The research adopted a qualitative case study methodology, collecting operational data through direct field observation, semi-structured interviews with the business owner, and literature review. System development followed the classical Waterfall lifecycle, encompassing requirements analysis, system design, implementation, testing, and maintenance. The application was constructed using Google Apps Script as the execution platform and Google Sheets as the relational cloud database, enabling a cost-effective serverless architecture without independent hosting requirements. Functional validation was conducted through black box testing across key functional requirements, including master data management, transaction logging, automated stock validation, role-based access control, and activity auditing. The results demonstrate that all planned test scenarios executed successfully, effectively preventing invalid stock deductions and enabling real-time inventory adjustments. The operational deployment facilitates structured stock tracking, mitigates manual recording errors, and streamlines periodic reporting, thereby improving operational efficiency at Hanbai Photocopy.

Nandini Ladivina Sari +20International Journal Software Engineering and Computer Science (IJSECS)
Scholarsynthesis-pnr

Deploying enhanced security for blockchain MFA wireless virtual-key access management for industrial frontiers

The increased adoption of automated devices and its utilization to advance data processing in many facets of societal tasks and demand operations has continued to birth accompanying security challenges. Physical infrastructures today integrates within them automated units to upscale traditional issues. The recent proliferation of automated door access with innovative embedded units employs a centralized design with single-factor authentication, which has proven inefficient for safeguarding restricted environments. Other challenges of data corruption, theft, mishandle of user credentials, all of which can be compromised due to its single point of access control. This study advances enhanced access security for virtual key door infrastructure via a blockchain-MFA strategy. It implements secure, scalable access control with a 3-layered framework of PIN + RFID + Biometric, validated using smart contracts on a blockchain network. The results demonstrated improvements in security, transparency, and auditability compared to traditional door access control, especially against adversarial attacks from single-point failures and credential vulnerabilities. Our prototype unit demonstrates 3-factor authentication for improved security that ensures forensic traceability of every authentication attempt, making it a viable solution for institutions with high-security needs.

Andrew Okonji Eboka +110International Journal of Informatics and Communication Technology (IJ-ICT)
Scholarsynthesis-pnr

Design of a Web-Based Attendance System Using the Agile Method at biMBA AIUEO Klender Unit

Attendance tracking for teachers and students at biMBA AIUEO Klender Unit is currently conducted manually using paper forms, rendering the process susceptible to physical document loss and time manipulation through timestamp camera applications. This conventional approach also poses compliance risks regarding Government Regulation Number 17 of 2025 and Minister of Communication and Digital Affairs Regulation Number 9 of 2026 concerning electronic system governance in child protection. This study aims to design and develop a web-based attendance system that integrates facial recognition technology to prevent fraudulent teacher attendance, while ensuring that student attendance records remain entirely under the supervision and operational authority of teacher accounts. The system was developed using the Agile software engineering framework, combined with qualitative data collection through observation, interviews, and literature study, alongside quantitative descriptive evaluation [cek metode: if the journal explicitly requires the term 'numerical method', replace with: alongside numerical quantitative evaluation] to determine the validity rates of Black-box Testing and user satisfaction levels via User Acceptance Testing. The application was constructed using Laravel, PHP 8.2+, and MySQL, enhanced with geofencing verification, automated WhatsApp Gateway notifications, DOMPDF-based report generation, and Progressive Web App accessibility. Black-box Testing demonstrated that all 18 test scenarios were completely valid (100%), while User Acceptance Testing across user categories achieved an overall average score of 91.6%, categorized as "Very Good." This study confirms that the proposed system effectively eliminates attendance manipulation loopholes, complies with statutory child protection regulations, and improves administrative monitoring and reporting efficiency for biMBA AIUEO Klender Unit.

D. Wijaya +20International Journal Software Engineering and Computer Science (IJSECS)
Scholarsynthesis-pnr

A Completeness-Driven Enterprise Architecture for Integrated Smart-Township Operations Using TOGAF ADM, ArchiMate, and the Zachman Framework

Master-planned townships represent a predominant model of urban expansion in Indonesia, each governed by an operational entity that concurrently coordinates residents, commercial tenants, estate-management teams, and third-party service providers. However, the operational systems supporting these developments remain predominantly fragmented—characterized by estate dues tracked on disparate spreadsheets, resident inquiries and incident reports handled via informal messaging channels, and gate access logged manually—depriving township operators of an integrated, auditable, and data-rich operational platform. Objective: This study establishes a completeness-driven enterprise architecture (EA) framework for integrated smart-township operations. Method: Grounded in Design Science Research, the architecture was developed using the TOGAF Architecture Development Method across Phases A to D, modeled in ArchiMate 3 using the Archi modeling tool, and systematically validated through cross-layer consistency verification, an As-Is versus To-Be gap analysis, and a Zachman 5W1H completeness assessment. Results: The Architecture Vision establishes end-to-end alignment between five stakeholder groups, six strategic operational goals, and their requisite resources. Three core processes—estate dues billing, facility booking with smart access, and resident service with security incident response—are traced from business workflows through application services to underlying cloud and cyber-physical infrastructure, with shared services providing modular integration boundaries. Technology-layer specifications formalize external payment gateways and IoT network connections. The Zachman completeness evaluation indicates that 20 of the 24 cells achieve full coverage, while the remaining four cells (three partial and one weak) converge exclusively within the temporal (When) dimension. Conclusion: The developed enterprise architecture provides an extensible, auditable blueprint for operator-managed townships with an empirically documented coverage profile, wherein the identified temporal gaps establish an explicit empirical trajectory for operational lifecycle scheduling and governance extensions.

Stephen Edlin +40International Journal Software Engineering and Computer Science (IJSECS)
Scholarsynthesis-pnr

S2CSR: Self-Supervised Contrastive Learning for State Recognition in Multivariate Time Series

The recent proliferation of sensors has enabled the collection of extensive time series data from cyber-physical systems. These time series can typically be segmented into distinct states, each carrying rich semantic information. For example, time series recorded during patient rehabilitation may represent sequences of activities such as standing, walking, and sitting. However, annotating these states remains both labor-intensive and costly. Although recent state-of-the-art recognition approaches utilize self-supervised learning to identify such states, many of them overlook informative frequency-domain characteristics, which are often crucial for distinguishing different states. As a result, these methods may fail to accurately identify state boundaries, particularly during periodic state transitions. In addition, existing methods often struggle to effectively capture inter-channel correlations and intra-channel temporal dependencies within multivariate time series. Furthermore, they are susceptible to false negative samples during contrastive learning, which can degrade the quality of learned representations. To address these limitations, we propose S2CSR, a cross-domain contrastive learning framework for multivariate time series state recognition. Specifically, S2CSR jointly performs representation learning and state discovery by introducing contrastive objectives at both the instance and temporal levels. Moreover, the framework encodes time series data in both the temporal and frequency domains, enabling the model to learn more robust and discriminative representations through cross-domain contrastive learning. To further improve representation quality, we design a false negative elimination mechanism to alleviate the adverse impact of incorrect negative samples during contrastive learning. Finally, a nonparametric clustering algorithm is employed to automatically discover latent states. Extensive experiments conducted on six benchmark datasets demonstrate that S2CSR consistently outperforms existing state-of-the-art methods in terms of both effectiveness and robustness.

Jie Zhang +20
Scholaraccelerators

SCARF: A Scene-Adaptive Depth-Guided G-3DGS Encoder Accelerator with Semantic Reuse and Fused Dataflow

Depth-guided, cost-volume-based Generalizable 3D Gaussian Splatting (G-3DGS) encoders reconstruct scenes as sets of Gaussian primitives directly from multi-view images without requiring per-scene optimization, thereby facilitating feed-forward novel-view synthesis on edge computing platforms. However, their shared encoder architecture presents two primary bottlenecks. First, the per-pixel depth search disregards semantic similarity within the 2D feature space, causing depth search operations to account for over 65% of memory traffic due to redundant exhaustive candidate scans. Second, the sequential pipeline of depth prediction followed by Gaussian generation neglects local 3D geometric continuity, resulting in more than 52% of generated Gaussians being redundant prior to instantiation. In this paper, we introduce SCARF, a dataflow-co-designed accelerator that aligns each bottleneck with a corresponding scene-level regularity. To mitigate the first bottleneck, Feature Similarity Depth Reuse reconceptualizes depth search as a problem of semantic locality: depth consistency is determined by affinity in feature space rather than physical address proximity. It transforms exhaustive per-pixel depth searches into anchor-centered local searches. To overcome the second bottleneck, Scene-Adaptive Early Sparsification disrupts the conventional serial depth-to-materialization boundary by probing each tile to infer local scene regularity. It then selectively activates only the minimally sufficient generation pathway, eliminating redundant Gaussian computations before primitive instantiation. On a 27.8 square millimeter 28 nm post-layout design, SCARF delivers a 2.94x geometric-mean end-to-end speedup with maximum PSNR degradation below 0.072%, 2.23 inferences per joule, and 34.5x GPU energy efficiency.

Zi-Rui Ma +70
Scholaraccelerators

FieldRT: Reconfigurable Architecture from Geometric Sensitivity to Streaming Acceleration

Geometric perception on edge must simultaneously achieve two goals that are rarely addressed together: preserve accurate spatial understanding for object localization and operate within stringent latency, memory, energy, and precision constraints. This coupling is particularly evident in oriented bird's-eye-view (BEV) detection, where global context, multi-scale interactions, and geometric fusion affect both prediction quality and hardware dataflow. Conventional deployment flows first compress a model using aggregate quality or complexity proxies and then map the compressed operations to a fixed accelerator interface. This separation risks removing geometric structures that appear redundant under coarse metrics but are critical for precise localization, while leaving hardware-level efficiencies in the retained geometry unexploited. We present FieldRT, an algorithm-hardware co-design methodology that uses a common analysis of geometric sensitivity and data lifetime to derive both the deployment profile and its spatial realization. We deploy FieldRT with Axis-Fusion Network (AFN), a sub-million-parameter oriented-BEV detector, and construct controlled profiles that vary the capacity and organization of its geometric processing. The experiments show that redundancy is task-dependent rather than uniform across metrics. Dense prediction can improve while real instance detection deteriorates; coarse localization may remain stable while strict-IoU precision degrades. Moreover, transformations that are individually safe can become detrimental when combined. FieldRT translates the retained geometric structure into a parameterized streaming architecture that shares field access and storage across spatial operators. The resulting custom-PL primitives achieve production-scale numerical closure on an AMD Kria KV260 and make otherwise infeasible multi-dilation execution realizable.

Hai-Chuan Li0
Scholaraccelerators

A Mixed-Precision CNN-LSTM Accelerator Featuring Long-Kernel Memory Reuse and Dynamic Quantization for Sleep Staging

This letter presents a mixed-precision CNN-LSTM accelerator for EEG-based sleep staging. To efficiently execute long temporal convolutions, a serpentine dataflow propagates activations between neighboring processing elements, reducing memory accesses by 6.2 times and 2.25 times compared with output- and weight-stationary dataflows, respectively. A hybrid quantization scheme applies runtime dynamic quantization to the LSTM and MLP layers and assigns mixed precision on a layer-by-layer basis, reducing model size by 4.16 times. Combined with a bit-flexible MAC architecture and channel-grouped memory scheduling, the accelerator supports INT2/4/8/16 inference. Fabricated in 65 nm CMOS, the accelerator consumes 0.40 mW and achieves a 258 times speedup over an iso-frequency 32b RISC-V processor. Evaluated across all 62 subjects in the public MASS-SS3 dataset using 31-fold leave-two-subject-out cross-validation, the mixed-precision approach with hybrid quantization maintains 5-class accuracy within 0.15% of the floating-point baseline when optimized for general sleep staging and N3 sensitivity within 0.51% when optimized for deep-sleep detection. The developed techniques are broadly applicable to EEG-based DL workloads.

Xin-Yu Chen +50IEEE Solid-State Circuits Letters
Scholaraccelerators

RecIS: Unifying Sparse and Dense Computation for Efficient Recommendation Training

Modern recommender models scale sparse representations and dense computation together. They process thousands of heterogeneous feature fields, long user-behavior sequences, and trillion-scale Embedding parameters, while their dense networks increasingly use large Transformer architectures. The resulting execution mismatch creates fine-grained operations within stages, load imbalance and synchronization delays across devices, and GPU resource bubbles between stages. As models grow, these costs become a primary limit to scalable recommender training. Although existing systems improve Embedding caching and data prefetching, these inefficiencies still coexist in the GPU-centric workloads we study and jointly limit end-to-end throughput. We present RecIS, a unified framework for GPU-centric recommender training. RecIS treats the complete path from input features to dense computation as one optimization target and coordinates three complementary mechanisms. First, adaptive fusion combines fragmented feature transformations with vectorized memory access. Second, a GPU Embedding engine uses hash-based sharding and hierarchical communication to balance request placement and eliminate redundant transfers; it also specializes aggregation for forward and backward index structures. Third, a multistage pipeline overlaps input preparation for future iterations with dense computation in the current iteration. Together, these mechanisms reorganize a fragmented sparse front end into a dataflow that continuously feeds dense computation. We evaluate RecIS on representative industrial recommender models across distributed training platforms. Compared with PICASSO and TorchRec, RecIS reaches up to $2.62\times$ and $1.15\times$ their throughput, respectively, while reducing step latency by up to 61.5\% and 12.9\%. RecIS provides an efficient and scalable training system for large recommender models.

Hua Zong +160
Scholaraccelerators

TADA: Technology Aware DNN Accelerator Performance and Energy Estimator

Deep neural network (DNN) accelerators for edge inference must meet tight energy, latency, and area (EDP) constraints, and their optimal hardware architecture and dataflow configuration is typically identified through design space exploration (DSE). However, on-chip memory technology including cell type, bank organization, and process node, is rarely treated as a first-class citizen in DSE. For example, increasing memory capacity can reduce memory accesses through greater data reuse, but at the cost of higher area and access energy, thereby limiting the compute units budget. Consequently, memory technology can fundamentally shift the optimal accelerator architecture for a given area and EDP budget. Existing machine learning (ML)-based DSE methods require repeating simulation, data generation, and model training for each new memory technology. We present TADA, a memory technology-aware DNN accelerator performance and energy estimator that predicts system latency and energy for a given workload, hardware configuration, and memory technology. TADA is trained on a synthetic dataset that generalizes across memory technologies. The dataset spans current and future technology trends while enforcing physics-based guardrails to preserve realistic area-energy trade-offs. Each memory technology is represented by a technology signature, a vector of read energy, write energy, and storage capacity across a range of areas, provided as input alongside the workload and hardware configuration. Across 21 unseen technology curves spanning SRAM, STTRAM, ReRAM, PCM, FeFET, eDRAM, and CTT, TADA predicts system energy and latency with 1.5-4.4% and 0.7-1.2% mean absolute percentage error, respectively. In three pairwise case studies, it matches the simulator-selected memory technology at 98.7-100% of area budgets, with a mean EDP loss of at most 0.21%, while achieving over 10^4x higher evaluation throughput.

Tanvi Sharma +30
Scholaraccelerators

Energy Optimization for FPGA-Based CNN And SVM Accelerator: A Comparative Design Framework

The research explores various hardware acceleration approaches, beginning with Graphics Processing Units (GPUs) that leverage their massively parallel architecture to accelerate matrix operations fundamental to neural networks. We examine the architectural advantages of GPUs, including their thousands of cores optimized for floating-point operations, high memory bandwidth, and specialized tensor processing units. The analysis extends to Application-Specific Integrated Circuits (ASICs) such as Google's Tensor Processing Units (TPUs), which represent a paradigm shift toward domain-specific architectures optimized exclusively for machine learning workloads. These specialized processors demonstrate significant improvements in performance per watt and throughput compared to general-purpose alternatives. The paper analyses key performance metrics including computational throughput measured in operations per second, energy efficiency expressed as operations per joule, memory bandwidth utilization, and latency characteristics across different acceleration platforms. Real-world benchmarking studies demonstrate that specialized hardware can achieve 10-100x performance improvements over traditional CPUs for common deep learning tasks, with energy efficiency gains often exceeding 1000x for inference workloads. Field-Programmable Gate Arrays (FPGAs) are investigated as a flexible middle ground, offering reconfigurable hardware that can be optimized for specific neural network architectures while maintaining adaptability for evolving ML algorithms. The study examines how FPGAs enable custom data path designs, variable precision arithmetic, and pipeline optimization that can significantly reduce inference latency and energy consumption. Additionally, we explore emerging technologies including neuromorphic computing architectures that mimic biological neural networks, in-memory computing solutions that eliminate the von Neumann bottleneck, and quantum computing approaches for specific ML applications.

Manish Kumar +20IOSR Journal of Agriculture and Veterinary Science
Scholarmodel-compression

Co-JDPRec: Collaborative Joint Dynamic Pruning for Efficient LLM-based Generative Recommendation

Generative recommendation leverages Large Language Models (LLMs) to model item semantics and user preference dynamics through sequence-to-token generation, but deep Transformer layers and wide Feed Forward Network (FFN) dimensions introduce substantial memory and latency overhead in online recommendation scenarios. Existing LLM pruning methods are mainly designed to preserve general language modeling ability, often overlooking the recommendation-specific states shaped by item semantics, collaborative signals, and temporal preference dynamics. To address this limitation, we propose Co-JDPRec, a collaborative joint dynamic pruning framework that shifts the pruning objective from parameter removal to recommendation-state preservation. Specifically, Co-JDPRec introduces a task-aware saliency criterion that combines collaborative alignment and sequence-order sensitivity, enabling the model to identify neurons critical for user-item relations and evolving preference dynamics. Guided by this saliency, Co-JDPRec jointly searches FFN width and layer depth to derive a compact heterogeneous architecture, followed by distillation recovery to restore recommendation performance. Theoretically, Co-JDPRec characterizes structural importance via first-order Taylor sensitivity under recommendation-aware objectives. Empirically, it achieves superior accuracy–efficiency trade-offs on both title-based and semantic-ID-based generative recommendation backbones, validating the effectiveness of recommendation-state-aware pruning.

Shun-Hang Zheng +50
Scholarmodel-compression

Tree-Compatible Personalized Ranking via Linear Latent-Trait Decomposition

Learning-to-rank systems often use a global scorer despite systematic variation in user preferences. We propose a model-agnostic latent-trait framework for probabilistic ranking objectives. A low-dimensional profile activates shared ranking directions linearly, enabling tractable profile inference with nonlinear document-feature effects. Our iterative route encodes all traits as G profile features in one ranker rather than learning or serving G factor models. Ranker fitting reuses an unmodified standard LTR package, and serving uses one model. An offline module probes the ranker to construct an additive surrogate for profile updates across neural and tree rankers. To our knowledge, this is the first continuous user-specific latent personalization under a ranking loss for standard LambdaMART without architecture changes. In simulation, the method outperforms non-personalized and history-encoder baselines, incurs a modest penalty when G is overestimated, and yields probe-surrogate updates that transfer to the trained ranker. In a retrospective offline Message Search evaluation, personalized LambdaMART outperforms its non-personalized baseline. TreeSHAP provides model-relative trait interpretations.

Pavel Metrikov +40
Scholarmodel-compression

The role of radiomics and deep learning in glioma MRI: advances, challenges, and LMIC-tailored opportunities for precision medicine

Glioma is the most common primary malignant central nervous system tumour in adults, with limited access to advanced molecular diagnostics in low- and middle-income countries, like Pakistan. Radiomics and deep learning offer cost-effective, non-invasive alternatives for glioma magnetic resonance imaging characterisation, but their clinical translation remains unoptimised for resource-constrained settings. A structured literature search from January 2018 to October 2025 was conducted across PubMed, Embase, Web of Science, Cochrane Library and Chinese biomedical databases with focus on radiomics and deep learning for glioma magnetic resonance imaging endpoints. Of the 2,132 initial records, 41(1.92%) studies were analysed. The narrative review synthesised advances in radiomics and deep learning for glioma magnetic resonance imaging segmentation, like convolutional neural networks, generative adversarial networks and transformer architectures, preoperative grading (area under the curve: 0.85-0.90), molecular typing, survival stratification and post-radiotherapy response differentiation. Key limitations included small cohorts, lack of external validation, and non-standardised imaging protocols. Radiomics and deep learning enhance glioma precision medicine, but face barriers in terms of standardisation and workflow integration. Prioritising lightweight models and multicentre validation from low- and middle-income countries will accelerate adoption in resource-limited settings, like Pakistan. Key Words: Glioma, Magnetic resonance imaging, Radiomics, Deep learning, Low- and middle-income countries.

Rong-Hong Liao +10JOURNAL OF PAKISTAN MEDICAL ASSOCIATION
Scholarmodel-compression

STLT: A Novel System for Energy-Aware Distributed Inference in the Edge-Cloud-Space Continuum

Low Earth Orbit satellites are pivotal for Earth observation tasks such as wildfire detection, capturing imagery for analysis by Deep Neural Networks (DNNs). On-board analysis reduces downlink volume, but severe energy constraints force a trade-off between computation and data transmission. Splitting the DNN between satellite and ground turns this balance into a design choice, where the split point determines combined computation and downlink energy. Existing approaches typically rely on static split-point decisions that fail to account for dynamic space conditions, or on complex algorithms, historical data, or prediction models that introduce significant overhead. To address these shortcomings, we present STLT, a two-stage system for energy-aware distributed inference between LEO satellites and a ground station. First, an offline analyzer characterizes the DNN architecture to identify a set of favorable split points, consequently reducing the search space. Second, an on-board algorithm performs lightweight energy-aware split selection, dynamically choosing from this pre-computed set based on the current state of the environment. We evaluate STLT on a real-world satellite telemetry dataset, showing that it reduces the split-candidate search space by 30.7% on average, lowers decision latency by 25.1% and energy by 27.5% on average, and delivers up to 20.1% more inference tasks than a state-of-the-art satellite-ground co-inference baseline while consuming 8.8% less energy overall.

Andrija Stanisic +30
Scholarmodel-compression

Evaluating SAM-based and Convolutional Architectures in Thin Retinal Vessel Segmentation: Model Capacity, Computational Cost and Foundation Pre-training

In retinal vessel segmentation, plain Vision Transformers (ViTs) are commonly expected to perform worse than convolutional neural networks (CNNs) on thin vessels, for reasons such as their coarse patch tokenization, their weak locality bias and their need for large training sets. Although this expectation is widely shared in the literature, how plain ViTs actually compare with CNNs on thin vessels has not been systematically quantified. We address this question by comparing a model with a plain ViT encoder (MedSAM), a hybrid Transformer (LiteMedSAM) and three CNNs of different sizes on three retinal datasets (DRIVE, HRF, FIVES) and, as a second-domain check, a microscopy dataset (VessMap). We measured the centerline recall and precision of thin vessels and compared the models while also considering their computational cost and how many epochs they needed to train. We found that the pre-trained MedSAM had a centerline recall of thin vessels that was not significantly different from that of the best CNN on all four datasets (-0.6 to +2.5 percentage points). A compact plain ViT obtained through architecture search but trained from scratch did not retain this performance. For pre-trained plain ViTs, our results do not support the expectation that they miss thin vessels, although similar thin-vessel performance can be obtained with much smaller models.

Carlos Eduardo Fontaneli +10
Scholarmodel-compression

Task-Dependent Performance of GNNs and Pretrained Models in ADMET Prediction

Structure-based prediction of ADMET (absorption, distribution, metabolism, excretion, and toxicity) properties promises fast, low-cost screening of drug candidates, yet model selection remains unclear because published comparisons of molecular representations often use unmatched tuning budgets and leakage-prone evaluation protocols, producing performance differences that may not reflect model quality. To address this gap, we benchmark graph neural networks (GNNs), frozen pretrained models (ChemBERTa, MolCLR), and fingerprint baselines across six Therapeutics Data Commons ADMET endpoints under scaffold-split evaluation. To make it really fair we gave every model family a matched 50-trial, validation-selected tuning budget and reported the median over ten seeds; architecture and hyperparameter selection use validation data only, with the held-out test split reserved for final performance estimation. Under this fair comparison protocol, no model family dominates and the differences between families fall within the seed-to-seed range on most endpoints. The GNN is clearly best only on Caco-2 permeability (R2 ≈ 0.48), and a controlled ablation shows this advantage comes entirely from auxiliary physicochemical descriptors rather than the graph representation; Morgan fingerprints are competitive on hERG (AUC ≈ 0.77) and give the best median on both clearance endpoints, while MolCLR is best on Tox21. For hepatocyte clearance and half-life, all models yield near-zero or negative R2. No single hyperparameter-optimization strategy, among seven evaluated, consistently outperforms random search at this budget, and we show that selecting the optimizer using test-set performance or reporting only single best-trial runs can substantially inflate a favored model’s apparent advantage. Reliable ADMET modeling therefore requires endpoint-specific model selection together with fair, leakage-controlled, seed-averaged evaluation, offering more trustworthy decision support for early-stage drug candidate prioritization.

Martin Stamenov +30bioRxiv
Scholarmodel-compression

Deepfake Generation and Detection Algorithms: A Comprehensive Survey of Multimodal Learning, Robustness, Explainability, and Trustworthy Synthetic Media Forensics

Deepfake generation now combines autoencoders, generative adversarial networks, diffusion models, transformers, neural rendering, and audio-conditioned synthesis. This survey presents a structured narrative synthesis of image, video, audio, and multimodal generation and detection, organised around the relationship between synthesis mechanisms and forensic evidence. Literature selection follows explicit source, search-term, eligibility, deduplication, and evidence-appraisal criteria. Generation methods are compared by conditioning requirements, identity preservation, temporal coherence, and computational cost. Detection methods are analysed through spatial, frequency, temporal, physiological, device, audio–visual, and provenance evidence, with attention to the assumptions that make each cue informative. Evaluation definitions distinguish reconstruction quality, distributional realism, identity and speech fidelity, classification performance, calibration, robustness, localisation, and efficiency. Results are interpreted within their dataset and protocol rather than pooled into an architecture-wide ranking. The comparative analysis separates demonstrated benchmark behaviour from proposed directions in open-set detection and multimodal synthesis. Practical recommendations prioritise held-out generators and datasets, realistic processing pipelines, explicit operating thresholds, calibrated confidence, faithful explanations, and provenance validation. Together, these requirements connect algorithmic performance to the evidence needed for trustworthy synthetic-media assessment.

Teerath Kumar +80Algorithms
Scholarmodel-compression

Human detection in UAV-captured imagery for search and rescue operations: a aystematic review of deep learning and image processing methods

The use of drones in search and rescue operations has become a widely adopted practice because of their availability and ability to capture a large, controlled area in a short amount of time. To increase the efficiency of detecting people using images captured by drones, various image processing techniques and deep learning methods have been utilized. This study aimed to identify and evaluate each of the machine learning methods discussed in recent research and provide insights into their applicability and effectiveness for human detection in diverse aerial scenarios. This study identified a paradigm shift towards end-to-end Convolutional Neural Networks (CNNs). The You Only Look Once (YOLO) architecture has emerged as the prevailing standard, optimally balancing real-time inference speed and detection accuracy, often achieving performances over 90% in precision and recall. While widespread operational deployment remains limited, recent documented successes of lightweight frameworks in active SAR missions demonstrate the life-saving potential of bridging the gap between theoretical deep learning models and practical emergency response.

T. D. de Aquino +40CEAS Aeronautical Journal
Scholaroptimization

GraphCraft: Skill-Centric Curation of Evolving GraphRAG Pipelines

GraphRAG pipelines build entity–relation graphs from text and use them for retrieval and question answering, but extraction noise, entity fragmentation, and unstable clustering can silently corrupt the evidence behind an answer. Correcting these errors is difficult to scale because their dominant forms depend on the corpus and repeated corrections remain manual. We present GraphCraft, an interactive system that records provenance-rich human corrections and compiles repeated corrections into reusable curation skills. Each skill identifies similar errors in subsequent corpus updates and proposes edits for curator review; proposals are previewed rather than applied automatically. The same journal links corrections and skill suggestions back to their graph context and source spans, while bi-temporal graph state distinguishes changes to the stored graph from changes in the underlying world. We evaluate a compiled skill on held-out ingestion batches and demonstrate the complete workflow from diagnosing an erroneous answer, through repairing its graph evidence, to reusing the correction on later data. GraphCraft is released as an open platform with offline deployment. Code and offline deployment: https://github.com/apsolutions/graphrag; demo video: https://vimeo.com/1141894235.

Ilia Karpov +50
Scholaroptimization

MemDTG: Dynamic Temporal Memory Graphs for Evidence-Admissible Agent Memory

Long-term agent memory is an online, non-monotonic retrieval problem. New interactions may reinforce, contradict, supersede, specialize, abstract, or expire earlier memories; the same fact can be a wrong current answer, a correct historical answer, or necessary transition evidence depending on query intent. Existing chunk, summary, and graph memories retrieve by semantic or graph relevance and treat time as post-hoc metadata, so they cannot determine whether a memory is observable, valid, lifecycle-admissible, and provenance-supported before it enters the candidate space. We propose MemDTG, a dynamic temporal memory graph that frames agent memory as temporal evidence admission. MemDTG separates immutable raw episodes from a mutable entity-state temporal graph and a memory-evolution graph over assertion versions. State slots localize mutable user and task states; an LLM-assisted slot-local resolver classifies semantic evolution only within bounded temporal histories; and deterministic commit rules update valid intervals, lifecycle states, and typed evolution edges. At read time, a query compiler instantiates observed-time, valid-time, version, edge-type, and role constraints before ranking, and QV-CTW retrieves evidence paths only inside the query-legal temporal subgraph. Evidence certificates record selected facts, sources, roles, and rejected stale or future alternatives, making retrieval auditable. On LongMemEval and LoCoMo, MemDTG achieves the best LLM-judge accuracy and Citation F1. Ablation and temporal error analyses confirm that provenance certificates and legal-subgraph pre-filtering are each essential for these gains, reducing stale-answer errors by 61.7% and future leaks by 80.1%.

Ya-Xin Deng +40
Scholaroptimization

APG: Agentic Static-Analysis Infrastructure on LLVM-IR Graph Databases

LLM-based coding agents increasingly rely on external program knowledge, yet access to static-analysis graphs alone does not ensure appropriate operation selection or evidence interpretation. We present APG, an infrastructure for semantics-grounded static program analysis. Built on SVF and LLVM, it decomposes compiler-derived program analyses into atomic capabilities that LLM agents can select, compose, and orchestrate for goal-directed, cross-graph investigation, with query safeguards and evidence traces preserving inspectable results and provenance. APG provides 38 atomic capabilities. On analysis requests arising during benchmark repair runs, 35 achieve first-attempt query-generation success rates above 50%, including 13 at 100%. Evaluation on 812 C/C++ defects shows higher repair success than the evaluated baselines. Compared with Passive Graph Push, APG improves first-iteration repair success by 17.6 and 18.0 percentage points on the two benchmark collections and reduces average LLM input-token consumption by 90.8%. Our tool is available at https://github.com/JasonZhongZexin/APG-ICSE-Demo and its demonstration video at https://youtu.be/vkuExZUjGnw

Ze-Xin Zhong +30
Scholaroptimization

FactorSpec: Serving Exact Relational Constraints as Speculative Drafts

Speculative decoding speeds up LLM serving by drafting tokens that the target model verifies in one pass, and its gain grows with the drafts the target accepts. Structured outputs should be an ideal case, because much of a record follows from relations among its fields: an itinerary's days follow from its cities'stays and flights. Serving engines, however, draft such text by proposing it from the prompt or a learned head, and their grammar masks enforce only syntax, so the model must satisfy the relations, which it violates: on NATURAL PLAN's trip problems, Qwen3-8B writes a valid itinerary for 10%, and retries that name the violated relation raise this to 14%. We present FactorSpec, which serves exact relational constraints as speculative drafts. FactorSpec compiles a JSON Schema and its relations into a factor graph with an exact feasibility check, drafts every token the constraint determines as certain, verifies each request at a width chosen for it, and computes the masks in worker processes outside the scheduler while the GPU runs. In SGLang, its drafts yield up to 10.5 tokens per verify step, up to 1.9x those of the fastest n-gram drafter, and it decodes up to 1.43x faster than speculators tuned per load, at a grammar mask's cost; it also leads in vLLM. Because every record is also valid, it serves up to 13.9x the valid records per GPU-second of syntax-only baselines.

Jian-Chang Su +20
Scholaroptimization

CipherFlow: Compiling AI Programs for Efficient Cryptographically Secure Inference

Cryptographically secure inference faces two practical challenges: making secure programming accessible to AI developers and data scientists, and achieving efficient execution across deployment environments. Developers must compose cryptographic primitives correctly, while performance depends on hardware, networks, and evolving ciphertext states. We present CipherFlow, a compiler framework that takes a high-level AI computation graph, available cryptographic primitives, and a target hardware and network configuration to generate an executable, latency-optimized secure inference program. CipherFlow automates cryptographic execution decisions through three techniques: MicroBench profiles deployment-specific primitive costs; state-expanded planning jointly optimizes homomorphic encryption (HE) and secure multi-party computation (MPC) placement, conversions, and ciphertext maintenance under cryptographic-state constraints; and BlockFold summarizes branched regions to scale planning to Transformer graphs. Evaluation using measured HE/MPC backend costs across CPU/GPU and network settings shows up to 16.5$\times$ speedup on BERT-base and over 22$\times$ on ViT compared with static hybrid placement baselines. CipherFlow separates model programming from cryptographic execution decisions, enabling developers to express AI computations while the compiler constructs efficient secure programs.

He-Dong Zhang +20
Scholaroptimization

SparseLift: Automatically Optimizing Opaque Sparse Paths in Industrial-Scale Recommendation Model Training

The sparse path, comprising sparse operations and collective communication, is a critical component of large-scale recommendation models. However, these operations are often represented as opaque nodes in the training computation graph, hindering whole-graph optimization. We present SparseLift, a low-overhead compiler for the sparse path of large-scale recommendation training. It exposes sparse computation and communication to whole-graph analysis to expand the optimization space and improve training efficiency. By analyzing model structure and runtime information, SparseLift reconstructs opaque nodes as analyzable subgraphs, recovering the resource semantics of sparse computation and communication within the graph. Building on the recovered semantics, we design incremental cost models and a constraint-aware planner to automatically select and organize subgraph optimizations into an execution plan that is reused during subsequent steady-state training. We evaluate SparseLift on an industrial video-platform recommendation task with 100B sparse parameters and the public DLRM benchmark. SparseLift improves steady-state throughput for ordinary industrial training steps by 16.5\% over the native baseline. On DLRM, it achieves $1.60\times$ and $1.26\times$ model-execution speedups over DLRM-Baseline and TorchRec, respectively.

Chang-Qing Shi +30
Scholaroptimization

XLA from the Trenches: Correctness on TPU Workloads at Scale

Accelerating modern foundational models has pushed domain-specific compilers toward increasingly complex, multi-tiered optimization pipelines. However, this heavy optimization focus introduces delicate tradeoffs between peak execution performance and correctness. Subtle codegen defects can trigger numerical degradation and memory or buffer corruption, invalidating multi-million-dollar training runs and degrading live serving quality without raising obvious compiler errors. In this paper, we present a battle-tested methodology for verifying and triaging Google's XLA:TPU compiler at hyperscale. We leverage fuzzing techniques, combining synthetic program generation with real production traces, while orchestrating AI agent swarms to autonomously generate stress tests, reproduce failures, and draft fixes. To pinpoint root causes in full workloads, we pair this with a hierarchical differential pipeline that bisects large graphs down to individual failing operations inside fused kernels. Deployed in production, our framework has already driven the resolution of over 142 compiler bugs (with 103 additional findings under review), from which we distill pragmatic lessons learned.

Ignacio Cano +210
Scholaroptimization

When Benign Is Unverified: Label-Signal Mismatch and Abstention-First Triage in Graph-Based Memory Forensics

Machine-learning triage for memory forensics is usually validated on corpora whose benign class is asserted by provenance and never verified. We report a validation study of a graph-based triage pipeline for Windows memory images: Volatility 3 output is compiled into typed heterogeneous graphs and scored by a temperature-scaled GINE classifier and two one-class encoders, fused with rule risk under an abstention gate. On MalVol-25 a transparent rule layer contradicts the benign label on every control (15/15 and 18/18 flagged; at least 11 of 15 paired controls rated HIGH or CRITICAL), and mean injection and command-and-control counts run higher on controls than on infected captures. We call this rule-label discordance label-signal mismatch; it shows the benign class is unverified, not compromised. Leave-one-family-out AUROC over matched pairs is provably three-valued, so a reported mean and standard deviation identify the fold composition. The learned model (0.786 +- 0.311; 9/4/1, scored in-sample by a frozen checkpoint) equals a logistic model over rule-derived attributes (0.786 +- 0.410; 11/0/3) in mean only, and its significance rests on dropping four saturated ties. The fusion path sends 97.7% of snapshots to analyst review, 95.3% with every gate disabled. Two public feature-level corpora show the same defect: on CIC-MalMem-2022 every benign row carries injected regions and one host-population feature reaches AUROC 0.997 under leave-one-family-out validation; on the Sihwail et al. (2021) sample the injection indicator fires on 53% of benign rows. We close with seven reporting artefacts for learned triage on corpora whose ground truth is asserted.

Viraj Chudasama +40
Scholaroptimization

Quantum Intelligence-Driven Dynamic Clustering for 6G Mega-LEO Satellite Constellation Networks

Mega low Earth orbit (LEO) satellite constellations are a key infrastructure for 6G global connectivity, but their high node density, fast orbital motion, and continuously changing inter-satellite topology make routing difficult. Clustering is widely used to reduce routing complexity by partitioning a constellation into routing domains. This paper evaluates, side by side, two quantum-ready clustering approaches for mega LEO constellations. The first is qc-kmeans, a hybrid quantum–classical compressive K-means algorithm that selects cluster centroids by solving small quadratic unconstrained binary optimization (QUBO) problems with the quantum approximate optimization algorithm (QAOA). The second, RS-QUBO, partitions the inter-satellite link (ISL) graph directly with a routing- and stability-aware QUBO that rewards compact, balanced routing domains and penalizes changes between consecutive snapshots. Both are compared with classical baselines under identical inputs, with every trial recorded. For qc-kmeans, selecting the best of repeated trials suggests better cluster quality than K-means, but over all trials the two are practically identical on a 10,687-satellite Starlink snapshot and on a synthetic benchmark; an ablation shows that the outcome is determined by the classical K-means seeding and refinement, while QAOA accounts for most of the runtime. For RS-QUBO, on a real, irregular 1454-satellite Starlink sub-network with time-varying links and 3% random link failures, it is the only method of the main comparison that keeps all routes inside their routing domains at every tested number of domains (0% routing failures, compared with 39–71% for K-means, 16–44% for a grid tiling, and 2–22% for recomputed METIS), with a delay stretch of 1.08–1.18 over the routed pairs and 4–11% of satellites changing domain between snapshots. A component ablation attributes the absence of failures to adapting the partition to the current links and repairing disconnected domains, which also removes the failures of recomputed METIS; the compactness objective lowers the delay stretch relative to cut-based partitioning in the recomputed pipelines at 8 and 32 domains but not at 16. Domain changes are costly for routing: under a link-state update model, RS-QUBO needs 0.6–1.2 times the update load of flat routing, mostly to announce satellites that change domain, whereas an incremental refinement of a static METIS partition is also free of routing failures and needs 0.3–0.4 times the flat load. QAOA solves 6–12-satellite instances of this QUBO better than random sampling in ideal simulation, but the benefit becomes marginal under an ibm_torino noise model, and the compiled circuits require up to 960 two-qubit gates. We conclude that, in the evaluated conditions, the benefit of quantum-ready clustering for satellite routing comes from the problem formulation and the classical pipeline around it, solved classically, and that domain stability matters as much for routing as domain quality; QAOA is demonstrated at small scale on current hardware and noise models.

B. Alamri +20Electronics
Scholaroptimization

SPARQLcirc: Probabilistic Query Evaluation with Provenance Circuits

Probabilistic query evaluation (PQE) computes the probability of every answer of a query over a tuple-independent probabilistic knowledge graph. We study exact PQE for SPARQL over graphs accessed through SPARQL 1.1 endpoints. Shared triples among the derivations of an answer make exact PQE #P-hard in general. Knowledge compilation of the Boolean provenance of every answer yields exact probabilities, and the provenance formulas of different answers can share subformulas. We propose SPARQLcirc, which rewrites a SPARQL query into CONSTRUCT queries that name every gate by its content, so that the results of independent queries form one Boolean provenance circuit for all answers. The rewriting covers joins, unions, projection, filters, OPTIONAL, difference, and transitive closures. SPARQLcirc compiles the circuit and computes the exact probability of every answer by weighted model counting. For the evaluation, we draw random probabilities for the triples of WatDiv and gMark and the statements of Wikidata and evaluate the CONSTRUCT queries on GraphDB, Fuseki, and Oxigraph. The ratio of the end-to-end time of PQE with SPARQLcirc to that of PQE by compiling the provenance expressions of the SPARQL provenance rewriting NPCS decreases with the number of derivations.

Jing-Cheng Wu +30
Scholaredge-ai

Advanced IoT and ML-driven personal safety technology for women

This article introduces VigilNet, a personal safety system powered by internet of things (IoT) and machine learning (ML) that identifies environmental irregularities instantaneously through intelligent wearable gadgets. In contrast to solutions dependent on extensive datasets and reactive alert systems VigilNet employs a few-shot learning (FSL) strategy utilizing siamese neural networks (SNNs) to achieve precise anomaly recognition, with limited training data. In the information and communication technology (ICT) realm, VigilNet is a smart IoT sensing and communication network with interconnected wearable sensor nodes continuously collecting physiological and environmental data for real-time processing. The SNNs based FSL module then fuses these multiple modal streams of sensor information and extracts the similarity-based threat assessment, which in turn drives an intelligent decision support system to alarm the operator to take timely action for any abnormal situations. VigilNet, thus, combines IoT connectivity, real-time data analytics, and machine-learning-driven decision support into a single personal safety system. The framework was tested on an IoT wearable dataset encompassing heart rate, body temperature, ambient noise and proximity data. Experimental results demonstrate that VigilNet achieves a threat detection rate (TDR) of 0.85, false alarm rate (FAR) of 0.10, and a response time (RT) of 3.5 seconds, outperforming conventional IoT–ML safety models. These findings validate the feasibility, efficiency, and robustness of the proposed architecture and confirm the potential of FSL-based techniques for next-generation women’s safety technologies.

P. Divya +20International Journal of Informatics and Communication Technology (IJ-ICT)
Scholaredge-ai

Innovation in Smart Sensor Technology for Automated Palm Sugar Cooking Control to Improve Production and Sales at Harapan Palm Sugar Producers

Palm sugar is an important natural sweetener produced by many small-scale home industries in Indonesia. However, conventional cooking processes rely heavily on manual observation, resulting in inconsistent temperature and viscosity control, inefficient energy utilization, and variations in product quality. This study aims to develop an Internet of Things (IoT)-based smart sensor system that integrates temperature and viscosity measurements for real-time monitoring and automatic control of the palm sugar cooking process. An applied experimental engineering approach was employed, consisting of requirement analysis, hardware and software development, sensor calibration, system integration, laboratory testing, and field evaluation at a partner home industry. The proposed system integrates a DS18B20 temperature sensor, viscosity sensor, ESP32 microcontroller, automatic heating controller, and web-based monitoring platform. Experimental results showed an average temperature measurement error of 0.26% and viscosity measurement error of 0.98%. The IoT communication system achieved a 99.2% successful data transmission rate with an average delay of 1.9 s. Compared with the conventional process, cooking time decreased from 122 to 101 min, fuel consumption decreased from 20 to 16 kg per production cycle, temperature stability increased from 72% to 96%, and defective products decreased from 12% to 3%. These results demonstrate that integrating temperature and viscosity sensing with automatic IoT-based control improves process stability, energy efficiency, and product consistency. However, the study was limited by single-site field validation and predefined control thresholds. Future research will focus on adaptive control using machine learning or fuzzy logic, multi-site validation, cloud analytics, and mobile notification features to improve system scalability and intelligence.

J. Hutagalung +20International Journal Software Engineering and Computer Science (IJSECS)
Scholaredge-ai

Integration Of Machine Learning And IOT Smart Traps For Fruit-Fly Monitoring In Gerga Citrus Orchards Of Pagar Alam

Fruit-fly infestation is an important constraint in Gerga citrus production because it can reduce fruit quality, yield, and economic value. Conventional monitoring based on manual trap inspection is labor-intensive and provides limited real-time information for pest-management decisions. This study proposes an Internet of Things (IoT) Smart Trap integrated with machine learning for automated fruit-fly detection, counting, and population monitoring in Gerga citrus orchards in PagarAlam, Indonesia. The proposed system integrates image acquisition, environmental sensing, edge computing, YOLO-based object detection, EfficientNet-B0 verification, DeepSORT-based tracking, cloud data transmission, and XGBoost-based population prediction. The research follows a CRISP-DM-based workflow covering data acquisition, preparation, modeling, evaluation, and deployment. The evaluation showed that the proposed YOLO–EfficientNet-B0 framework achieved a precision of 97.1%, recall of 95.6%, F1-score of 96.3%, mAP@50 of 97.8%, and mAP@50–95 of 81.4%. Thetracking evaluation achieved a counting accuracy of 95.2%, while the XGBoost model obtained an MAE of 3.21, RMSE of 4.56, and R² of 0.89. These results indicate the potential of combining automated image detection, environmental sensing, tracking, and population prediction within an IoT Smart Trap for continuous fruit-fly monitoring. The proposed framework can provide a foundation for early-warning systems and support more precise Integrated Pest Management (IPM) decisions in Gerga citrus orchards.

Inka Rizki Padya +20
Scholaredge-ai

Computational Botany: A Review of Machine Learning, IoT, and Bioinformatics in Plant Sciences

The intersection of computer science and botany has catalyzed a paradigm shift in plant taxonomy, pathology, and agricultural management. This paper reviews the current literature on computational applications in plant sciences, focusing on three primary domains: Deep Learning for automated disease detection, Internet of Things (IoT) sensor networks for physiological monitoring, and high-performance computing in plant genomics. By synthesizing recent methodologies, this review identifies existing technological bottlenecks—such as dataset biases and processing latencies—and outlines future trajectories for integrating artificial intelligence into sustainable botanical research and precision agriculture.

Nishchal Sharma +10International Journal for Research in Applied Science and Engineering Technology
Scholaredge-ai

Lightweight Intrusion Detection for Resource-Constrained IOT Devices Using Machine Learning

Billions of low-cost Internet of Things (IoT) devices run on microcontrollers with kilobytes of memory and no room for conventional intrusion detection, which has made them easy targets for botnets such as Mirai. This paper presents the design of a lightweight, machine-learning-based intrusion detection system (IDS) that runs on the IoT device itself. A hook in the device's TCP/IP stack updates a few counters for every packet in constant time; every 2 s these are turned into 20 traffic and host features, using 64-bit bitmaps to estimate distinct addresses and ports. A depth-limited decision tree or an 8-bit quantized multilayer perceptron classifies each window as benign or malicious, and a 3-of-5 vote across consecutive windows decides whether to block the source and raise an MQTT alert. An optional gateway tier names the attack type. A review of 16 related studies shows that most machine-learning IDS work is evaluated offline on datasets, uses features that a microcontroller cannot compute, and rarely reports how models generalise across datasets. Design analysis shows that the multilayer perceptron needs 490 parameters and 464 multiply–accumulate operations per window, a full depth-8 tree at most 2.6 KB, and the feature state under 0.3 KB, all within the budget of a Class 1 constrained device; the voting rule reduces expected false alarms from 43 to under 0.001 per day at a 0.1% per-window false-positive rate, assuming independent windows. A validation protocol using TON_IoT, CICIoT2023, Edge-IIoTset and N-BaIoT, including cross-dataset tests and on-device measurements, is provided

Rakhi Kumari +30International Journal for Research in Applied Science and Engineering Technology
Scholaredge-ai

Artificial Intelligence in Smart Cities: Intelligent Solutions, Applications, Challenges, and Emerging Research Trends

Artificial Intelligence (AI) has emerged as an important technology for developing smart cities by improving urban services, resource management, public transportation, energy efficiency, environmental monitoring, and citizen services. The increasing population of urban areas creates challenges related to traffic congestion, pollution, waste management, electricity consumption, and efficient public administration. Traditional city management systems often depend on manual monitoring and separate information systems, which can limit real-time decision-making and coordination. This paper presents a structured review of AI applications in smart cities and proposes an integrated AI-based framework for intelligent urban management. The proposed framework combines the Internet of Things (IoT), machine learning, data analytics, cloud computing, and intelligent decision-support mechanisms to collect, process, and analyze urban data. It focuses on traffic management, smart energy systems, waste collection, environmental monitoring, and citizen services. The paper also discusses the benefits, implementation challenges, limitations, and future research opportunities associated with AI-enabled urban development. Existing research indicates that AI can support more efficient and responsive city services; however, data privacy, cybersecurity, infrastructure costs, interoperability, algorithmic bias, and digital inclusion remain important concerns. The proposed framework provides a conceptual foundation for developing sustainable, intelligent, and citizen-centered smart cities.

P. A. W. Yash Yogesh Harale +10International Journal of Advanced Research in Science, Communication and Technology
Scholaredge-ai

Artificial Intelligence and the Internet of Things in Education: A Scientometric Mapping of Growth, Thematic Evolution, and Intellectual Structure in a Decade of Scopus-Indexed Research

The convergence of artificial intelligence (AI) and the Internet of Things (IoT) has been advanced as the technical foundation of adaptive, sensor-rich learning environments, yet the structure and maturity of the resulting literature have not been mapped systematically. This study reports a scientometric analysis of 824 Scopus-indexed documents published between 2016 and 2025, retrieved with a TITLE-ABS-KEY query combining AI, IoT, and education and screened in accordance with the PRISMA 2020 statement. Performance analysis and science mapping were conducted with the bibliometrix R package (Biblioshiny) and VOSviewer, and comprised source, authorship, and country analyses, keyword co-occurrence, a strategic thematic map, thematic evolution across three periods, author co-citation, and a logistic life-cycle model. The corpus was produced by 2,668 authors across 525 sources, and annual output rose from a single document in 2016 to 274 in 2024. Publication is markedly dispersed: computing proceedings series and engineering outlets, rather than education journals, head the Bradford core; conference papers account for 47.1% of research documents; and author productivity departs sharply from Lotka’s law (β = 5.07), with 95.4% of authors contributing a single document. India and China dominate corresponding authorship, whereas international co-authorship remains modest at 20.87%. Artificial intelligence, the Internet of Things, and machine learning form a single central keyword cluster, while educational terms occupy peripheral clusters. The strategic map contains basic and niche themes but no motor theme, and thematic evolution shows continuous lineages for AI and IoT but none for pedagogy. Author co-citation reveals an intellectual base drawn from Industry 4.0, cyber-physical systems, and machine-learning research, with education-specific authors at its periphery. A logistic model signaled saturation (K = 869), but sensitivity analyses showed this to be an artifact of the partial 2025 year; on complete or annualized data, the field is at or just before its inflection point. The field is therefore expanding but unconsolidated, technologically driven, and reliant on an intellectual base imported from outside education. Implications for research agendas, publication strategy, and capacity building are discussed.

Phanuwat Ruangkulsap +50Higher Education Studies
Scholaredge-ai

Privacy-Preserving Deep Learning: Differentially Private Neural Network Training Against Training-Data Inference

This paper addresses the critical security challenge of protecting sensitive training data in deep neural networks against sophisticated adversarial attacks. We present a comprehensive framework that enables privacy-preserving machine learning while maintaining model utility, focusing on defense mechanisms against white-box adversaries with full access to model parameters and training procedures. Our approach integrates differential privacy into the stochastic gradient descent optimization process through a privacy accounting mechanism that tracks higher-order moments of privacy loss, yielding substantially tighter privacy guarantees compared to traditional composition methods. The framework implements gradient norm clipping to bound individual training examples' influence and adds calibrated Gaussian noise to sanitize parameter updates, creating a provably secure training pipeline. We evaluate our methodology on standard image classification benchmarks, demonstrating that deep neural networks can achieve high accuracy while maintaining strong privacy protection. Experimental results show that models trained under this framework exhibit minimal overfitting, with training and test accuracy gaps remaining small, a direct consequence of the privacy-preserving regularization effect. Systematic parameter sensitivity analysis reveals the interplay between privacy budgets, noise levels, batch configurations, and model architecture choices. The implementation provides a practical privacy-preserving training pipeline built on a widely used machine learning framework, offering organizations a deployable approach to training neural networks on privacy-sensitive datasets with a formal differential privacy guarantee against white-box inference of training records. This research establishes foundational implementation and evaluation techniques for privacy-preserving deep learning applicable to security-critical domains including healthcare, finance, and mobile device applications where training data confidentiality is paramount.

Pramod Prakash0
Scholarai-hardware

A Theoretical Framework and Efficient Algorithm for Computing Condensed Multi-Level Bisimulation Graph Summaries

Summary graphs have shown to be a valuable tool in graph analysis and processing, with applications in query optimization, retrieval, and visualization, among others. Many summarization approaches are based on the concept of stratified bisimulation, which determines vertex equivalence based on outgoing paths of length k. Summaries based on bisimulation typically also store information about all paths of length less than k, and thus explicitly encode the 1, 2, ..., k bisimulations. To this end, we introduce condensed summaries: a formal model for representing stratified bisimulations in a compact manner. By leveraging repeating structures between consecutive bisimulation levels, condensed summaries represent all levels of bisimulation up to some k, while containing O(|V|) vertices, independent of k. We also prove that condensed summaries are equivalent to an established, uncondensed representation. Furthermore, we provide an algorithm for computing condensed summaries, along with an optimization that can yield a 20-fold speed-up. To demonstrate the efficacy of our model and algorithm, we provide experimental results on various edge-labelled graphs with up to 28~billion edges and a bisimulation depth of up to 4,757. All experiments successfully reached full bisimulation in less than 10 hours, while those with up to 1~billion edges only took minutes or even seconds, with no more than 64 GiB of memory. The resulting summaries were all substantially smaller than the original graphs, sometimes with a 95% reduction in vertices and edges. Code is available at: https://github.com/R-van-Bakel/Multi-Summaries.

Ruud van Bakel +40
Scholarai-hardware

Curvature Graph Mergeability: Predicting Model Compatibility from Per-Layer Hessian Geometry

Merging independently trained neural networks --- via weight averaging, model soups, or task-vector arithmetic --- is increasingly central to large-scale model reuse, yet the success of any merge hinges on whether two parameter vectors occupy the same loss basin, a property currently certified only by expensive multi-point interpolation barriers. We introduce Curvature Graph Mergeability (CGM), a scoring framework that predicts the interpolation barrier from only cheap, per-layer Hessian curvature features at each endpoint. CGM extracts two scalar curvature statistics per layer using Hutchinson's trace estimator and the power method, organizes them into a chain-structured factor graph whose latent compatibility states encode qualitative regimes of layer alignment, and computes a mergeability score from the negative variational free energy of the chain. For deep linear networks under balanced factorization, we prove that the interpolation barrier is upper-bounded by a multilinear function of per-layer curvature mismatch, and we identify a diagonal-dominance condition under which the score is monotone in mismatch; the bound is stated as an implicit joint bound on the per-layer mismatch vector, with the regime of validity made explicit. Across four meta-training architectures and four held-out OOD architectures (ResNet-110, ViT-B/ImageNet, WRN-28-10, 3-layer LSTM), CGM attains Spearman ?=0.86 on leave-models-out splits and ?=0.81 pooled OOD, outperforming zero-order distance, CKA, Fisher overlap, REPAIR, midpoint-barrier, and architecture-agnostic curvature-MLP baselines at Holm--Bonferroni-corrected p<0.01. It is approached only by Git Re-Basin + barrier (?=0.90), which costs ~14,500× more per pair; CGM closes the gap to permutation-based merging at 35× amortized speedup over direct 11-point barrier when screening 50 candidates, with negligible memory overhead. Per-layer posterior marginals additionally pinpoint specific layers that act as merge bottlenecks.

Ze-Hua Cheng +20
Scholarai-hardware

Resource-Efficient Generative AI: Rethinking Knowledge Management and Delivery

Generative AI is increasingly embedded in information retrieval, knowledge management, and recommender systems, yet its adoption is constrained by finite compute, memory, latency, energy, and data-governance budgets. This tutorial presents a resource-aware and sovereign perspective on generative AI for scalable knowledge access, framing efficiency as a multi-objective optimization over accuracy, latency, and cost. We structure the tutorial around: intelligence building, which addresses training and adaptation under explicit resource budgets, and intelligence delivery, which focuses on serving-time inference under strict tail constraints (e.g., P99 latency, peak memory) typical of large-scale retrieval and decision-support pipelines. Core techniques, including parameter-efficient fine-tuning, sparsity and pruning, quantization, distillation, mixture-of-experts routing, and speculative decoding, are analyzed as controlled approximations that trade utility and predictive uncertainty. The tutorial equips participants with methods and deployable design patterns for building cost-efficient, privacy-preserving generative AI systems aligned with CIKM's core themes of scalable, trustworthy knowledge access under real constraints.

João Gama +10
Scholarai-hardware

LuWu: Adaptive Layer-wise Heterogeneous KV Cache Quantization

The KV cache is essential for efficient autoregressive inference in Large Language Models (LLMs), but its memory cost increases with context length, layer depth, and request concurrency, limiting long-context and high-throughput serving, particularly in Computing Power Network (CPN) environments. KV cache quantization reduces this cost by lowering the storage bit-width of keys and values. However, existing methods often rely on uniform precision or static mixed-precision policies, which cannot fully capture layer-wise differences in quantization sensitivity or decoding-time changes in precision requirements. This paper proposes LuWu, an adaptive layer-wise heterogeneous KV cache quantization method. LuWu first performs offline calibration to measure attention-output errors under different K/V precision pairs, clusters layers with similar error response patterns, prunes unsuitable candidate configurations, and searches an initial heterogeneous precision configuration under accuracy and bit-width objectives. During online inference, LuWu uses the Layer Activity Score (LAS) to adjust the precision of selected layers, while a hierarchical residual-code representation supports efficient promotion and demotion between precision levels. Experiments on representative reasoning, long-context, and language modeling tasks show that LuWu achieves a favorable accuracy--bitwidth trade-off. Compared with representative state-of-the-art K8V8 baselines, LuWu reduces the average equivalent bit-width by 29.5%--38.1% while maintaining competitive inference accuracy. These results highlight the effectiveness of adaptive layer-wise heterogeneous quantization for memory-efficient LLM inference.

He-Xuan Lv +20
Scholarai-hardware

BPB-GNN: Resource-Efficient Distributed GNN Training via Boundary-Aware Prefetching and Cost-Balanced Batch Construction

Distributed graph neural network (GNN) training is essential for large-scale graph learning, but existing mini-batch frameworks still suffer from substantial inter-partition communication overhead and indirect, coarse-grained GPU memory control. In distributed mini-batch training, remote sampled neighbors cause frequent feature fetching, while fixed seed-node batch sizes cannot accurately reflect the computation graph expansion induced by multi-layer sampling. This paper presents BPB-GNN, a resource-efficient distributed GNN training system that jointly optimizes inter-partition feature access, hierarchical storage management, and mini-batch construction. First, BPB-GNN introduces a training-centric boundary-aware prefetching strategy to localize high-value remote neighbor features associated with boundary nodes in training-node receptive fields. Second, it uses a DRAM/NVMe SSD hierarchy with priority-aware placement and dynamic migration to manage prefetched features under limited DRAM capacity. Finally, it constructs seed-node batches according to estimated sampling-induced load costs rather than equal seed-node counts, thereby balancing batch-level workloads and reducing peak GPU memory usage and its inter-batch fluctuation. These resource-efficient mechanisms may also improve the efficiency of large-scale GNN computation in computing power networks and JointCloud environments. Experiments on real-world graph datasets show that BPB-GNN achieves up to 5× speedup over the baselines while reducing peak GPU memory usage by up to 36%.

Zhi-Chao Chen +40
Scholarai-hardware

ML Inference Acceleration via Scalable In-network Computing

Modern inference engines have significantly advanced the scalability of large models through innovations in serving techniques, proving instrumental in driving the widespread adoption of AI. However their true potential is constrained by traditional server architecture where reliance on host CPUs and operating systems introduces unnecessary overhead and degrades performance. We identify a significant opportunity to eliminate these bottlenecks by offloading data management directly to the network layer by proposing a unified architecture based on our previous research and ongoing work using programmable Data processing units (DPUs)/SmartNICs for in-network computing. Our previous work, NICTokenizer is a framework that offloads the initial tokenization phase to programmable network dataplane eliminating the host OS Kernel overhead lowering the overall tokenization latency by 61.3% and increasing the token throughput by 407.62% compared to conventional server based implementations and was previously published as a poster at SIGCOMM' 25. Alongside we introduce CacheFlow a disaggregated prefix cache pool which utilizes the DPU memory and GPUDirectRDMA for efficient sharing, storage and retrieval of KV cache across a cluster. CacheFlow reduces the average Time to first token (TTFT) by 82% imporves overall throughput by 39% and decreases memory storage requirements by 44% compared to standard inference engines like vLLM. As a natural progression our future research explores offloading semantic routing to a DPU enabling the network plane to dynamically select the appropriate model experts for each incoming sequence.

Shaunak Galvankar0
Scholarai-hardware

GSPARC: A Mode-Aware GPU-Based Method for Large-Scale Sparse Tensor Contraction

Sparse tensor contraction (SpTC) is a core operation for high-order sparse data analytics in recommendation, knowledge graph analytics, large-scale tensor mining, and scientific computing. Existing methods are limited by SpTC-misaligned representations, overlapping outputs across partitions, and memory-inefficient accumulation over large output spaces. We propose GSPARC, a GPU-based SpTC method that combines a new mode-aware bit-wise representation with mode-decoupled output partitioning and memory-efficient accumulation. These techniques improve index matching, eliminate inter-partition merging, and avoid full output-space allocation. Extensive experiments demonstrate that GSPARC outperforms state-of-the-art SpTC methods in contraction speed and scalability. Our code is available at this link: https://github.com/seohyeon0322/GSPARC.

Seohyeon Kim +30
Scholarai-hardware

Learn to Compress or Compress to Learn? Shannon and Turing meet again in the LLM era

The year 1948 gave birth to two transformative pillars of modern computing: Claude Shannon's information theory, which gave rise to data compression, and Alan Turing's writings on intelligent machinery, which ignited the field of artificial intelligence. Today, nearly eight decades later, we are witnessing a profound hybridization of these once-distinct fields. As we push to efficiently store, process, and retrieve massive datasets across diverse data types, users, devices, and systems — from vector and graph databases to cutting-edge agentic memories — AI and compression can no longer operate in silos. In this talk, we investigate how Learning is fundamentally reshaping the fields of Indexing and Compression. We will focus on two challenging research directions: designing Learned Data Structures that fuse classical indexing with machine learning to achieve unprecedented efficiency, and developing next-generation Data Compressors driven by powerful autoregressive and diffusion neural models to reach new frontiers of efficacy. We will also examine the reverse direction — how Compression can reshape Learning through model compaction and comprehension. By unpacking the distinct computational advantages and current practical pitfalls of these lines of investigation, we will outline a roadmap for these research frontiers and hopefully inspire new cross-disciplinary collaborations.

P. Ferragina0
Scholarai-hardware

Parallelism-Aligned Pruning and Round-Aware Context Reuse for On-Device Multimodal Agents

On-device multimodal agents are limited by two deployment mismatches: pretrained attention structures may not divide evenly across the compute cores of an edge accelerator, causing irregular partitions; and standard prefix caching keeps historical visual tokens in the persistent KV cache, whose capacity grows with the number of rounds. We present PARC, which combines parallelism-aligned structured pruning with lifetime-aware KV-cache reorganization. At the model level, it uses first-order Taylor importance and grouped-query-attention-aware selection to make attention dimensions divisible by the target core count while preserving the query-to-key/value ratio and head dimension. At runtime, it separates the system prompt, persistent text history, and transient visual observations by lifetime, updating text history during tool execution so that historical images leave the persistent cache and more interaction rounds fit a fixed budget. On 2,062 GUI actions, the incremental interface shows no accuracy regression, and task-aware recovery training restores 98.7% of the action-type accuracy gap and the coordinate hit rate to the original level. Under seven-core deployment, hardware-aligned pruning reduces vision-encoder latency by 11.4%, improves decoding throughput by 23.8%, and lowers per-heap high-water memory by 12.7%; a parameter-matched control confirms that the speedup stems mainly from core alignment rather than parameter reduction.

P. Jiao +50