Xolotl: run DeepSeek-V4.1-Flash, all 552B, as a local coding agent on one gaming PC (RTX 4090 + 110 GB RAM): ~33 tok/s on agent tasks, ~41 in fast mode, 512K context. Plus GLM-5.3-Flash, Qwen3.8 and LFM3-Omni. MoE experts tiered across VRAM, pinned RAM and NVMe, with Xolotl Direct NVMe->VRAM. OpenAI + Anthropic API.
What does a transformer compute when it adapts to unknown physics in-context? Pre-registered mechanistic study: Bayes-optimal until the floor, sufficient statistics in the key-value memory, an in-context controller that explores but fails on unstable plants.
Archive Execution Architecture: PureBasic listings lowered through a 6502 IR to ROPE/AGC assembly, executed in a container and banked-memory runtime; Pascal kernels lowered to SPIR-V for CUDA-free GPU compute.
Mobility-driven IoT computation substrate: simulator, compiler and testbench for computing across thousands of small energy-harvesting devices linked by steerable free-space photonics
Next-generation networks must provide strong confidentiality while sustaining high throughput and operating under constrained computational and cryptographic resources. Encrypting entire packets is often unnecessary when only specific fields contain sensitive information, and it may unnecessarily consume key material, particularly in quantum key distribution-assisted networks. Moreover, regulatory and operational requirements may require access to explicitly defined network metadata, which blanket encryption can obscure together with protected payload data. This paper presents ADASEC-SDN, an SDN-controlled, FPGA-oriented selective encryptor that applies per-flow, byte-granular OTP-based encryption while preserving the visibility of designated non-sensitive metadata. The proposed data plane is implemented at the RTL level in SystemVerilog and comprises flow classification, key provisioning, key-byte scheduling, XOR-based encryption, and packet forwarding. The implementation is functionally verified using ModelSim simulations together with a Python- and PCAP-based testbed, confirming correct flow classification, selective encryption, and packet routing. Performance experiments further demonstrate benefits of the proposed mechanism in terms of throughput. The synthesis results confirm the feasibility of resource-aware selective confidentiality for high-speed programmable networks, while testing in a real hardware environment and integration of additional security services remain subjects of future work.
Register-transfer-level (RTL) design is a critical but labour-intensive stage in digital hardware development. Large language models (LLMs) have recently shown the ability to convert natural-language specifications into Verilog code, thereby reducing manual design effort. This paper presents a review of LLM-based Verilog generation and verification approaches, organised around deployment requirements and verification strategies. The review covers classical electronic design automation (EDA) foundations, general-purpose cloud coding assistants, Verilog-specialised models and benchmarks, feedback-driven and agentic EDA frameworks, open-source verification back ends, and resource-constrained deployment using quantisation and lightweight inference engines. The literature indicates that cloud- and GPU-based systems remain dominant, while compiler and simulator feedback can improve the syntactic validity of generated designs. However, functional correctness remains difficult because compilation alone does not verify intended hardware behaviour. Resource-constrained deployment is feasible in selected cases; for example, RTLCoder reports a 4-bit quantised model of approximately 4 GB that can operate on a laptop. Nevertheless, consistent reporting of latency, memory use, energy consumption, offline operation, privacy, and reproducibility is still limited. This review identifies major open challenges, including functional verification, benchmark quality, scarcity of high-quality Verilog training data, security of generated hardware, small-model efficiency, and standardised resource-aware evaluation. The findings suggest that reliable AI-assisted RTL design will require closer integration of LLMs with deterministic EDA tools, stronger verification oracles, improved benchmarks, and transparent evaluation practices.
Amol Dhote +50International Journal of Research Publication and Reviews
This paper describes the design and FPGA implementation of a high-performance complex multiplier developed using Xilinx IP cores in the Xilinx ISE Design Suite. Complex multiplication is widely used in digital signal processing and communication systems, where both speed and hardware efficiency are essential. In this work, the real and imaginary parts are computed simultaneously using vendor-optimized multiplier IP cores and dedicated FPGA DSP slices. By taking advantage of these built-in resources instead of designing the multipliers from basic logic elements, the overall design becomes more efficient, faster, and easier to implement. This approach improves throughput, reduces delay, and provides a practical solution for real-time complex arithmetic operations in spartan 6-XC6SLX45-CSG- 324C FPGA based systems. The Simulation results are validated Through the Xilinx-Isim simulator.Xilinx XST synthesizer is used for synthesis results, Xpower analyzer is used for power analysis.The design uses 37mW of power and the timing delay for the design is 2.944ns. Here we use 32 slice registers, 18 Look-up tables
B. Rehman +50International Conference on Circuit, Power and Computing Technologies
Modern secure hardware platforms and DMA-based systems require encryption architectures that provide low power consumption, efficient data processing, and reliable operation. Although the Advanced Encryption Standard (AES) is widely adopted for hardware security, the choice of operating mode significantly influences hardware efficiency and overall system performance. This work presents the RTL implementation and comparative evaluation of six AES-128 operating modes, namely Electronic Codebook (ECB), Cipher Block Chaining (CBC), Cipher Feedback (CFB), Output Feedback (OFB), Counter (CTR), and Cipher Feedback-64 (CFB-64), using Verilog HDL and Cadence Genus targeting a 90 nm CMOS technology library. Based on synthesis analysis, AES-CFB-64 was identified as the most suitable architecture owing to its low power consumption and favorable timing characteristics. To further improve system efficiency, the selected AES-CFB-64 architecture was integrated with a custom FSM-based DMA controller and optimized using pipelining, clock gating, and error monitoring mechanisms. The proposed architecture enables efficient data transfer, reduced switching activity through clock gating, and improved operational reliability through error monitoring. Comparative results demonstrate that the proposed design achieves a 19.9% reduction in area and an 85.7% reduction in power consumption compared with the baseline CFB-64 architecture while maintaining positive timing slack and correct functional operation. Simulation and synthesis results validate the effectiveness of the proposed AES-DMA architecture for low-power and secure cryptographic applications
This paper presents a novel 10T2R resistive random-access memory (RRAM) based nonvolatile SRAM (NVSRAM) bitcell for energy-efficient and reliable nonvolatile operation. The proposed architecture employs decoupled bilateral RRAM paths to eliminate the DC short-circuit issue present in conventional NVSRAM designs, thereby improving restore reliability, stability, and robustness. The NVSRAM bitcell is implemented using TSMC 65 nm CMOS technology, while the RRAM device is adopted from the Stanford Verilog-A compact model calibrated with $\mathbf{H f O}_{x}$-based RRAM characteristics. Simulation results demonstrate femtojoule-range store and restore energy consumption, achieving a 14.41% reduction in store energy and a 31.16% reduction in restore energy compared with existing NVSRAM bitcell. In addition, Monte Carlo and process variation analyses confirm stable and reliable operation under variability conditions. The proposed design is therefore well suited for low-power embedded systems, edge AI, IoT, and instant-on computing applications.
Sharada Katailiha +20International Symposium on VLSI Design and Test
High-order silicon micro-ring resonator (MRR) filters are crucial for dense wavelength-division multiplexing (WDM) systems. However, their practical implementation is severely hindered by a "tuning disaster" induced by fabrication variations and thermal crosstalk. Furthermore, traditional optical simulators lack the capability for closed-loop automated tuning with complex peripheral control circuits. This paper proposes a novel paradigm to achieve the automated closed-loop tuning of an 8-channel flat-top MRR filter entirely within a standard Electronic Design Automation (EDA) environment. By developing a rigorous electro-optic-thermal multi-physics behavioral model using Verilog-A, we seamlessly bridge the optical interference mechanisms with electrical driving signals. Leveraging the native Cadence Virtuoso Advanced Design Environment (ADE) optimization engine, we demonstrate an automated closed-loop optimization strategy that simultaneously calibrates 23 independent control voltages to overcome physical non-idealities. Simulation results demonstrate that the proposed method successfully synthesizes a box-like broadband spectrum spanning from 1551 nm to 1556 nm. The optimized filter exhibits excellent performance with an in-band ripple of < 1 dB and an insertion loss of < 5 dB. This work provides a robust engineering solution for the automated tuning of complex photonic integrated circuits.
Yong-Bo Meng +50International Test Conference in Asia
In contemporary System on a Chip (SoC) and embedded designs, data transfer speed and reliability are crucial because data transfer should take place with minimal CPU intervention while preserving integrity. Traditional Direct Memory Access (DMA) controllers are primarily throughput-oriented and do not commonly incorporate advanced features such as adaptive transfer methods or fault tolerance techniques. This paper presents the design and implementation of a fault tolerant, adaptive DMA controller that employs methods of data integrity based on Error Correcting Code (ECC) (Single Error Correction (SEC) - Double Error Detection (DED)). The proposed architecture increases data transfer reliability by allowing correction of single-bit memory errors and detection of double-bit errors while performing high speed transfers. Adaptive burst length control adjusts the data transfer length based on responses from the system environment providing improved bandwidth utilization and reduced latency. Each component of the DMA controller is modeled with a finite state machine and modular SEC-DED encoder/decoder units. The design was completed using Verilog HDL, functional verification via ModelSim RTL simulation and synthesis via Xilinx VIVADO targeting Spartan-7 FPGA. The proposed DMA controller has achieved enhanced efficiency, increased fault tolerance, and improved architectural robustness, making it well-suited for high-performance embedded and high-performance SoC systems.
S. Arun +10International Conference on Circuit, Power and Computing Technologies
Large Language Models (LLMs) are increasingly being explored for register-transfer-level (RTL) generation, hardware verification, debugging, and design automation tasks. Unlike traditional electronic design automation (EDA) workflows that rely heavily on formal methods and handcrafted heuristics, LLM-based systems introduce natural-language-driven hardware synthesis and reasoning capabilities. Recent work has illustrated promising results in Verilog generation, testbench synthesis, simulation-guided refinement, and retrieval-augmented hardware design. However, existing approaches remain fragmented across different architectures, benchmarks, and evaluation methodologies, while challenges such as functional correctness, silent failures, scalability, and hardware-aware reasoning remain unresolved. This paper presents a focused overview of LLM-assisted RTL generation and verification systems. We organize existing methods into direct generation, retrieval-augmented, verification-aware, and agentic multi-stage architectures, and discuss commonly used benchmarks, evaluation metrics, and tool flows. Finally, we highlight major limitations and outline future research directions toward reliable and production-ready AI-assisted hardware design systems.
Conventional CMOS-based circuits are advancing toward higher reliability, reduced leakage, and robust electrostatic discharge protection. Typical Electrostatic Discharge (ESD) protection structures employ static ballast resistors and diode-based clamps, and struggle to balance low leakage and high-speed performance, while maintaining sufficient transient suppression at nanoscale nodes. We implemented current-controlled memristors as an adaptive current limiting element for ESD protection to enhance the device reliability and energy efficiency in advanced ICs. Using experimental results, a current-controlled model was developed in Verilog-A, and then integrated into an ESD protection circuit under the Human Body Model configuration. The results demonstrated that the memristor effectively limits transient current surges by dynamically adjusting its resistance during ESD events. The findings highlight the potential of current-controlled memristors as scalable, low-power, and self-adaptive elements for next-generation ESD protection in nanoscale CMOS technologies.
Hao Qian +50Midwest Symposium on Circuits and Systems
Direct Digital Synthesis (DDS) has become a practical technique for generating precise waveforms in communication, instrumentation and embedded electronic systems. This paper presents the implementation of a multi-waveform DDS architecture on an Artix-7 FPGA platform. The developed system employs a 32-bit phase accumulator, a lookup-table-based waveform generation unit and an 8-bit R-2R digital-to-analog converter to produce sinusoidal, square and triangular waveforms. Frequency control is achieved through an ESP8266 module that transfers frequency tuning words to the FPGA using SPI communication. A reference clock of 100 MHz was used during implementation, resulting in a theoretical frequency resolution of 0.0233 Hz. Functional verification was first performed through simulation and subsequently validated through hardware measurements using a digital storage oscilloscope. Experimental results showed output frequencies of 7.40 kHz for sinusoidal and square waveforms and 4.63 kHz for triangular waveforms, corresponding to frequency accuracies above 98%. Resource utilization analysis indicated that the design occupied only a small portion of the available FPGA resources. The obtained results demonstrate that the proposed architecture provides a low-cost and reconfigurable solution for waveform generation in educational laboratories and embedded signal-processing applications.
The potential of electric vehicles (EVs) as an important flexible resource in the electricity market is increasingly prominent, but existing dispatch strategies often overlook the coordination mechanism between distribution network physical constraints and multilevel market signals. To address this issue, this paper aims to develop a real-time intelligent charging optimization strategy for EV-grid collaboration to balance the profits of charging operators, the preferences of vehicle owners, and grid safety. This paper integrates model predictive control (MPC) and dynamic programming (DP) to construct an MPC local control + DP global planning framework for multiperiod charging and discharging optimization, and designs four integrated scheduling strategies based on the EV2Gym platform. The main novel contributions of this study are twofold: first, a three-level price linkage model is designed to clarify the interaction rules between the regional day-ahead market and the local real-time market; second, grid factors such as transformer power limits and photovoltaic accommodation constraints are quantitatively integrated into the optimization framework. Experiments demonstrate that compared with existing technologies, the proposed strategy increases profits by 8%, decreases the risk of transformer overload by 15%, and improves the PV accommodation rate by 20%, while successfully controlling battery degradation within an acceptable range. This study provides an effective and practical solution for the implementation of large-scale intelligent charging.
This study aims to examine how user-generated content (UGC) influences medical travel decision intention and to explore the mediating role of destination trust in this relationship. The research seeks to clarify the psychological mechanisms through which digital social influence shapes potential travelers' choices in a high-risk medical tourism context. A quantitative research design was adopted using a structured questionnaire administered to potential international travelers. Out of 517 distributed questionnaires, 209 valid responses were retained for structural equation modeling (CB-SEM) using IBM SPSS Amos 22 to test the conceptual model and evaluate the indirect paths. The empirical findings reveal that user-generated content exerts a significant positive effect on both destination trust and medical travel decision intention. Furthermore, destination trust acts as a crucial full mediating mechanism connecting UGC to behavioral intention, demonstrating that online peer-generated information reduces perceived risk and uncertainty during destination selection. Theoretically, this study integrates multiple perspectives to show that trust is more influential than raw information in high-stakes decisions, while contributing to geographical diversity by incorporating the Algerian context. Practically, the findings suggest a necessary strategic shift from traditional promotional marketing to experience-based marketing. Destination managers and healthcare stakeholders in Algeria should invest in digital transparency, patient evaluations, and structured trust-building systems similar to those in established destinations like Turkey. From a geographical standpoint, UGC reshapes the spatial mental image of destinations, indicating that digital presence now competes with physical distance in redirecting international patient flows. Ultimately, effectively leveraging UGC and managing online reputation are key to strengthening destination competitiveness and fostering health-tourism clusters.
Abdesselam Belahoual Aouad +20Geo Journal of Tourism and Geosites
In sensitive scenarios such as security surveillance and emergency search and rescue, Unmanned Aerial Vehicle (UAV) face recognition faces simultaneous challenges from severe physical degradation, scarcity of annotated samples, and privacy compliance constraints. Existing research mainly focuses on either low-quality face recognition under plaintext conditions or template protection and secure inference under privacy-preserving conditions. However, privacy-preserving target-domain adaptation for UAV face recognition remains underexplored. This paper proposes Pri-UAVFace, a Minimal Secure Multi-Party Computation (MPC) adaptation framework for privacy-preserving UAV face recognition. Rather than migrating the entire training workflow into MPC, it learns general identity representations and degradation robustness from public plaintext data, while confining residual target-domain adaptation to the ciphertext domain. To this end, it combines physical simulation and consistency learning for public-domain pre-learning, followed by MPC-based few-shot adaptation using Parameter-Efficient Fine-Tuning. Furthermore, an encryption-friendly discriminative head, named MPCCosineHead, is designed to reduce the protocol interaction overhead caused by non-linear operators. Evaluated on the real-world DroneFace dataset, Pri-UAVFace achieves an AUC of 0.9943 and a TAR@FAR=10-3 of 0.9308 under a 5-shot setting. Through joint method and protocol optimization, it reduces the number of MPC communication rounds from 864 to 152 and shortens single-step latency by 75%. These results demonstrate effective UAV target-domain adaptation with minimal MPC utilization, providing a feasible path for privacy-sensitive edge multimedia sensing.
Olfactory feedback enhances immersion in virtual reality (VR), yet its deployment is constrained by the limited number of odor sources. We investigate whether visual cues can influence users' interpretation and affective evaluation of the same physical odor in immersive environments without increasing hardware complexity. We conducted a formative study to identify odor characteristics susceptible to visual reinterpretation, a pre-study to calibrate odor intensity, and a within-subject VR experiment manipulating contour recognizability (concrete vs. abstract) and lighting conditions (achromatic lighting, cool-toned lighting, warm-toned lighting). Participants evaluated virtual objects paired with two base odorants in terms of odor matching, perceptual clarity, and affective responses. Results demonstrate the feasibility of visually guided odor reinterpretation, with an average matching success rate of 80.45% across the three phenylethanol-associated objects. Our findings show that visual design systematically influences users' interpretation of odor categories and their experience of pleasure through ease of recognition, authenticity, and the color atmosphere created by lighting. We further discuss how visual cues can influence odor interpretation and how cross-modal olfactory illusions affect the virtual experience.
Dual-Camera Compressive Hyperspectral Imaging (DCCHI) introduces an auxiliary panchromatic (PAN) observation to mitigate spatial information loss in compressed hyperspectral measurements. Despite recent advances in deep unfolding for DCCHI, two challenges remain: physically coupled PAN-guided reconstruction and effective modeling of spatial degradations caused by coded modulation and spectral dispersion. To address these issues, we propose PPU-Net, a Physically-coupled PAN-guided Unfolding Network. Derived from a joint optimization model, PPU-Net unfolds the reconstruction into three interpretable steps: PAN-consistency updating, hyperspectral-consistency updating, and learned refinement. It explicitly enforces PAN consistency through the spectral response operator and its adjoint, establishing principled physical coupling between the two observations. Furthermore, we propose a Dispersion-Guided Self-Attention (DGSA) tailored to capture both spatial structures and dispersion-induced artifacts. DGSA adopts a dual-branch design with shared query, key, and value features: a structural branch models local spatial structures with enlarged receptive fields, while a dispersion branch leverages stripe-shaped windows aligned with the dispersion direction to suppress ghosting artifacts. Experiments on both simulated and real datasets demonstrate that our method consistently outperforms state-of-the-art approaches in reconstruction quality and computational efficiency. Code is available at https://github.com/M1zuk1/PPU-Net.git
Industrial Physical AI extends data-driven intelligence from purely digital spaces into the physical world by embedding physical laws, multimodal perception, and causal reasoning into model design, opening a path toward reliable, deployable intelligent systems for safety-critical industrial environments. The International Workshop on Industrial Physical AI (WIPAI 2026), co-located with the 34th ACM International Conference on Multimedia (ACM MM 2026) in Rio de Janeiro, Brazil, provided a focused forum for researchers and practitioners to discuss the latest advances, challenges, and opportunities in multimodal-driven physical AI for real-world industrial scenarios. The workshop addresses the gap between general multimedia research and industrial deployment, treating multimodal technology as the perceptual foundation of Industrial Physical AI. After a rigorous review process, eight papers were accepted for oral presentation, spanning multimodal sensing and monitoring of manufacturing processes, physics-grounded agents and world models, LLM orchestration for physical assembly, vision--language--action policies and their execution runtimes, the physical fidelity of feed-forward 3D reconstruction, and topological validation of robot-swarm dynamics. This summary describes the motivation, scope, and technical themes of the workshop and overviews the accepted contributions.
Large language model (LLM) accelerators are commonly optimized for fixed model structures and isolated inference workloads, whereas emerging artificial-intelligence operating systems (AI OSs) primarily virtualize agents, memory, and tools in software. Reusable hardware, phase-aware serving, operating-system-level composition, and automata-based formalisms each exist separately in prior work; this architecture paper treats them simultaneously to ask whether LLM inference can be exposed as a secure, reusable hardware service. We introduce HCDP-LLM, an AI-OS-managed hierarchical common datapath that composes reusable tensor, reduction, nonlinear, quantization, memory, and key–value (KV)-cache modules through model capsules and a Moore finite-state controller. The architecture is formalized as a finite-precision multi-tape computational transducer, enabling deterministic checkpointing, migration, isolation, and Turing-machine-based bit-time, space, and I/O analysis. An explicit derivation on this transducer yields quadratic prompt-attention work, context-linear per-token attention and KV-cache storage, and separates compute-dominated prefill from bandwidth-dominated decoding. The formal model separates finite-state physical realizations from a scalable family with extensible memory that can support universal computation; a worked example derives from it the admission limit, lane allocation, and rebinding granularity of a 7B-parameter model on one reconfigurable accelerator. No prototype measurements or simulations are reported; a pre-specified evaluation protocol with a control-plane-disabled baseline, which isolates the abstraction-layer cost, defines how a field-programmable gate array (FPGA) or system-on-chip prototype must test the resulting position: Hardware LLMs as runtime-composable, weight-programmable inference fabrics rather than permanently hardwired models.
Retail distribution networks in Saudi Arabia must coordinate variable outlet demand, constrained transport capacity, retailer receiving windows and geographically diverse delivery conditions. This critical narrative review evaluates machine-learning-assisted route-to-market optimization in the context of Saudi Vision 2030 and the National Transport and Logistics Strategy. It synthesizes research on vehicle routing, neural combinatorial optimization, predictive demand modeling and learning-assisted local search, and proposes an integrated architecture connecting commercial data, predictive models, constraint-aware optimization and human-supervised dispatch. The review distinguishes operational feasibility from algorithmic benchmark performance and develops a framework for evaluating distance, on-time-in-full service, fleet utilization, cost, reliability and emissions. Two conceptual figures and comparison tables support the discussion; all numerical scenario values are explicitly illustrative rather than empirical findings. The evidence suggests that hybrid approaches are plausible candidates for deployment when prediction quality, governance, commercial service policies and optimization robustness are jointly addressed. Future research should prioritize Saudi-specific datasets, controlled evaluation, heterogeneous fleet constraints, multimodal interfaces and transparent reporting of uncertainty.
Mohamed Saad Hassan Ismail0Iconic research and engineering journals
The following research presents the development of “black box” models for the dynamic characterization of a Yazaki WFC SC20 absorption chiller. Unlike previous studies based on proprietary software, this research uses open-source software technologies such as Python and PyTorch to improve the reproducibility and scalability of the modeling. Two deep neural network architectures were designed and compared: a Long Short-Term Memory (LSTM) network and a hybrid convolutional-recurrent model (CNN-LSTM). To optimize performance, the Optuna framework was used, enabling intelligent hyperparameter search and input variable selection. The models were trained and validated with real data from the CIESOL center’s solar thermal installation from 2014 and 2015. The results demonstrate that the architecture that best describes the absorption chiller is the CNN-LSTM.
The use of information technologies to solve problems of agricultural production is highly relevant under current conditions. At present, there are a sufficient number of solutions based, among other things, on convolutional neural networks. However, the available neural networks are either heavyweight, making it impossible to deploy them on edge and mobile devices, or lightweight, but without an analysis of the accuracy of solving the required task, in particular segmentation, classification, and detection of crop objects. Therefore, the development and validation of a convolutional neural network with an analysis of the influence of architectural parameters on the accuracy and computational complexity of the neural algorithm is highly relevant. Classification of agricultural crops was chosen as the research problem. The object of the study is digital images of grape leaves. The dataset is divided into four classes: healthy leaves and leaves affected by three types of diseases. The following parameters were varied: the number of convolutional blocks, the initial number of channels, and the use of batch normalization layers. Accuracy, loss value, and the amount of required computational resources were used as evaluation criteria. Based on a grid search over architectural parameters, the most optimal configuration of the convolutional neural network was identified. It was found that the developed configuration outperforms analogous models both in terms of accuracy and computational complexity and can be used for monitoring agricultural crops on edge and mobile devices under field conditions.
The high computational overhead of deep learning architectures often hinders real-time flood detection in resource-constrained IoT deployments. This paper proposes a framework for high accuracy flood detection based on systematically optimized K-Nearest Neighbors (KNN) algorithm. The proposed model integrates real-time telemetry from rainfall, water level and humidity sensors, and prioritizes low-latency inference without compromising predictive precision. Using a grid search-based hyperparameter strategy and 10-fold walk-forward cross-validation, the model was able to achieve a peak accuracy of 98.79%. Comparing with non-optimized Convolutional Neural Networks (CNN) and Long Short-Term Memory (LSTM) networks, the tuned KNN model is more suitable for deployment in infrastructure limited regions. The results show that systematic optimization of lightweight algorithms can match the performance of deeper architectures in environmental monitoring applications under resource constraints.
Graph neural networks (GNN) based on message passing are provably no more powerful than the one-dimensional Weisfeiler--Leman colour-refinement test (1-WL): two graphs it cannot tell apart receive identical representations, however deep or wide the network. A common remedy augments node or edge features with precomputed structural descriptors, most often counts of a fixed small subgraph such as triangles or longer cycles, but such counts require committing in advance to the size of the substructure counted, a choice usually made blind to the data. We study a descriptor that avoids this choice. The edge-girth of an edge is the length of a shortest cycle through it, and its multiplicity is the number of such shortest cycles; together they form a per-edge invariant that reports cycles of arbitrary length, computable exactly by a single breadth-first search per edge. Injected into a gated message-passing architecture, EGAGNN, it reaches a test MAE a factor three below the closest gated comparator on the ZINC-12k regression benchmark at 104k parameters; against bounded cycle-counting descriptors under the same architecture, it matches only a dictionary counting cycles up to length eight, using twice as many channels, while a dictionary capped at length four performs no better than no structural information at all. On graph discrimination we prove a matching limitation: on graphs where every edge sees the same number of shortest cycles of the same length, the descriptor becomes constant and any model built on it collapses back to the 1-WL bound. This holds without exception across all 400 pairs of the BREC benchmark: not one of the 90 such pairs is distinguished.
The weaponization of information has become a defining feature of contemporary conflict, and generative artificial intelligence has lowered the cost of producing disinformation at scale. This paper examines how deep machine learning models for automated information retrieval and fact verification can be trained, deployed and rigorously evaluated as instruments of national security, with particular reference to Pakistan. Drawing on securitization theory and the resilience framework, it proposes and implements a human-supervised pipeline of claim screening, lexical–dense evidence retrieval, deep verdict prediction, calibration and conformal triage. The analysis uses only real data: a compiled cross-national dataset on organized manipulation, platform takedowns, fact-checking capacity, public perception and benchmark performance, together with two human-annotated claim corpora (12,836 political statements and 1,535 climate claims with 7,675 claim–evidence pairs). Six deep architectures were trained in Python, and the evidence was analysed with fractional-logit quasi-maximum likelihood with wild bootstrap inference, logistic diffusion and unknown-break structural change tests, ordered probit, two-way clustered fixed-effects models, double/debiased machine learning with neural-network nuisance functions and Oster coefficient-stability bounds. Organized manipulation grew by 42.5% a year between 2017 and 2020, whereas fact-checking capacity saturated near 468 projects and has declined since late 2023. On real claims, text-only deep models did not outperform classical baselines; gains came from source metadata; between a quarter and two-fifths of correct verdicts lacked supporting evidence; and a low-cost typographic attack erased the best model’s advantage. Calibrated conformal triage, however, delivered guaranteed coverage. The paper concludes that automated verification should be built and judged as a strategic triage capability that supports, rather than replaces, accountable human analysts.
Cyberattacks on energy, finance, health and logistics systems now produce losses large enough to register at the level of national economies, yet the quantitative tools that link threat intelligence to macroeconomic exposure remain fragmented. This paper develops a hybrid framework that couples advanced econometrics with deep and machine learning across three modules: threat intelligence, vulnerability assessment and predictive resilience. Using a compiled dataset of published series for 2015–2025 (NVD/CVE disclosures, FBI IC3 complaints and losses, IBM/Ponemon breach costs), ITU Global Cybersecurity Index scores for 25 economies and 24 documented macro-relevant incidents, we estimate HAC and wild-bootstrap elasticities, ARDL bounds and DOLS long-run relations, Bayesian conjugate regressions, sup-F break and Granger tests, quantile, Tobit and fractional-logit models, and a Poisson–lognormal loss distribution with extreme-value diagnostics. A from-scratch LSTM deep ensemble, deep MLP, random forest, gradient boosting and an ARX–LSTM hybrid are evaluated by rolling-origin forecasting with Diebold–Mariano tests and exact Shapley attribution. Reported losses grew 35.6% a year against 22.4% for vulnerability disclosures; disclosures Granger-lead losses. Gradient boosting gave the lowest out-of-sample error (log RMSE 0.113 against 0.362 for a random walk); the LSTM ensemble reached 0.191. The hybrid projection puts US-reported losses at USD 52.0 billion by 2028 (80% interval 41.0–65.9). Incident severity is lognormal with a heavy right tail, and parameter uncertainty roughly doubles the 99% Value-at-Risk. Projected expected losses range from 0.15% of GDP in the United States to 0.26% in Nigeria, with Pakistan and Ukraine close behind.
Video surveillance systems play a vital role in modern security, transportation, and public safety infrastructure. Most of the analytical functions of an intelligent surveillance system, from detection and tracking to ROI-aware compression, begin by isolating a Region of Interest (ROI), yet the algorithms that perform this step are scattered across the computer-vision, video-coding and sensor-network literatures, use the term “ROI” in different senses and report results on metrics that cannot be compared directly. This narrative review classifies and critically analyzes ROI extraction algorithms for video surveillance published between 2020 and 2026. A comparison with survey papers published between 2021 and 2026 shows that none is centered on the extraction step itself. The reviewed methods are organized by explicit criteria (hand-crafted or learned features, learned or online-estimated decision rules, annotation dependence and output granularity) into background-subtraction, classical machine-learning, deep-learning and hybrid families, and the deep-learning family is further divided into six groups. Five semantic variants of the ROI are distinguished, and reported results are compared only within the metric family appropriate to each variant. Widely used video surveillance datasets are compiled, described and compared. The analysis shows that the evaluation base of the field is narrow, with fourteen of the compared studies reporting their principal results on CDnet 2014 and none reporting run-to-run variance. Because the compared studies differ in dataset, subset, training split and averaging, the reported values indicate operating ranges and do not show that any method family, deep learning included, is generally superior to another. Research directions are derived from these limitations, including cross-dataset evaluation with temporal-stability measures, codec-aligned ROI extraction for edge devices, motion-gated promptable segmentation, extraction from moving cameras, and the joint design of ROI extraction and privacy protection.
Smart campuses generate massive heterogeneous data from IoT sensors, cameras, and administrative systems, yet existing infrastructures operate in silos, leading to fragmented visibility, reactive maintenance, and suboptimal resource utilization. This paper presents a unified AI Digital Twin framework that integrates real-time IoT telemetry, centralized data ingestion, predictive machine-learning analytics, and an interactive 3D digital twin interface to enable data-driven campus operations. The architecture comprises four layers: (1) edge sensor networks (DHT22, PIR, RFID, PZEM-004T, HC-SR04) collecting environmental, occupancy, electrical, and water-level data via MQTT/HTTP; (2) a FastAPI-backed central platform persisting telemetry in SQLite and synchronizing a room-level digital twin state; (3) AI/ML models (RandomForest regressor for occupancy forecasting, IsolationForest for anomaly detection, LogisticRegression for asset-failure probability) trained on historical observations; and (4) a web-based dashboard with Three.js 3D visualization, Chart.js KPIs, and an AI-assisted conversational interface. The system enforces data-quality guarantees: no synthetic feeds at staleness, explicit unavailability markers, and predictions always labeled as such. Deployed on a live testbed at PRMCEAM, the platform demonstrates real-time monitoring of 12 campus rooms, predictive maintenance alerts for electrical anomalies, 24-hour occupancy forecasts with mean absolute error below 3 persons, and resource optimization recommendations reducing estimated idle power consumption by 18%. Results confirm that a unified digital twin foundation enables scalable, intelligent campus management while maintaining transparency regarding data limits and prediction confidence.
P. P. Tijare +50International Journal for Research in Applied Science and Engineering Technology
HeartGuard is a prototype system that uses Internet of Things (IoT) technology to monitor cardiovascular health and predict risks. It combines physiological sensing, ECG data capture, wireless communication, database storage, and patient medical profiles within one integrated platform. The hardware centers on an ESP32 microcontroller and includes sensors for heart rate, oxygen saturation (SpO₂), temperature, and ECG. The system displays data locally on an OLED screen and also sends it to a web application, allowing remote data viewing. The software integrates real-time sensor readings with patient details such as age, sex, blood pressure, cholesterol, chest-pain type, exercise-induced angina, ST depression, and resting ECG results. For risk prediction, HeartGuard uses a Random Forest machine learning model, while ECG arrhythmia classification relies on a convolutional neural network (CNN). Currently, these models are developed using publicly available data from Kaggle. In the future, HeartGuard aims to train models directly on data collected by the deployed sensors. The prototype presents cardiovascular risk as a percentage calculated at the application level, but future versions plan to use calibrated probabilities directly from the Random Forest output. The system includes interactive interfaces for risk visualization, ECG analysis, patient records management, and generates clinical-style PDF reports.
Kapil Gupta +40International Journal for Research in Applied Science and Engineering Technology
A hybrid perception-control framework for navigating rough terrains by remote YOLOv8n-based terrain classification combined with on-board safety-based control of rover is proposed. The system uses ESP32-WROOM-32 module for controlling motors, ultrasonic sensors for detecting obstacles, infrared sensors for detecting cliff edges, and watchdog communication, whereas ESP32-CAM is responsible for sending terrain images to a laptop where the inference algorithm works. Four-class terrain data consisting of 800 images of grass, sand, gravel, and soil is employed for fine-tuning YOLOv8n using stratified dataset. Terrain predictions control movements depending on traversability coefficients, while safety controls take precedence over remote controls when an obstacle or communication link is detected. The evaluation over 60 trials of outdoor navigation has resulted in terrain classification with an accuracy of 78.3% and successful navigation with an 80.0% success rate. The perception-to-action latency was 170±23 ms, fail-safe time was 1.02±0.04 s, and emergency braking was 310±45 ms. The power consumption of the rover was 4.22±0.35 W, and the estimated operational time was around 4.5 h. Such values characterize the performance of the limited resources rover that implements the separation of complex computation-based inference from motion control in embedded systems.
Rapid evolution of Artificial Intelligence of Things (AIoT) has allowed intelligent data-driven decision-making under several cyber-physical systems (CPS) like smart cities, smart grids, industrial automation, and healthcare monitoring. AIoT incorporates artificial intelligence (AI) methods with IoT devices to develop large volumes of sensor data and improve system efficacy, automation, and predictive proficiencies. Among different forms of attacks, the false data injection attack (FDIA) has proven that high-priority cyber-related problem and has received more attention recently. The FDIA detection is an important security process in the smart grid controller system and is resolved by data-driven machine learning (ML) algorithms. Now, Federated Learning (FL) is a well-known ML method that assists in addressing data privacy problems by permitting users to collect its raw data locally and transferring it local model parameters only to the aggregator server to collectively train a shared global model. This paper develops a Federated Learning-Based Framework for Detecting Attacks Using an Ensemble Classification Model (FLFDA-ECM) in distributed cybersecurity systems within the Artificial Intelligence of Things (AIoT) environment. The major intention of the FLFDA-ECM paper is to propose an FL-based cybersecurity framework for detecting and classifying FDIA in AIoT- enabled distributed systems while preserving data privacy. At first, the data pre-processing contains dual levels, such as handling missing data and feature scaling to regularize the input data. In the feature selection stage, the hybrid approach integrates minimum redundancy maximum relevance and ReliefF techniques to recognize the most appropriate attributes. Furthermore, an ensemble-based weighted majority voting approach incorporating bi-directional gated recurrent unit, variational autoencoder, and temporal convolutional network models is employed for attack detection. Finally, the parameter fine-tuned model has been executed by the Adan optimizer to enhance the classifier solution. The performance evaluation of the FLFDA-ECM technique occurs using a benchmark dataset. The empirical outcomes specified the boosted performance of the FLFDA-ECM model in comparison with recent techniques.
In the construction industry, severe schedule and cost overruns, retrospective performance tracking, and fragmented communication are persistent problems. The Construction 4.0 paradigm has encouraged a trend toward data-driven, real-time project management that depends on the Internet of Things (IoT), artificial intelligence (AI), and building information modelling (BIM). This systematic evaluation, which focuses on multidimensional resource optimization, schedule, and cost, critically evaluates the use of IoT technology into construction project management. By combining content primarily released between 2019 and 2026, the analysis looks at how continuous sensor data transforms traditional, reactive management frameworks—such as Earned Value Management (EVM) and the Critical Path Method (CPM)—into proactive, predictive control systems. Evidence indicates that while individual sensor technologies and machine learning delay-prediction algorithms have reached a high degree of maturity, there are currently few comprehensive platforms connecting these various data silos. This paper outlines a comprehensive nine-layer conceptual framework for fully integrated predictive construction management, identifies major implementation challenges, and explains sensor architectures, predictive models, and digital twin integrations that are revolutionizing the sector.
The integration of machine learning and computer vision has advanced automated quality control systems in the coffee processing industry. However, limited industrial-scale validation and real-time implementation remain challenges in intelligent coffee bean defect detection. This study aims to identify research trends, technological developments, and research gaps through a Systematic Literature Review (SLR) and bibliometric analysis. Data were retrieved from the Scopus database, comprising 1,630 publications from 2023 to 2026, and analyzed using VOSviewer, Bibliometrix, and Biblioshiny. The results revealed increasing publication trends from 2023 to 2025, with machine learning, deep learning, artificial intelligence, and computer vision emerging as dominant research themes. Keyword co-occurrence analyses highlighted the prominence of neural networks, support vector machines, and automated inspection technologies. The literature synthesis further identified Convolutional Neural Networks (CNN) and YOLO as relevant approaches for coffee bean classification and defect detection. Despite these advancements, challenges remain in real-time deployment, industrial integration, and model validation. This study contributes to identifying research gaps and proposing a research roadmap toward intelligent quality control systems integrating machine learning, computer vision, and Internet of Things (IoT) technologies to support efficient and automated coffee bean inspection within Industry 4.0.
Smart Universities increasingly rely on connected learning platforms, smart classrooms, Internet of Things (IoT) devices, cloud services, and digital administrative systems. This integration expands the attack surface and creates security requirements involving educational data, heterogeneous infrastructure, and the continuous availability of academic services. This survey examines AI-based intrusion detection through three complementary perspectives: campus-specific threats and architectural layers, detection approaches and their deployment trade-offs, and operational evaluation and governance requirements. It reviews machine learning, deep learning, ensemble and hybrid models, lightweight detection, federated learning, and explainable AI, relating their capabilities and limitations to device, edge, network, cloud, and application layers. The analysis highlights gaps between benchmark performance and evidence of suitability for operational university environments, including limited campus-specific validation, incomplete reporting of latency and resource requirements, and insufficient assessment of privacy safeguards and explanation usefulness. Conventional detection metrics remain necessary but should be complemented by false-alert workload, end-to-end latency, temporal robustness, communication costs, and analyst-centered evaluation. The survey provides a campus-oriented synthesis, an operational evaluation matrix, and a staged research and deployment roadmap prioritizing documented datasets, accountable data governance, baseline assessment, and monitored pilots before broader federation or automated response. AI-based intrusion detection offers promising capabilities for Smart Universities, but its practical value requires validation under representative campus conditions and alignment with institutional resources, privacy obligations, and incident-response processes.
The Internet of Things (IoT) comprises a vast number of digitally networked devices. Machine learning (ML), deep learning (DL), and federated learning (FL), all subfields of artificial intelligence (AI), have recently been investigated as cybersecurity solutions for protecting these networked devices from intrusions and cyberattacks. However, existing ML and DL techniques for Sustainable IoT data carry significant privacy implications. Because of its inherent suitability for IoT settings, AI-based cybersecurity, and FL in particular, has become critical, yet traditional FL-based approaches themselves fall short in several respects. We need a framework that prioritizes privacy while enabling effective FL on edge-generated data. This research aims to address three persistent problems in traditional FL-based systems: class imbalance, unequal data distribution across devices, and computational complexity. To this end, we propose a novel Multi-Agentic AI framework in which each device decomposes its local workload across three cooperating: Agent 2 performs data preprocessing, including WGAN-based oversampling to correct class imbalance; Agent 1 performs data distribution, applying a vertical federated learning (VFL) partitioning scheme to address the unequal data distribution that typically affects FL-based IoT systems; and Agent 3 performs classification, training a deep autoencoder and an MLP on the resulting features while keeping the computational load on each device manageable. Experiments on the Modbus ICS network and ToN-IoT datasets show that our framework improves on the performance of comparable ML models while overcoming the limitations of existing FL-based techniques. Our VFL approach delivers consistent data distribution and higher attack detection accuracy, reaching 99.56% and 99.62% on the two datasets, respectively. Overall, this multi-agent agentic AI methodology offers a practical way to address the core challenges of FL and meaningfully improve AI effectiveness in privacy-sensitive sustainable IoT domains.
The transition toward sustainable and resilient cities requires urban transport systems that combine efficient mobility management with environmental responsibility, digital innovation, and adaptive governance. This study investigates the concept of Green-Line Transport as an integrated framework for improving urban mobility through traffic modeling, data-driven analysis, and smart transportation technologies, using Petroșani, Romania, as a case study. The empirical component of this study combines document analysis, field observations, video-based traffic measurements, and calibrated transport modeling to assess current mobility conditions and alternative transport scenarios. Building on these empirical results, a conceptual Green-Line Transport framework is proposed in which real-time mobility data, Internet of Things (IoT) technologies, Intelligent Transportation Systems (ITSs), and machine-learning-based analytics represent prospective digital components for future implementation rather than technologies empirically deployed or evaluated in the present case study. This study further proposes a data-supported governance framework through which traffic-model outputs and mobility indicators can inform mobility-policy coordination, multimodal integration, infrastructure prioritization, environmental objectives, and resilience-oriented decision making. Governance is, therefore, addressed as a conceptual and decision-support dimension of the proposed framework rather than as an independently investigated institutional process. The analysis considers traffic flows, mobility patterns, user behavior, accessibility, and the interaction between conventional and sustainable modes of transport within the urban system. A conceptual Green-Line Transport framework is proposed as a closed-loop mobility-governance architecture that extends existing SUMP, ITS, smart-mobility, and data-driven approaches by linking strategic mobility objectives with calibrated traffic modeling, real-time monitoring, predictive analytics, resilience assessment, adaptive decision making, intervention, and continuous performance reassessment. The framework also supports the identification of mobility vulnerabilities and priority interventions for local authorities and urban planners. This study demonstrates the potential of integrating digital technologies and advanced traffic analytics within a proposed governance-oriented decision-support framework for more efficient, sustainable, and resilient urban transportation. The proposed approach provides a transferable methodological basis for medium-sized cities seeking to implement smart mobility strategies and advance the transition toward sustainable and resilient urban development.
On-device LLM agents are increasingly attractive yet remain bottlenecked by the prefill stage, where attention compute and memory access dominate latency under long agent contexts. Existing sparse attention designs are poorly suited: hard block routing may drop critical constraints, while token pruning yields fragmented accesses that conflict with block-based KV paging. We propose LinkCLAW, a spatiotemporal collaborative sparse attention framework for long-context, tool-augmented on-device agents. Spatially, LinkCLAW introduces Link Sparse Attention (LSA): learnable link tokens appended to each block compress block-level semantics and mediate cross-block interaction, enabling heterogeneous connectivity that combines high-fidelity link-to-token access for hotspot blocks with low-rank link-to-link interactions for background blocks. Temporally, Link-guided Asynchronous Prefill (LAP) removes the ''wait-for-all'' barrier by splitting each job into a block prefill (ordinary tokens, run upon arrival) and a chained link prefill (in logical order), overlapping compute with I/O under out-of-order arrivals without requiring special position encoding. We also introduce an asynchronous agent benchmark with heterogeneous job latency and out-of-order returns. Experiments on seven benchmarks show that LinkCLAW matches or exceeds dense-attention accuracy while achieving 2.64×, 3.11×, and 1.47× speedups in single-layer attention latency, tail prefill, and end-to-end TTFT, respectively, for on-device agent workflows on a single consumer-grade GPU.
Fiber-optic parametric amplifiers in a phase-sensitive amplifier (PSA) configuration are well known for their low noise properties. Quantum mechanics dictates that these amplifiers can have a 0 dB noise figure (NF) (ratio of signal-to-noise ratio at the input and output of the amplifier) under certain conditions. However, for a dual-pump configuration, when the system is highly nonlinear, fully quantum calculations of NF become non-trivial. In such cases, semi-quantum models of NF evaluation exists, however, they are often complicated and resource intensive. To simplify that, in this work, first we use a semi-quantum NF calculation framework for the generation of numerically simulated input-output dataset of NF over a large parameter spaces. Then we use the data to train an optimized deep neural network (DNN) for rapid NF prediction. Utilizing the DNN, we demonstrate about 400 times reduction in the computation time compared to the standard semi-quantum NF evaluation approach with a mean absolute error (MAE) of 0.0015 dB. Using clustering we demonstrate a rapid exploration of the parameter space with respect to high gain (20 dB) and low NF (0.3 dB) targets and retrieve associated design parameters. These results illustrate the efficacy of machine learning-based approaches for rapid search of high-dimensional parameter spaces especially in compute intensive photonic systems.
Dilute Magnetic Semiconductors (DMSs) are the promising materials for spintronic, optoelectronic and photovoltaic technologies. This can be achieved by integrating the electron's spin with its intrinsic charge transport. Amongst oxide based host semiconductors, Zinc Oxide (ZnO) stands out due to its key features: wide direct energy bandgap (3.37 eV), large exciton binding energy (60 meV) and high thermal stability. Doping with transition metal (TM) or rare-earth (RE) elements into the ZnO crystal lattice induces spinpolarization and Room-Temperature Ferromagnetism (RTFM). This review highlights the crystal polymorphism, photoluminescence (PL) mechanisms, defect-mediated magnetic exchange interactions and related experimental advances in ZnO based DMSs. We also discuss the key technological applications ranging from spin field-effect transistors (Spin-FETs), spin light-emitting diodes (Spin- LEDs) and magnetoresistive random-access memory (MRAM) to transparent optoelectronics, photocatalysis and biological/gas sensing. Additinally, major challenges and prospective applications in spintronics, quantum computing, optoelectronics and catalysis are addressed.
Large language model (LLM) agents increasingly rely on external memory stores to retain information across conversations, typically via dense vector embeddings searched with approximate nearest-neighbor methods. The fruit fly's olfactory circuit computes a locality-sensitive hash of sensory input using a strikingly different mechanism - sparse random projection followed by winner-take-all competition - and prior work (Dasgupta et al., 2017) showed this scheme, dubbed FlyHash, preserves neighbor structure better per bit than conventional random-hyperplane hashing (SimHash) on several benchmarks. We apply FlyHash, to our knowledge for the first time, to LLM agent memory retrieval, and ask three questions a practitioner actually needs answered: does the quality advantage hold under a correctly matched-bit protocol (it does, robustly, across 10 seeds and four datasets spanning two modalities, with one specific, previously undocumented comparison pitfall identified and corrected along the way); does that advantage translate into real memory and latency savings, including against the dense-vector indexing (HNSW) practitioners default to today (no: at equal bits FlyHash's codes are larger and slower to compare than SimHash's, and HNSW dominates both hash methods at ordinary scale; a sparse-set comparison we introduce recovers most but not all of a naive latency gap, leaving hashing's practical case narrower and more specific than the quality result alone implies); and does the advantage survive genuine, non-circular agent-memory relevance labels rather than the nearest-neighbor-by-embedding proxy this literature otherwise relies on (it does). We report all three answers precisely, including the two that narrow rather than flatter the headline result, and release our implementation, a working agent-memory demo, and a full account of the methodological pitfalls we found and corrected.
Achieving passive protection of quantum information at finite temperature is a central problem at the intersection of quantum information and quantum many-body physics. Yet a general, computable criterion for determining self-correcting behavior directly from the form of local interactions remains elusive. Here we develop a unified algebraic-geometric framework for certifying partial self-correction in three-dimensional translation-invariant local stabilizer codes, encompassing the cubic code as a special case. From the algebraic curves defined by the codes'defining polynomials, we construct valuation matrices on the boundaries of their normalized projective closures and derive a computable rank criterion. Together with a polynomial-time decoder, this criterion guarantees super-Arrhenius memory lifetimes for code families satisfying it under the specified local Davies dynamics. We apply the framework to certify several classes of partially self-correcting quantum memories, including nearest-neighbor models with stabilizer weight as low as six and long memory lifetimes at modest system sizes. These results establish an algebraic-geometric connection between local stabilizer structure and the recoverability of quantum information at finite temperature, providing a systematic approach to certifying and designing passive quantum memories.
Near-Threshold Computing (NTC) offers a substan- tial energy savings, i.e., up to 10x, for digital circuits. However, due to Process, Voltage, and Temperature (PVT) variations, low voltage operating digital processors face increased susceptibility to timing errors. This work equips conventional massive MIMO algorithms with computational error detection mechanisms, en- abling reliable operation down to near-threshold voltages. While conventional hardware based Timing Error Detection (TED) systems introduce substantial overhead, this work modifies MIMO algorithms to seamlessly integrate low overhead Algorithm Based Fault Tolerance (ABFT). A co-simulation framework combining transistor level SPICE simulations and workload aware fault injection is developed to investigate reliability, Bit Error Rate (BER), energy savings, and overheads of modified algorithms operating at low voltages. The transistor-level characterization was performed in PSpice using generic Level-1 Shichman-Hodges CMOS transistor models with Monte Carlo parameter variations. No specific CMOS fabrication process node was assumed. Conse- quently, the reported energy savings are analytical estimates and are not specific to a particular CMOS fabrication technology. We evaluate Gauss Seidel (GS), ADMM based infinity norm detection (ADMIN), and Conjugate Gradient (CG) detectors using circuit derived fault models. Results demonstrate remarkable algorith- mic resilience. Although undetected computational errors in some trials near the Point of First Failure (PoFF) were observed, BER increases by only 0.7% to 1.4%, while beyond PoFF, ABFT reported nearly all erroneous cases. The proposed checksum based ABFT detects most arithmetic and memory errors, with an estimated overhead below 2% for a 64 x 16 MIMO system, decreasing further for larger configurations.
The numerical simulation of kinetic plasma models in six-dimensional phase space poses significant computational and memory challenges, particularly when considering modern high-performance computing architectures. In this work, we propose an adaptive dynamical low-rank approach for the efficient solution of the 6D Vlasov--Poisson system. Our strategy is an extension of the projector-splitting dynamical low-rank integrator. The algorithm changes the rank based on physical observables (e.g., the electric field and energy) as opposed to controlling the error in the distribution function. Our approach is memory efficient in the sense that only rank $r$ objects are required to implement the scheme. This makes the algorithm suitable for performing high-dimensional computations on graphic processing units. Comparisons with rank adaptive Basis-Update&Galerkin integrators exhibit both significant memory savings and reduction in run time. We show, by presenting a variety of 6D benchmark experiments implemented in the C++/CUDA low-rank framework Ensign, that our technique allows us to run simulations with a significantly smaller rank, while still preserving the physical fidelity. Thus, the approach substantially reduces memory footprint and computational cost and enables simulations that are prohibitively expensive for full-rank or less memory-efficient low-rank schemes.
Looped Transformers increase effective depth by re-executing shared layers, but three limitations hinder their adoption: memory grows with the loop count, compute can only be adjusted in whole-block steps, and the loop count is hard to change after training. We propose LooM (Loop-invariant Memory), which addresses each in turn. In a causal encoder-decoder (CED), LooM loops only the decoder layers after the first, which read a global memory written once per token, and each of them publishes the key-value (KV) entry for a token only on its first visit, so the cache does not grow with the loop count. These layers loop independently, providing every budget from 3 to 21 layer executions per token, and one model trained on randomly sampled budgets covers all of them without a router. With enough loops, LooM outperforms a single-pass baseline trained on the same data on language modeling, zero-shot accuracy and long-context perplexity, while using the same cache size. Our analysis further shows that the training budget distribution decides what extra depth buys. These results suggest that a CED is a natural place to loop: memory is written once, and depth becomes a budget chosen at deployment.
Recycling a target model's own predictions offers a training-free source of proposals for speculative decoding. In vision-language models, however, existing recycling indexes predictions by the last token alone, conflating predictions computed under different images and prefixes; we find that such reuse fails most often at image-sensitive decoding states. Longer keys reduce this aliasing but match fewer states. We introduce \method, which stores target predictions under one-, two-, and three-token keys: contextual memories live only for the current request and are thus implicitly conditioned on its image, while token-level memory persists across requests to retain coverage. At each lookup, entropy-confidence weighting fuses the retrieved predictions according to their key specificity and concentration. Best-first search composes the fused transitions into a candidate tree, supplemented by suffix-matched continuations, and one target pass verifies the tree under greedy decoding. Without training or an auxiliary draft model, \method\ achieves mean speedups over autoregressive decoding of $1.66\speed$ and $1.56\speed$ across seven single-turn benchmarks on Qwen2.5-VL-7B and LLaVA-1.6-7B, exceeding the strongest training-free baseline, Token Recycling ($1.51\speed$ and $1.37\speed$). Our implementation is available in https://github.com/Senmo996/RECAP-official-code.
The rapidly growing energy cost of artificial intelligence presents an urgent challenge to conventional computing architectures based on separate processing and memory units. Neuromorphic devices, which integrate memory and logic into a single unit and operate in an event-driven manner, offer a promising alternative. A potential avenue to realize neuromorphic architectures is based on materials that can switch between distinct resistance states reversibly, at low voltage, and at room temperature. Transition metal oxides with oxygen-vacancy driven metal–insulator transitions (MITs) are compelling candidates, but the microscopic mechanisms governing these transitions have remained poorly understood, limiting the ability to design and optimize materials for device applications. In this Account, we describe the authors’ effort at developing, and validating experimentally, a quantitative, first-principles understanding of ion migration and oxygen vacancy engineering in cobaltite perovskites, using La1–xSrxCoO3-δ (LSCO) as the primary model system. We started by asking a basic mechanistic question: what drives the MIT in LSCO? Using density functional theory calculations, we found that the MIT is not simply determined by local bonding changes at oxygen-deficient cobalt sites, but rather by cooperative, global structural distortions. Building on this mechanistic foundation, we developed a model to predict the electrical bias required to trigger the MIT electrochemically. The model expresses the threshold voltage in terms of the oxygen vacancy formation free energy, the high-frequency dielectric constant, and the permanent polarization. The identification of these key quantities enables the prediction of concrete design guidelines for minimizing the operating voltage. We also conducted experimental work to map the topotactic phase landscape of LSCO thin films, where we observed the emergence of the Ruddlesden–Popper phase under the most reducing conditions, extending well beyond the well-known perovskite–brownmillerite transition. This result was also confirmed thermodynamically by first-principles calculations. To further refine experimental work on oxygen-deficient phases, we developed a combined experimental and theoretical protocol for quantifying oxygen stoichiometry in thin films through the analysis of X-ray absorption spectra, providing a practical diagnostic tool. Our investigations were further extended in three directions. First, we discovered that direct hydrogenation using H2 gas provides a low-temperature route to phase control, with hydroxyl bond formation catalyzing oxygen vacancy creation at neighboring lattice sites. Second, we predicted that a metallic interface could emerge between two coexisting insulating phases of LSCO as a result of orbital reconstruction driven by charge transfer and symmetry mismatch. This metallic state, which can be switched on or off by strain or gate voltage, opens a new route to low-energy resistive switching. Third, we showed that the cooperative distortion framework established for LSCO generalizes beyond cobaltites: Grenier-phase iron perovskites R1/3A2/3FeO2.67 exhibit tunable ferroelectricity by the same mechanism of concerted octahedral and tetrahedral unit distortions. Further, combining theory and operando synchrotron X-ray diffraction, we identified “hidden” Grenier phases that form transiently during the perovskite-to-brownmillerite transition in LSCO; their metallic or insulating character depends nonmonotonically on Sr doping concentration, a trend that explains the nonmonotonic dependence of the threshold voltage on Sr content observed experimentally.