laravel-ffmpeg
This package provides an integration with FFmpeg for Laravel. Laravel's Filesystem handles the storage of the files.
10 repos · 26 papers · 62 scholar
Showing 98 of 98 items
This package provides an integration with FFmpeg for Laravel. Laravel's Filesystem handles the storage of the files.
Halo is an open-source framework built by White Circle for training large language and multimodal models
Type-safe tools, on-device inference, multi-agent workflows, memory, and guardrails, in one Swift-native runtime
Exploratory data analysis and advanced analytics in SQL, from the first look at the data to a customer report and a product report. The data and every query.
Optical Neuromorphic Computing for Real-Time Pathfinding: A GPU-Accelerated Eikonal Solver with Directional Memory
Brian2-compatible spiking neural network simulation and native training with a Rust engine, CPU/GPU/browser/MPI backends, AtlasIR, and experimental neuromorphic model conversion.
X-Star Autonomous Brain (HADL v5.8): Non-Token Latent Thought Matrix, Recurrent Toroidal Central Router ("Otak Berputar"), Neuromorphic Spiking Gates & Rigid Symbolic Registers for PyTorch & vLLM.
Subthreshold differential memristor crossbar (analog compute-in-memory) in Cadence Virtuoso, gpdk045: layout, DRC/LVS, PEX
Complete neuromorphic IC design in LTspice & Electric VLSI: 12T LIF neuron, bio-inspired auditory feature extraction, STDP synaptic plasticity, and 0.6V subthreshold optimization across 180nm/90nm/65nm nodes.
Embedded systems rely on sensor measurements that traverse a complex hardware-firmware stack (e.g. peripheral interfaces, communication buses, memory, and drivers) before reaching the applications that use them. At worst, an attacker with access to this path may successfully intercept, overwrite, fabricate, or replay readings. This may cause systems that act on such data, such as flight controllers and industrial processes, to behave incorrectly or unsafely. Existing defences, such as trusted execution environments and trusted I/O paths, still depend on trusted driver software and correctly configured access control, which remain fragile in practice. In response, we present SensorGuard-RV, a hardware-first approach that delivers sensor samples into processor registers without trusting intermediate software, buses, or memory. After configuration, a write-once ownership switch hands the sensor to a fixed-function controller that performs hardware-level authenticated encryption on sensor samples under a key inaccessible to software. We develop a new RISC-V instruction then instructs the CPU to decrypt the sample and writes it to a register only after verifying its tag and rejecting replays, including across resets. We implement SensorGuard-RV on the NEORV32 core with an ADXL362 SPI accelerometer on an Artix-7 FPGA. At 100 MHz, a protected sample reaches a register in 5,191 cycles (51.9 $μ$s), of which the two Ascon authenticated encryption engines account for 30. Because fixed-function acquisition removes per-byte software polling, the protected loop sustains 19,015 transactions/s, 5.3% above a software-driven SPI baseline. The added logic (3,155 LUTs and 2,215 registers; 5.0% and 1.7% of the device) is dominated by the Ascon engine and preserves the 100 MHz clock, while estimated power rises from 118 to 368 mW.
As the computing power density of large-scale AI clusters rises, network communication increasingly limits system performance. GPU-initiated communication requires low-overhead request submission to overlap communication with computation. However, GPUDirect RDMA (GDR) still relies on CPU involvement in its control path, while InfiniBand GPUDirect Async (IBGDA) requires GPUs to assemble Work Queue Entries (WQEs), consuming additional SM resources and memory bandwidth. We propose OpenRDMA, an open-source multi-path reliable RDMA transport architecture designed for GPU-friendly communication. First, OpenRDMA provides a minimal GPU doorbell mechanism: the GPU issues an 8-byte token, and a dedicated FPGA NIC reconstructs the complete WQE from preconfigured context. Second, building on Multipath Reliable Connection (MRC), OpenRDMA supports per-packet multi-path transport and out-of-order arrivals, while selective retransmission, ordered DMA placement, and GPU-visible flag polling provide reliable delivery and safe data consumption. Third, a hardware-software co-designed retransmission mechanism keeps the GPU submission path lightweight and reduces FPGA hardware state overhead: the NIC handles fast data transfer, while CPU software manages complex timeout detection, state tracking, and recovery policies. We describe the architecture and key hardware components of OpenRDMA; implementation of the FPGA prototype is underway.
ML code should run where researchers choose, not where it was first written. Today, machine learning code mainly lives in two ecosystems, PyTorch on NVIDIA GPUs and JAX on Google TPUs. Migrating repositories between them means building and testing across frameworks and hardware. This migration is (1) a practical need, since it unlocks different hardware for researchers, and (2) a natural benchmark: it involves different frameworks and hardware, covers whole repositories, and is long-horizon and challenging. We introduce XStack, a benchmark that evaluates agents on migrating ML repositories between these two ecosystems. Each agent receives a PyTorch repository and a clean TPU machine and must deliver a JAX repository. Its output passes if it meets our grading criteria, based on the source repository's published result, authenticated and calibrated by us. Migration by agents remains unreliable: Claude Opus 5 passes 75%, GPT-5.6-sol 45%, and every other agent at most 40%. XStack is a step toward freeing ML code from its ecosystem. We release XStack to make progress toward reliable migration across frameworks and hardware measurable.
Quantization reduces the memory and bandwidth demands of large language model (LLM) inference but can degrade accuracy; offloading higher-precision data can instead stall generation. We introduce StrataFlux, a framework that challenges this tradeoff by recovering accuracy lost to low-bit weight and key-value (KV) cache quantization while retaining quantization's memory and speed advantages. StrataFlux co-designs refinement representation, selection, and delivery to make host RAM and solid-state drives useful sources of additional precision. Its selective precision streaming reduces transfer volume, overlaps delivery with computation, and reuses fetched precision across decoding steps so generation need not wait for every transfer. We evaluate 2-4-bit weights and 2-3-bit KV-cache on dedicated GPUs and unified-memory platforms. For 2-bit Llama-3.1-8B-Instruct weights, StrataFlux reduces WikiText-2 perplexity from 9.46 to 8.40 while maintaining 95.4% of the quantized base's throughput on RTX 5090. With 2-bit weights and KV, StrataFlux raises Qwen3-8B's AIME 2024/2025 accuracy from 11.0% to 47.1% at unchanged GPU memory use. At 128K context, Llama-3.1-8B-Instruct with 3-bit weights and 2-bit KV achieves comparable perplexity and speed to vLLM's 4-bit AWQ weights and 8-bit KV on RTX 5090, using approximately half the GPU memory. These results show that GPU-resident precision need not set the accuracy ceiling, nor offloading-tier latency dictate decoding speed.
AI agents store text as embedding vectors and compress them at scale. A compression is accepted when each item's nearest neighbors survive, but agents also consolidate, delete, and compare memories, tasks that depend on which items lie far apart. We evaluate both ends with two label-free fidelities from our prior work: the local fidelity $L$ at 10 neighbors and the global fidelity $G$ at 30 to 50 percent of the memory. At 96 bytes per passage, rescored binary codes reach an $L$ of 0.97 yet move 10.1 percent of passages to another theme; PCA with 8-bit codes, at an $L$ of 0.77, moves 1.7 percent. Rescoring leaves the stored codes, and thus $G$, unchanged. We prove that across the $G$ band, $1-ν_k$ is at most 2 to 2.18 times the square root of the near-far rank-violation rate, and that this rate follows the distance noise of the compression (correlation 0.979). Our locality-aware quantization (LAQ) turns the bound into a hinge objective on the codebooks of an OPQ index, keeping its bytes, rotation, and search. To our knowledge, LAQ is the first quantizer whose objective comes from a proven bound on the far end. Reconstruction alone raises $G$ in 0 of 52 held-out runs; the far hinge raises it in all 52. LAQ raises $G$ in all 52 ($p=8.5\times 10^{-4}$ over 13 memory-encoder groups), including the external FinDER memory, has the highest $G$ of seven equal-size quantizers in all 12 Multi-Doc-2025 runs, scales to a million passages, and keeps evidence recall.
We present \textsc{Ginger}, a state-of-the-art open-source C++ library for graph-based approximate nearest neighbor search (ANNS). We developed \textsc{Ginger} at Google to push the speed and memory efficiency of graph search on large and high-dimensional embedding datasets. \textsc{Ginger} surpasses existing ANNS libraries with regards to query throughput and indexing time while matching the memory footprint of the most space-efficient competitors. \textsc{Ginger} builds upon recent advances in proximity graph construction (PiPNN) and vector quantization (TurboQuant). We introduce Nested TurboQuant (NTQ), a two-level quantizer which supports low precision graph navigation and high precision reranking in a single code. This representation achieves near floating point precision with $2.6\times$ less memory. We also compress the size of our proximity graphs by up to $2.5\times$ without impacting query throughput. We combine graph reordering with a new lossless compression codec called AdaptiveGap, which builds upon StreamVByte. At $95\%$ recall, \textsc{Ginger} speeds up query throughput by an average of $1.8\times$ over each dataset's fastest competitor and by at least $2.3\times$ over any single library. This lead grows to $2.9\times$ over compact libraries on high-dimensional embeddings.
LLM pruning must preserve quality while meeting hardware- and workload-specific constraints on latency, throughput, and memory. Existing pruning workflows often target parameter counts, requiring practitioners to manually select model sizes that meet latency, throughput, and GPU memory limits. However, parameter count alone does not predict these metrics, which depend on the model architecture, workload, hardware, and serving engine. We introduce AutoCompress, a hardware-guided compression framework that directly incorporates such constraints into differentiable architecture search. AutoCompress jointly optimizes global and per-layer structure to produce heterogeneous architectures, guided by serving-engine cost models validated against end-to-end latency, throughput, and memory measurements. We evaluate AutoCompress on long-context local coding, batched datacenter serving, and latency-sensitive in-game interaction. Compared to the state-of-the-art Minitron pruning framework, AutoCompress improves average accuracy across eight benchmarks by 6.57, 2.35, and 0.57 percentage points on RTX 5070, RTX 5080, and H100, respectively. AutoCompress also compresses Nemotron 3.5 Super from 120B to 74B parameters for single-B200 BF16 serving at 128K context, retaining 95% of its mean benchmark score.
Self-attention carries a quadratic cost in token count, which limits how far Vision Transformers can scale with input resolution. We introduce SSoG, a separable sum-of-Gaussians spatial operator that replaces dense self-attention with a learned mixture of axis-factorized Gaussian kernels, reducing the per-layer cost from $\mathcal{O}(N^2)$ to subquadratic in $N$. Unlike dot-product self-attention, whose Lipschitz constant is unbounded outside a compact input domain, SSoG with fixed kernels is globally Lipschitz continuous, which, via the Picard--Lindelöf theorem, guarantees that the continuous-depth variant has a unique solution on the whole integration interval, a global guarantee that dense-attention counterparts lack in general. On CIFAR-100 at $224\times224$, in a matched discrete-Transformer setting and in the continuous setting, SSoG improves Top-1 accuracy by 2.2--4.1 points across three model widths while using 8.8--16.6\% fewer parameters, with depth and training recipe held fixed. In a looped-Transformer setting, SSoG improves Top-1 accuracy by 5.4--8.1 points. We further probe inference-time scaling by evaluating on inputs up to $1024\times1024$ at matched batch size: in the continuous-depth (Neural ODE) architecture, SSoG reaches up to 154\% higher throughput and over 60\% lower peak memory than dense attention, with the gap widening as resolution grows, consistent with the difference in asymptotic cost. In the discrete and looped settings, SSoG also achieves over 30\% higher throughput than dense attention at this resolution. Together, these results establish SSoG as a practical, parameter-efficient, and theoretically grounded alternative to dense attention across discrete, weight-shared, and continuous-depth Transformer variants.
We identify an intriguing phenomenon when linearly interpolating between permutation-aligned neural networks: test accuracy briefly improves near the interpolation endpoints before dropping toward the midpoint, which we term BUMP (Brief Unexpected Mid-interpolation Peak). We observe that BUMP appears consistently across datasets, architectures, training regimes, and remains robust under substantial perturbations such as early checkpoints and extreme pruning. Moreover, controlled ablations suggest that it cannot be attributed to trivial regularization effects. While similar near-endpoint gains are visible in prior interpolation plots, they have not been explicitly documented or systematically and thoroughly analyzed. Leveraging this effect, we propose a simple, novel ensembling scheme: weight-prediction mixed ensembling that combines weight mixing with prediction-space mixing. Across our experiments, the resulting ensembling method consistently outperforms standard logit ensembling in test accuracy and yields improved calibration trade-off, highlighting complementary benefits of weight- and prediction-space mixing.
Dense optical flow allows small drones to avoid obstacles. However, few dense optical flow networks are designed for microcontrollers (MCUs) with neural accelerators, whose on-chip SRAM limits the input resolution. We present PACE-NAS, a macro-to-micro neural architecture search framework that decouples memory from latency: a macro design fixes the backbone and its peak SRAM, and a micro design searches operators and depths to trade accuracy against compiler-estimated latency. We apply it to EdgeFlowNet on a Grove Vision AI V2 board (Ethos-U55), where profiling locates the peak SRAM at the full-resolution output head. Replacing transposed convolutions with bilinear upsampling and convolutions lowers the peak SRAM by 18%, and two lightweight context modules lower the endpoint error (EPE) without raising the peak. All searched subnets share this peak, so the search needs no memory constraint. We name the Pareto-optimal models MCUFlowNet; the fastest variant, MCUFlowNet-S, runs 1.76x faster than a retrained EdgeFlowNet at a 1.6% higher EPE and, in simulated forest flights at 2 m/s, reaches the goal without collision in 51.5% of trials, against 41.8%. PACE-NAS also applies to semantic segmentation, where a searched PIDNet-S variant runs 1.17x faster at similar accuracy. Code is available at https://gitlab.kuleuven.be/emedia-research-lab/esat-stadius/mcuflownet/mcuflownet.
Efficient deployment of large language models (LLMs) on mobile devices requires balancing performance with compute and memory constraints. Post-training compression is an effective approach to managing this trade-off, while requiring much less computation than in-training counterparts. We propose DictLLM, a dictionary-based compression scheme in which weight matrices within grouped layers are decomposed into a shared dictionary and layer-specific sparse coefficients, directly obtained from a pre-trained model without further training. Specifically, we cast DictLLM as a four-step post-training optimization. First, we obtain shared dictionaries via singular value decomposition (SVD) of input-scaled weight matrices, amortizing rank across grouped layers; second, we sparsify the layer-specific coefficients to the target compression ratio using a dictionary-aware Hessian; and third, we refine the dictionaries given the resulting sparse coefficients. Finally, we aggressively quantize both components using factor-specific Hessian information: input covariance for the dictionaries and the dictionary-aware Hessian for the coefficients. The resulting representation reduces memory footprint by storing a single shared dictionary per layer group, along with sparse layer-specific coefficients. On language modeling and zero-shot reasoning benchmarks, DictLLM consistently outperforms existing post-training low-rank compression methods and pushes the Pareto frontier between model size and performance. DictLLM compresses the original LLaMA-7B model by 5.1x with only a 2.9-point increase in WikiText-2 perplexity, achieving a 3.6x higher compression ratio than the state-of-the-art low-rank method.
Quantization and pruning reduce the cost of deploying language models, but their effects are usually reported as benchmark scores. Such scores record only the loss that remains at the output. Inside the model, compression damages some components more than others, and the rest of the network can partly compensate. The observed loss therefore reflects both the damage and this compensation, which benchmark scores cannot separate. How the two combine, and whether their balance changes with model size, remains unclear. We study this compensation for in-context retrieval, a capability carried by a small, identifiable set of attention heads. To separate the two, the expected loss from ablating these heads is decomposed exactly into the part restored by the rest of the network, attributed to the components responsible, and the part that reaches the output. This decomposition is then compared across model sizes. For damage of comparable size, smaller models restore most of the loss, whereas larger models retain most of it, because the compensating responses offset less of it as models grow. Compressing the heads produces the same pattern. Compression that is nearly harmless in small models can therefore be costly in large ones.
Large Language Models (LLMs) incur severe memory-access overheads due to the cost of loading large weight matrices and key-value (KV) caches. Analog In-Memory Computing (AIMC) is an emerging solution that enables computation directly on static weights stored in memory, substantially reducing energy and latency. 3D-IMC further stacks memory arrays into vertical tiers, increasing capacity but revealing a new challenge: how weights are placed in memory determines which matrix-vector multiplications (MVMs) can execute concurrently, affecting latency and throughput. The concurrency aspect of 3D-IMC weight-mapping remains underexplored in the literature, leading to sub-optimal utilization of compute resources on 3D-IMC chips. In this context we propose CAMP (Context-Aware Mapping and Placement), an inference-context-aware weight mapping algorithm for 3D-IMC-based LLM accelerators. CAMP captures the potential resource contention resulting from the architecture, targeted network, and runtime context in a conflict matrix, enabling throughput-optimized weight mapping. This mapping is refined through a profiling-driven feedback loop that efficiently prunes the search space and returns optimized mappings within practical runtimes. CAMP is an offline, software-based solution requiring no hardware changes and improving 3D-IMC compute utilization from 42.4% to 60.3%, delivering up to 2.0x higher 3D-IMC throughput and 1.74x higher end-to-end throughput over state-of-the-art mapping algorithms.
Expert pruning compresses Mixture-of-Experts (MoE) language models by removing the experts that a calibration set marks as unimportant. Studies of high-resource languages largely find that English calibration transfers well to other languages. We test this for Lithuanian: we pruned four MoE models with REAP, removing 25% or 50% of the experts with English, Lithuanian or mixed calibration, and evaluated both languages on reading comprehension, translation and language modelling. English calibration does not transfer to Lithuanian: at 25% pruning, English comprehension fell by a median of 1.7 points but Lithuanian by 21.4. The transfer is asymmetric. Lithuanian calibration removed most of the Lithuanian loss and never significantly hurt English comprehension, which was on average 2.5 points better than under English calibration; only English language modeling got slightly worse. Routing suggests why: in the last layers, Lithuanian sends 55-72% of its expert selections to just eight experts, against 19-39% for English, and English calibration removes some of these experts. Compressed models should be calibrated, and evaluated, in every language they must serve.
Millimeter-wave radar resolves the micrometer displacements of a moving body as micro-Doppler, enabling gait, gesture, activity, pose and vital-sign sensing. Each task relies on a downstream model trained on real data, yet radar datasets lag orders of magnitude behind vision and language. Brute-force collection will not close that gap, and existing synthesis methods (forward-only simulators, learned generators and static-scene inverse renderers) cannot fit a real capture of a moving body, leaving a sim2real gap. We present 4D Point Splatting (4DPS), a physics-based differentiable renderer that fits an explicit moving scene to a real capture. Using the synchronized camera to initialize the room and body, we represent the scene as oriented points carrying material, normal and velocity, and splat each point into range and Doppler bins. Our output is the raw complex return a real radar records, so we apply the same signal processing on the rendered data to obtain any representation for training any SOTA downstream model. We fit 235 sequences on HuPR and 100 on RT-Pose, two datasets with different radars. Our reconstructions set the state of the art, with median correlations of 0.940 and 0.830 on range-azimuth, 0.829 and 0.695 on range-Doppler, and 0.988 and 0.966 on micro-Doppler. Our renders also set the state of the art in downstream sim2real gap. On RT-Pose, a pose model trained on them reaches 21.9 cm MPJPE on real data, against 16.7 cm for real training and 127.6 cm for the RF-Genesis baseline, where both the rendered- and real-trained models learn nearly the same features. Our custom CUDA kernels fit a sequence in 5.7 minutes on HuPR and 12.9 on RT-Pose on one H100, making dataset-scale inverse rendering of mmWave radar possible for the first time. This work can enable synthetic data generation for any human-centered downstream task and model, and scene editing for data augmentation.
Existing context models for learned image com- pression rely primarily on internal priors from the current image, overlooking common patterns and typical textures shared across natural images. To address this limitation, we propose an Internal-External Prior Fusion-Alignment (IE-PFA) context model built upon quadtree spatial-channel joint modeling, where a learnable dictionary shared by the encoder and decoder serves as an external prior. Specifically, an RWKV-based Spatial- Channel Prior Fusion (RSCPF) module efficiently fuses inter- nal and external priors along spatial and channel dimensions with linear complexity, while a training-only Self-Distilled Prior Alignment (SDPA) mechanism aligns them without inference overhead. Moreover, our model supports variable-rate coding over 64 quality levels within a single network and is accelerated by custom CUDA operators. Experiments on seven benchmarks demonstrate state-of-the-art rate-distortion performance, achiev- ing an average BD-Rate saving of 21.65% over VTM-22.0-Intra, with only 29.9M parameters and the fastest coding speed among advanced neural codecs (61 ms encoding and 52 ms decoding). Code is available at https://github.com/stcnjtech/DCVC-X-Intra.
In large language model (LLM) serving systems, the key-value (KV) cache grows linearly with sequence length and concurrency, so long-context inference easily hits the high-bandwidth memory (HBM) capacity wall and results in low concurrency. Offloading the KV cache to host dynamic random-access memory (DRAM) frees HBM space and raises concurrency, but the additional cache management and host-to-device (H2D) load overhead inflates inference latency, higher concurrency fails to translate into effective throughput. We present FlashOffloading, a host-DRAM-centric inference architecture that fully offloads the KV cache in both the prefill and decode phases. FlashOffloading builds rolling HBM cache for prefill, a hot cache for decode, a reverse sparse index with zero extra miss-lookup overhead, and a compute--load overlapped kernel, optimizing KV offloading latency across the entire pipeline and delivering throughput gains under a time-per-output-token (TPOT) service-level objective (SLO). In experiments deploying GLM-5.2 with multi-token prediction (MTP3) speculative decoding on an Ascend910C prefill--decode (PD)-disaggregated platform, FlashOffloading improves prefill throughput by up to 1.38\(\times\) and decode throughput by up to 1.91\(\times\) under a 30 ms TPOT constraint. FlashOffloading is integrated into an open-source serving framework: https://github.com/vllm-project/vllm-ascend/tree/releases/v0.30.0.
Low latency is critical to user experience. As NVIDIA Blackwell and Rubin GPUs advance multi-die scaling, data placement and cross-die communication become increasingly important. However, the latency implications and underlying costs of cross-die communication and die-aware execution remain largely unexplored. In this report, we present a comprehensive evaluation of the Locality Domain (LD) feature introduced in CUDA 13.4. Conducted on a single node equipped with eight NVIDIA B300 GPUs, our study examines three aspects: memory management and access, computation, and communication, to identify the sources of latency and, more importantly, explore how LD can mitigate them. Our measurements show that the benefits of LD depend on data access patterns. Through die-aware adaptations and integration with widely used libraries and tools, we achieve speedups of up to 43.5% for NCCL 2.32.3 (AllGather), 111.3% for nvbandwidth (DtoD), and 37% for DeepGEMM (EP). We also identify non-negligible LD-related GPU memory allocation overhead when allocation sizes exceed 32 MB due to barrier and TLB invalidation, we propose a fix to the GPU driver kernel module (submitted to NVIDIA) and show that overhead could be reduced 7x ~ 10x without requiring any user-code changes.
The Common Community Physics Package (CCPP) is a framework designed to simplify the exchange of physics packages between state-of-the-art atmospheric models. A core software component of CCPP is a code generator: a bespoke Python-based application that uses regular expressions to transform an input specification into Fortran source code. While the code generator initially provided a simple transformation capability, the demands of state-of-the-art atmospheric models have resulted in significant increases in its complexity. Furthermore, ongoing code modernisation efforts to enable GPU-based computing have placed additional demands on the code generator for which it was never designed. Domain Specific Languages (DSLs) have proven to be highly effective in High Performance Computing (HPC), offering the ability to simplify code modernisation efforts. However, traditionally the effort and specialist skills required in the development of a DSL have been a blocker to adoption. In recent years, however, there have been great advances in compiler technologies, and frameworks such as MLIR and xDSL have the potential to greatly lower this barrier to entry. In this paper, we present xDSL-CCPP, a new application-specific code generator based upon the latest MLIR and xDSL compiler technologies, for CCPP. Demonstrating a reduction in bespoke tooling code of around three times, from approximately 15,000 LoC to 5,000 LoC, we highlight the opportunities provided by these compiler frameworks and demonstrate how they significantly lower the barrier to development of HPC application-specific DSLs whilst maintaining application performance and promoting flexibility.
Attention is a core building block of frontier models and often dominates compute-bound prefill at long context. FP8 matrix instructions offer twice the peak throughput of BF16, but they accelerate only matrix multiplication, and the softmax between the two matrix multiplications of attention runs on vector units that FP8 does not accelerate. On the AMD Instinct MI355X, the vendor's AITER kernels reach 39% of peak FP8 matrix throughput at 32K tokens, compared with 49% in BF16. Attention is also more numerically sensitive to quantization than feed-forward layers, and unlike NVIDIA's Hopper and Blackwell GPUs, the MI355X does not track the completion of matrix instructions in hardware, so overlapping the softmax with the matrix multiplications depends on instruction order. We present ZYFA, prefill kernels for the MI355X built on three techniques: (1) computing each FP8 probability directly from its score with a multiply-add and an integer conversion instead of an exponential; (2) a ping-pong schedule in which one group of threads computes its softmax while another executes its matrix multiplications; and (3) scaling each block of probabilities up by a power of two before rounding, and normalizing by the sum of the rounded probabilities. On attention inputs captured from three models and timed at sustained clock speeds, ZYFA achieves 1.16-1.76x the throughput of AITER. On the Gaussian inputs of kernel benchmarks at 32K tokens, it achieves 1.26-1.37x at head dimension 128 and 1.99x for DeepSeek-style multi-head latent attention. The sustained clock of the MI355X depends on operand values, so throughput on random inputs does not reliably predict throughput on real model activations. At head dimension 128, ZYFA's numerical error is within a few percent of that of Composable Kernel and Triton and 62-78% lower than that of AITER's. The source code is available at https://github.com/Zyphra/zyfa.
Zirah is a hardware signer built around an STM32F407 microcontroller, and its purpose is to find out, on inexpensive hardware, where a post-quantum signing key is exposed and what actually protects it. The device draws every random bit from a reverse-biased 2N3904 junction in avalanche breakdown, amplified by an LM358 and screened by SP 800-90B health tests, computes ML-DSA-65 (FIPS 204, hedged) on the STM32 with two-share arithmetic masking of the secret-key multiplications, and adds ECDSA P-256 from an ATECC608 secure element over the same digest, once a person has matched an eight-digit digest prefix on its OLED with the PC and pressed a button. A desktop application and a command-line tool drive the board and verify its output, and because only a 32-byte manifest digest reaches the device, it signs files of any type and size, returning signed copies for PDF, Office Open XML, ZIP, PNG and JPEG files and detached records for every other format. NIST's SP 800-90B tool rejected the IID assumption for 1,000,000 raw samples and assessed 6.898 bits of min-entropy per byte on the non-IID track, against the 1 bit the firmware credits, while spectral lines near 116 Hz that hold about 5 % of the variance remain unexplained. At 168 MHz the masked signature took a median of 157 ms over 20 runs and a complete signing operation under 0.8 s, and the signer matched the pq-crystals reference in 1,976 checks natively and on an emulated Cortex-M4. The ATECC608 can neither store the 4,032-byte ML-DSA-65 key nor compute with it, so that key sits in ordinary flash behind readout protection, and on an earlier board, before that protection was enabled, a debugger copied it out and the copy produced a valid ML-DSA-65 signature without the device, although the leakage results come from an instruction-level emulator and nothing here is certified.
The rapid expansion of the Internet of Things (IoT) has increased the volume of data generated at the network edge, where many devices operate under severe memory, processing, and energy constraints. In this context, Tiny Machine Learning (TinyML) enables machine learning (ML) tasks to be executed on microcontrollers, smart sensors, and other resource-constrained platforms. However, the literature remains heterogeneous regarding the algorithms evaluated, energy-measurement methods, reported metrics, influencing factors, and optimization strategies. This study presents a systematic mapping of energy-aware TinyML research. It identifies the main models and learning approaches, classifies energy-evaluation methodologies and metrics, examines the technical factors affecting energy consumption, and organizes the reported optimization techniques into a structured taxonomy. The findings highlight the frequent use of convolutional and lightweight architectures, direct measurements and analytical estimations, and factors related to model architecture, hardware, communication, numerical precision, and input characteristics. The optimization strategies were grouped into model compression, hardware-aware design, system and communication optimization, and data and processing optimization. The results also reveal limited standardization in experimental platforms, measurement procedures, and reporting practices, limiting reproducibility and direct comparison across studies. This mapping provides an updated overview of the field and guidance for developing and evaluating more energy-efficient TinyML solutions.
Analog compute-in-memory (CIM) enables energy-efficient AI inference by performing in-situ matrix multiplication. The growing demand for deploying larger models motivates chiplet integration to expand CIM capacity. However, simulation abstractions based on local operator costs and aggregate communication metrics obscure the impact of weight-mapping topology when CIM system scales up. We develop a phase analysis that quantifies how strongly mapping topology affects system performance across hardware configurations and how mapping optimization translates into latency gains. To evaluate and exploit this design space, we present Isobar, a simulation platform integrating CIM decompositions, heterogeneous execution, and resource modeling. Validated against silicon measurements and cycle-accurate simulation, Isobar provides reliable cost estimates that enable effective mapping search across model and hardware configurations. For Gemma-7B prefill, Isobar-guided mapping reduces communication time by 23.3% and end-to-end latency by 18.6% relative to the strongest baseline. Isobar's system-level modeling show that analog CIM retains energy-efficiency potential relative to GPU when scaled to thousands of dies, and identify topology-aware mapping as a key design dimension for efficient execution.
Complete-block offloading reduces model residency but requires each transferred block to fit the GPU budget. We present Precision-Preserving Subtensor Inference (PSI), a fixed-arena runtime that combines sub-block weight streaming with separately managed intermediate state for the BF16 LTX-2.5 22B joint audio-video transformer. The contribution is a construction whose managed GPU footprint is determined by reusable pages, live state, and bounded workspace, together with controls that separate resource use, transport correctness, and numerical fidelity. Its partitioned operators preserve the dense real-arithmetic construction under stated assumptions. Across nine matched full-prompt cases on RTX A6000, modified PSI reduces sampled transformer-process peak memory from 8,578 to 360 MiB relative to official MATH, while mean time for the eleven transformer invocations rises from 247.11 to 349.00 s. Its managed arena is 80 MiB. Complete native generation and a complete-block FLASH path both succeed under a fixed 3 GiB available-memory cap. The matched A6000 comparison and separate DGX Spark service suite both achieve 100% success in complete audiovisual decoding. FFN reuse lowers mean block time by 4.0% with unchanged BF16 outputs; a fixed-arithmetic one-token transport control preserves all saved states bitwise. Targeted bias and attention changes improve local and one-token numerical comparisons, while mean endpoint error increases across nine full-prompt cases. The construction enables BF16 audiovisual inference with an explicit transformer-memory budget and makes numerical fidelity a separate, testable property of the implementation.
Multisensory Physical AI systems operating in isolated conditions must make timely and reliable decisions within strict energy and hardware budgets. In this context, the present analytical study investigates the feasibility of reliable multimodal sensor fusion using a neuromorphic approach based on memristor-based Integrate-and-Fire neuron circuits driven by aggressively compressed representations of sensory information. Rather than relying on fixed processing and learning parameters, the proposed framework introduces scenario-dependent adaptability into individual stages of the neuromorphic decision-making pipeline, including sensor-data compression, nonlinear spike encoding, reliability assessment, and memristive learning-forgetting dynamics. By coupling multiple sensor inputs through adaptive memristive synapses, the framework investigates how reliable and correlated observations can strengthen learned associations, while weak or obsolete relationships gradually decay through nonlinear forgetting. The resulting neuronal responses are evaluated across ten representative driving scenarios derived from the NuScenes dataset by comparing their encoded output patterns with predefined synthetic ground-truth codes representing the expected actuator responses. The simulations demonstrate that scenario-adaptive activity-vector sizing achieves a 61.5% reduction in retained sensory information, while improving mean bit-wise agreement from approximately 55% to 61.2% compared with a fixed maximum-resolution configuration. These findings highlight the potential of introducing scenario-aware adaptability directly into the building blocks of a neuromorphic system, providing a foundation for future hardware implementations that could significantly transform the capabilities of Physical AI systems operating under stringent resource and environmental constraints.
Tensor Network (TN) kernel machines accelerate supervised learning by using low-rank TNs to reduce computation and memory. However, most TN-based methods are deterministic with manually tuned ranks and feature dimensions, while the existing Bayesian TN kernel machine is restricted to the CANDECOMP/PARAFAC decomposition (CPD), whose single rank is shared by all modes. We extend this framework to the tensor train (TT) and propose Bayesian Tensor Train Kernel Machines (BTT-KM), whose hierarchical ARD-type priors on the TT cores infer the TT ranks and feature dimensions from data. Since each TT core is coupled to the others only through its neighbors, its closed-form posterior update requires propagating the contributions of all other cores along the chain. Mean-field variational inference yields a Bayesian alternating least squares algorithm with the same computational complexity as its deterministic counterpart and a closed-form predictive distribution. A synthetic experiment illustrates that BTT-KM recovers the true TT ranks and feature dimensions from an over-parameterized initialization.
Internet of Things (IoT) sensors and devices heavily rely on cloud-based processing, especially in healthcare monitoring, for sensor data analysis, which incurs high data transfer costs and energy consumption. To reduce constant data transmission, always-on smart sensors can use lightweight brain-inspired hyperdimensional computing (HDC) models to efficiently classify sensor data that require further resource-intensive cloud processing. The proposed architecture maintains HDC algorithm accuracy with 8-bit integer precision for complex datasets, where 1-bit quantization is not feasible. We propose SenseHD, an all-digital CMOS always-on smart sensor accelerator for energy-constrained smart sensing applications. SenseHD achieves an extremely small power consumption of 39.21– $92.59~\mu$ W depending on design parameterization, supporting 8-bit hypervectors (HVs) and an energy-efficient near-memory fused item memory (IM) + multiply–accumulate (MAC) systolic array encoder. We introduce a novel systolic similarity search unit (SSU) that utilizes a small systolic array to calculate HV cosine similarity metrics, consuming only 110.89 nJ/inference. By intelligently transmitting only relevant data, SenseHD reduces system-level energy consumption by up to $8.8\times $ compared to a cloud-centric baseline. Notably, the accelerator overhead is negligible, consuming only 1.2% of the total system energy.
In this paper and its companion [DEP2] we construct and study special generating sets for $p$-arithmetic unitary Lie groups, extending the seminal works of Lubotzky, Phillips and Sarnak to higher dimensions. The real embeddings of our sets form optimal topological generators for compact unitary Lie groups, and their $p$-adic embeddings give rise to Cayley graphs and complexes over finite unitary groups, which exhibit the cutoff phenomenon. In the current paper we prove two main results: First, we construct $p$-arithmetic lattices in $PU_4$ and $PU_8$ which act transitively on their corresponding Bruhat-Tits buildings. These groups are used to construct Golden Gate sets on two and three qubits, which have optimal circuit length and efficient compiling algorithms. Second, we study the cutoff phenomenon for quotients of the Bruhat-Tits building of a general semisimple $p$-adic group. We show that cutoff occurs for quotients which satisfy either the Ramanujan property or the weaker Sarnak-Xue Density Hypothesis (SXDH). For this purpose we bound the operator norms of a large class of Hecke operators in terms of the rate of decay of unitary representations. In the companion paper [DEP2], we prove such a SXDH result for unitary groups.
Electrochemical sensors are becoming increasingly important in a wide range of applications, generating a demand for highly optimized electronic interfaces. Development of such interfaces with high sensitivity and robustness, while maintaining low power consumption, depends on the ability to accurately model and simulate the behavior of the sensors being interfaced with. Electrochemical systems are difficult to model using standard electrical components in conventional circuit simulation frameworks, due to complexities such as nonlinear behavior and memory effects. Hardware description languages (HDLs) can help overcome this limitation by allowing for behavioral modeling of electrochemical systems based on underlying differential equations, within circuit simulation frameworks. This approach is used to model an electrochemical sensor, using Verilog-A, which is simulated along with a CMOS-based potentiostat interface. The model exhibits expected behavior in both time and frequency domains, with orders-of-magnitude performance improvements over previous implementations. The time-domain simulation outputs are compared to experimental data, with good agreement observed. The ability of the model to simulate complex electrochemical phenomena such as redox cycling is also demonstrated.
Low-light conditions can degrade agricultural vision imagery, while edge camera nodes may have limited computational resources. This study developed and characterized a streaming, fixed-point FPGA datapath for Retinex-inspired low-light image enhancement and evaluated the underlying algorithm in software. The register-transfer-level (RTL) design combines maximum-channel illumination estimation, edge-aware refinement with piecewise linear Gaussian weighting, table-based reflectance recovery and gamma remapping, and batch-normalization-style output normalization. It processes one pixel per clock cycle without a frame buffer. Functional simulation matched a bit-accurate arithmetic model at every compared pixel. Open-source synthesis, placement, and routing targeting an AMD Xilinx Artix-7 FPGA yielded 5901 LUTs and met a 100 MHz timing constraint. Image quality was evaluated only in software. On 15 LOL-v1 test pairs, a floating-point reference model differing from the RTL in two processing steps improved mean PSNR from 7.77 to 15.03 dB; an evaluation using the RTL equations achieved 9.35 dB, increasing to 16.64 dB after reselecting a normalization constant. The findings indicate a compact streaming realization of the pipeline, while highlighting the sensitivity of image quality to normalization. The design was not programmed or tested on a physical FPGA board, and real agricultural low-light images were not evaluated.
This paper presents a polynomial I/Q envelope generator for superconducting qubit control that requires no waveform lookup table. The architecture computes DRAG-like in-phase and quadrature envelopes from a compact support polynomial, without storing pulse samples. Programmable parameters control pulse duration, in-phase amplitude, and quadrature strength. A 16-bit VHDL implementation on a Zynq UltraScale+ FPGA meets a 500 MHz timing constraint after placement and routing, using 57 LUTs, 20 flip-flops, and 5 DSP blocks. The generated pulse family is compared with Gaussian-DRAG through normalized spectral analysis and closed-system simulations of a three-level transmon. These results characterize the hardware cost and simulated quantum control performance of the proposed generator.
This paper presents the broadband characterization of Thin-Film Resistors (TFRs) embedded in an RF silicon interposer for mm-Wave heterogeneous integration. On-wafer S-parameters measured up to 110 GHz evaluate the high-frequency stability and parasitic effects of the Titanium Nitride (TiN) films. To facilitate RFIC design, a geometry-scalable lumped-element π-network model is developed. Intrinsic sheet resistance and parasitic reactances are extracted from de-embedded measurements, demonstrating highly stable, non-dispersive behavior deep into the W-band. These parameters are seamlessly integrated into a Verilog-A module. The model’s predictive accuracy is validated against unseen intermediate geometries, showing an excellent correlation with the measured data. Ultimately, this work provides designers with a reliable, ready-to-use tool for wideband systems on RF silicon interposers.
All-Digital Phase-Locked Loops (ADPLLs) have emerged as an effective alternative to conventional analog PLLs due to their superior scalability, robustness against process–voltage-temperature (PVT) variations, and compatibility with advanced CMOS technologies. This paper presents the design, simulation, synthesis, and physical implementation of a fully synthesizable Time-to-Digital Converter (TDC)-based All-Digital Phase-Locked Loop comprising a TDC-based phase-frequency detector, a digital proportional-integral (PI) loop filter, an NCO-based digitally controlled oscillator (DCO), and a programmable feedback divider for integer-N frequency synthesis. The novelty of the proposed work lies in the realization of a complete RTL-to-GDSII implementation flow in 45 nm CMOS technology, enabling efficient integration into modern SoC designs. The proposed ADPLL achieves frequency locking within $10.25 \mu \mathrm{s}$ with a residual phase error of ±6 LSB, peak-to-peak jitter of 0.30 ns, RMS jitter of 50 ps, and frequency stability of 99.95%. Post-layout results demonstrate a compact area of $\text{2215.382} \mu \mathrm{m}^{\mathrm{2}}$ using $\text{91}$ standard cells, total power consumption of $\text{90.69} \mu \mathrm{W}$, and positive setup and hold timing slack of +7.502 ns and +1.891 ns, respectively, with zero timing violations. Compared with existing ADPLL implementations, the proposed design achieves lower power consumption, reduced implementation complexity, and reliable closed-loop operation, making it suitable for low-power and scalable clock generation applications in modern VLSI systems.
The growing demand for real-time deep learning applications has significantly increased the need for efficient hardware accelerators capable of delivering high performance with low latency and minimal resource utilization. However, conventional CPU- and GPU-based implementations of Convolutional Neural Networks (CNNs) suffer from high power consumption, increased latency, and inefficient hardware utilization, making them unsuitable for resource-constrained and edge-based systems. To address these limitations, this paper presents a resource-efficient pipelined FPGA accelerator for CNNs using fixed-point optimization. The proposed design focuses on optimizing core CNN operations, including convolution, activation, and pooling, through a modular and scalable hardware architecture. A pipelined dataflow mechanism is introduced to enable parallel processing and reduce computational delay, while fixed-point arithmetic is employed to minimize hardware complexity and improve resource efficiency. The architecture incorporates optimized multiply-accumulate (MAC) units for convolution, comparator-based ReLU activation for low-overhead non-linearity, and an efficient max-pooling unit for feature map reduction. The complete design is implemented using Verilog HDL, simulated using Xilinx Vivado, and synthesized using Xilinx 14.1 tools to validate functional correctness and hardware performance. Experimental results demonstrate efficient utilization of FPGA resources, with the design occupying only 188 LUTs, 132 registers, and 9 DSP slices. Timing analysis confirms stable operation with all constraints satisfied and a positive worst negative slack (WNS), ensuring reliable high-speed performance. Compared to conventional implementations, the proposed accelerator achieves reduced hardware area and improved processing efficiency. These results highlight that the proposed FPGA-based CNN accelerator provides an effective balance between performance, resource utilization, and power efficiency, making it highly suitable for real-time embedded and edge AI applications.
A house in today's world is packed with several devices, but most of them do not have their power consumption monitored in real time. Such a situation results in unnecessary wastage and even delays in detecting defective equipment. This paper proposes an energy monitoring framework that monitors power consumption of individual appliances within an edgebased smart home environment. The architecture incorporates ACS712 current sensors for load current measurement and ZMPT101B AC voltage sensors for grid voltage measurement, which are connected to an ESP32 microcontroller to collect voltage, current, and power data. These acquired electrical parameters are then conveyed to the processing unit on the PYNQ-Z2 Field Programmable Gate Array (FPGA) board. The cost function is implemented using Verilog on Vivado, with the output being displayed as the tariff so that power monitoring can be performed in real time.
This paper describes an FPGA-based edge vision node for Industrial IoT front ends. The design integrates OV5640 image acquisition, SCCB camera configuration, SDRAM buffering, RGB-to-gray conversion, Sobel edge extraction, and LCD output in a single timing-coordinated hardware path. Rather than treating the edge detector as a stand-alone kernel, the implementation emphasizes the complete sensor-buffer-process-display chain required by an embedded vision node. The prototype is evaluated using board-level functional outputs, Modelsim waveforms, RTL integration evidence, and Quartus synthesis results. The implementation occupies 2027 of 10320 logic elements on a Cyclone IV E EP4CE10F17C8 device and uses no embedded multipliers, indicating that the complete pipeline can fit in a modest FPGA while leaving headroom for later extensions.
This paper presents the full RTL-to-GDSII implementation of a 1X4 packet-switching router in a 45 nm technology node. The router architecture comprises synchronous FIFO buffers, a synchronizer for clock-domain alignment, a finite-state machine for packet routing, and registers for header storage. RTL implementation is done in Verilog and functional verification is performed using Universal Verification Methodology (UVM). The physical design flow is carried out using Cadence Genus for synthesis and Cadence Innovus for placement, clock tree synthesis (CTS) and routing. A major focus of this work is timing optimization through CTS, which achieves 100% removal of setup violations and an 89.5% reduction in hold-violating endpoints from pre-CTS to post-CTS stages. The final design uses 2,094 cells, occupies $6589.61 \mu \mathrm{m}^{\mathrm{2}}$, and consumes $\text{8. 0 2} \mu \mathrm{W}$. This work demonstrates a complete, industry-standard RTL-to-GDSII methodology for packet-switching routers with explicit emphasis on CTS-driven timing closure in an educational 45 nm node.
Leg-Arm Reuse Robots (LARR) employ a single pair of limbs for both locomotion and manipulation, requiring dynamic switching between compliant and rigid physical regimes. Standard imitation learning (IL) outputs kinematic trajectories but lacks an explicit mechanism to schedule the mutually exclusive physical constraints, while purely control-based pipelines schedule constraints through rules fixed at design time. We propose the Physically-Grounded Learning and Control (PGLC) framework, which vertically integrates executable physical primitives across the entire pipeline. Rather than learning physical abstractions from scratch, PGLC jointly learns motion and constraint scheduling from perception. Specifically, the policy learns when each expert-designed primitive should be activated. These primitives are injected into demonstration data by an enhanced teleoperation interface, learned as scheduling flags by a high-level policy, and instantiated in real time by a dynamically configurable quadratic programming (QP) controller. On a physical LARR prototype, PGLC achieves a 0.925 success rate on speed bump terrain, compared with 0.55 for the strongest learning baseline, and a 0.875 overall success rate on the bimanual pick-pour-place task, where baselines without scheduling remain below 0.45. Ablation studies attribute these gains to the physically-grounded data and the dynamic scheduling mechanism.
This paper presents the design and implementation of a multi-layer privacy-preserving network router built on the Raspberry Pi 5 platform. The proposed system integrates several complementary privacy-enhancing technologies, namely the TOR anonymization network, encrypted DNS protocols (DNSCrypt, DNS over HTTPS, DNS over TLS, and Oblivious DNS over HTTPS), a local recursive DNS resolver (Unbound), and DNS-level filtering via Pi-hole. A total of nine distinct operational modes were developed, each offering a different trade-off between the level of privacy protection and network performance — ranging from basic advertisement filtering to full traffic anonymization combined with multi-layer DNS encryption. A physical control interface featuring an OLED display and push buttons was implemented, enabling mode switching without remote shell access. Performance evaluation using the DNS Performance Testing Tool (dnsperf) demonstrated that even the most advanced configurations maintain acceptable response latencies and high query resolution rates. The proposed solution addresses the gap between expensive, closed-source commercial devices and complex open-source systems that require advanced technical expertise, offering an accessible and cost-effective privacy protection tool for individual users and high-risk groups. The total hardware cost does not exceed 200 USD, representing a fraction of the price of commercial solutions with comparable functionality.
Ensuring reliable contact execution across different peg-in-hole assembly tasks remains a key challenge in robotic assembly. This paper proposes a hierarchical learning framework for multi-category peg-in-hole assembly that functionally separates long-horizon geometric motion generation from contact-level physical adaptation. A task-conditioned multimodal diffusion policy learns a shared geometric prior from multi-task demonstrations and generates smooth, geometrically feasible motion trajectories. During physical interaction, a contact-triggered residual adapter trained via force-constrained reinforcement learning performs bounded local pose correction and stiffness modulation. This design preserves cross-task reuse of the geometric prior while localizing online contact adaptation to residual adapters trained separately for each task. Real-robot experiments across nine geometric and connector insertion tasks show that the proposed method improves assembly success and interaction robustness over learning-based and control-based baselines, while component ablations and sensitivity analyses further clarify its practical scope.
Ultra-high field (UHF, B0 ≥ 7 T) MRI provides substantial gains in signal-to-noise ratio, spatial resolution and contrast. Routine clinical deployment has been limited by challenges of transmit field ( B 1 + ) inhomogeneity and specific absorption rate (SAR). Parallel transmission (pTx) can mitigate these challenges but it requires per-channel complex B 1 + maps to design effective pTx pulses. We review established and emerging approaches to obtain these maps. We begin by summarising the underlying physical principles and theory. Then we describe methods for magnitude B 1 + mapping such as actual flip-angle imaging (AFI), saturation-prepared B 1 + mapping (e.g., satTFL) and dual refocusing echo acquisition mode (DREAM) that are the basis of widely-applied per-channel complex B 1 + mapping methods, as well as methods of pedagogical and/or historical relevance. Next, we describe how relative B 1 + mapping with low-flip angle gradient-recalled echo (GRE) scans can determine the relative phases (and relative magnitudes) of B 1 + for each channel. We recap the important interferometric approach that is now widely used. We explain how magnitude and phase maps are combined into complex per-channel B 1 + maps. We discuss remaining arbitrary factors such as the position-dependent transmit reference phase. We discuss recent practical implementations such as Sandwich TFL, 3DREAM, compressed-sensing SA2RAGE and respiratory-resolved B 1 + mapping. We explore the workflow improvements that could accrue if per-channel B 1 + maps do not need to be acquired for every subject. This includes deep-learning-based approaches to synthesize B 1 + maps from the normal localiser scans, the universal pulses (UPs) approach to precalibrate pulses and variants such as standardised universal pulses (SUPs), Fast-Online Customised (FOCUS) and machine-learning online customisation. We hope to encourage greater openness and harmonisation of these critical calibration steps for parallel transmit pulse design. This will become increasingly important as 7-T pTx MRI is adopted for multisite cross-vendor clinical research studies and to deliver the full scientific potential of new extremely high field human MRIs.
Vision-language-action (VLA)-based embodied agents combine large AI models with physical perception and actuation, creating safety and security risks when unauthorized actions lead to physical hazards. These systems therefore require context-based access control that accounts for runtime conditions while meeting latency constraints. However, existing work has focused primarily on access policies, leaving the system-level implications of alternative access-control designs underexplored. This short paper implements a context-based authorization framework for VLA-based embodied agents and evaluates motion thresholds, session key validity periods, and centralized versus decentralized authorization deployments. Our results quantify how these design choices affect authorization frequency and end-to-end latency and show that decentralized authorization substantially reduces latency and variability.
Metamaterials offer unprecedented control over physical fields through their microstructure, but current design methods cannot meet the demands of scenarios involving active sources, multiphysics field coupling, and multifunctional requirements. Moreover, existing approaches are generally developed for specific functions, making it difficult to integrate different functionalities and select the functional priorities within a single structure. To overcome these limitations, this study presents a systematic inverse design framework for multifunctional multiphysics metamaterials via multiobjective optimization. To uniformly explore the Pareto front, which clearly exhibits trade-offs among functionalities and facilitates the selection of functional priorities, we employ dynamic normalization into the decomposition-based multiobjective evolutionary algorithm. A robust anisotropic unit-cell database is constructed and used for generating metamaterials with optimized microstructures for multifunctionality. Three complex cases are studied that can simultaneously achieve mechanical cloaking plus stress minimization, thermal cloaking plus stress minimization, and direct current cloaking plus thermal stress minimization. Furthermore, the proposed framework is employed in potential engineering problems, for example, to realize mechanical cloaking by embedding functional components in structural materials, while minimizing the surrounding stress on functional components for mechanical protection when the assembled structures are compressed, thereby decoupling the interaction between functional and structural components. This systematic framework expands the capabilities of metamaterials in complex multiphysics coupling, offering a robust blueprint for future intelligent, structurally and functionally integrated systems.
High-level quantum mechanical (QM) simulations are important for understanding chemical and physical systems, but these methods scale unfavorably with system size, and this limits calculations to model representations of the real systems. To simulate realistic systems, approaches such as hierarchical quantum mechanics in quantum mechanics embedding (QM/QM) can be used to localize the accurate electronic structure methods to a region of chemical interest, with the wider environment represented using a less expensive embedding potential. However, implementing embedding procedures into established QM software packages is often challenging due to their monolithic design and complex software structure. Here, we present a modular and open-source approach that enables QM/QM embedding using established electronic structure codes with reduced development overhead. We demonstrate the new software, EmbASI, using the projection-based embedding theory scheme. Calculations performed for pentanol and organic dimers demonstrate the efficacy of the hierarchical density functional theory (DFT) embedding approach with the charge localization and partitioning procedures implemented in the wrapper environment; when using a truncated local atomic orbital basis, wavefunction-in-DFT embedding calculations demonstrate speed-ups of 32× for a 70 atom system, evaluated with RPA@PBE-in-PBE relative to a full RPA@PBE calculation, while introducing minimal errors to the total energies of the embedding simulations (<1 kJ mol-1). We then demonstrate capabilities for realistic extended zeolite models, achieving similar minimal errors with significant compute speed-up and memory reductions. The results demonstrate the capabilities of EmbASI for immediate applied QM/QM modeling and provide a broader foundation for multiscale QM simulations using established electronic structure infrastructure.
Learned indexes have significantly improved the query performance of in-memory databases by transforming traditional structural traversals into cumulative distribution function model predictions. However, existing dynamic learned indexes rely heavily on the core assumption of ''local linear data distribution.'' Under highly skewed dynamic workloads in real-world scenarios, this linear assumption frequently fails, forcing the system to trigger extremely expensive model retraining and horizontal data shifting. This incurs severe cascading splits and leads to a sharp degradation in dynamic write throughput. To break this architectural bottleneck, this paper proposes BLADE, a bit-level learned index architecture designed for dynamic mixed workloads. BLADE abandons the strong reliance on complex mathematical models and achieves a structural decoupling of routing addressing and underlying data storage at the physical level. In the upper routing module, BLADE employs a pure integer bitwise partitioning mechanism for stateless memory addressing, eliminating the overhead of floating-point model computation. In the underlying leaf nodes, the system inherits the contiguous array layout with reserved gaps to absorb insertions efficiently. To address local congestion caused by data skew, BLADE introduces a vertical splitting strategy, which redistributes data by extending the routing level downwards, effectively avoiding large-scale physical data copying at the bottom layer. Comprehensive experimental evaluations demonstrate that under read-only and read-heavy workloads, BLADE achieves point lookup throughput of 8.2×, 4.1×, 3.2×, and 1.3× that of B+-trees, PGM-index, ART, and ALEX, respectively. Under high-frequency write dynamic mixed workloads, its throughput reaches up to 4.5×, 4.0×, 2.3×, and 1.2× that of PGM-index, B+ Tree, ART, and ALEX. Furthermore, by replacing learned internal-node models with deterministic bitwise routing, BLADE eliminates the iterative model fitting overhead, accelerating the index initialization and bulk-loading speed by 16× to 48.8× compared to the most advanced learned indexes. Under the tested skewed insertion patterns, BLADE's maximum observed RSS is only 1.17× that of its Uniform baseline. Parameter-sensitivity experiments further validate the default 30% target gap and 16~MiB maximum leaf capacity as a balanced configuration for insertion performance and space efficiency.
Online anomalous sub-trajectory detection is a core technology for applications such as real-time traffic monitoring and fraud prevention. Existing methods primarily suffer from three key limitations: 1) weak capability in fine-grained anomaly localization, 2) insufficient constraints regarding road network topology, and 3) difficulties in interpreting the spatiotemporal physical logic of anomalous behaviors. To address these issues, we propose OSTAR, an online generative framework for anomalous sub-trajectory detection driven by topology-aware autoregressive reasoning. Specifically, OSTAR first utilizes a macro-micro-aware perception module to capture local movement dynamics. Building upon this, it embeds road network topology into its autoregressive modeling to capture physical movement constraints. During inference, rather than outputting an uninterpretable anomaly score, it dynamically constructs a physically coherent reference path to serve as a transparent baseline. Finally, by decoupling the detection into independent spatial and temporal dimensions, OSTAR explicitly pinpoints ''spatial detours'' and ''temporal abnormal dwelling.'' This design effectively overcomes the opacity of traditional black-box scores. Extensive experiments on two real-world datasets demonstrate that OSTAR achieves excellent accuracy in fine-grained anomaly localization while maintaining competitive inference efficiency, meeting the real-time requirements of industrial deployment. The source code is publicly available at https://github.com/lzxcompute/OSTAR.
Memristor‐based neuromorphic computing has emerged as a promising hardware paradigm for energy‐efficient, scalable artificial intelligence edge applications. This review surveys recent progress in memristor‐based neuromorphic computing, in‐memory computing architectures, learning algorithms, and hardware implementations for edge intelligence. Volatile and nonvolatile memristive dynamics are discussed as key mechanisms for emulating biological neurons and synapses, providing foundational building blocks for crossbar‐based hardware neural networks. Moving beyond conventional neural‐network acceleration, this review examines the integration of memristive technologies into edge‐specific applications, focusing on emerging in‐sensor computing, optical in‐memory computing, and multimodal sensing frameworks. Furthermore, we outline the deployment of hardware‐compatible, brain‐inspired algorithms—including artificial, spiking, and recurrent neural networks—and evaluate their physical implementation within memristive architectures. The review also highlights the evolution of memristor‐based hardware from single‐device demonstrations to array‐level accelerators and CMOS‐integrated system‐on‐chip platforms, with the associated hierarchical challenges and solutions. By bridging the gap between individual device physics and full system engineering, this review provides a comprehensive roadmap for developing highly adaptive, low‐latency, and autonomous hardware platforms for the next generation of localized edge intelligence.
ESRGAN provides high-quality image super-resolution but requires intensive computation, large memory capacity, and frequent off-chip data access, limiting its deployment on resource-constrained edge devices. This paper proposes an FPGA-based accelerator architecture for ESRGAN. The computational characteristics of the Residual-in-Residual Dense Block (RRDB) are analyzed, and a hardware-software co-design strategy is adopted. Mixed-precision quantization is applied to reduce parameter storage and intermediate feature-map memory while maintaining reconstruction quality. A parallel multiply–accumulate array, pipelined convolution unit, and on-chip buffering architecture are designed to improve data reuse and reduce external memory access. Modular control and multi-stage pipeline scheduling further increase computational parallelism and hardware utilization. The proposed accelerator is implemented and evaluated on an FPGA platform using standard super-resolution datasets. Experimental results demonstrate that it achieves image quality close to the floating-point model while improving inference speed, energy efficiency, and resource utilization compared with general-purpose processor implementations. The architecture provides an effective solution for low-power, real-time deployment of ESRGAN on edge vision systems.
The decode phase of Mixture-of-Experts (MoE) large language models (LLMs) remains primarily memory-bound because both attention and MoE layers exhibit low arithmetic intensity. Despite the massive compute capability of modern GPUs, decoding performance is constrained by memory bandwidth, with high power overheads compromising energy efficiency. Moreover, scaling memory bandwidth in conventional GPU–HBM architectures is constrained by the GPU shoreline limitation. To address these challenges, we propose CASCADE, a memory-centric accelerator that combines low-power, compute-enabled custom HBM (cHBM) chiplets with two-dimensional (2D) parallelism. Through this hardware–software co-design, CASCADE achieves up to 2.88× decoding speedup and up to 4.33× higher energy efficiency than state-of-the-art accelerators.
Neural networks controlling behavior of A-life agents can evolve over multiple generations. While effective in making the agents survive, such networks are generally opaque to a human observer. Furthermore, the neural weights cannot be transferred between agents with different neural architectures. Distilling such neural networks into compact symbolic policies can help humans understand agent behavior in A-life. Additionally, it allows the agents to share their knowledge in a compact fashion, improving population-level learning. Distilling a neural policy into a symbolic policy is typically performed by searching the space of symbolic policies to find one that approximates the neural policy. However, even with a domain-specific language the space of policies can be large, necessitating incomplete stochastic search such as genetic algorithms. Yet the distillation remains computationally expensive. We propose to instead use a transformer pre-trained on readily available random percept-action data. Doing so speeds up distillation substantially. We derive the transformer architecture via a neural architecture search. Empirical results compare different distillation methods and the influence of the symbolic space on A-life population dynamics through self-reflection and teaching and learning.
The explosion in the volume of genomic data generated at a single-molecule resolution, largely due to the development of high-throughput sequencing technologies and state-of-the-art biomedical instrumentation, has created an urgent demand for intelligent, scalable, and interpretable computational models. Classic machine learning approaches frequently suffer from high-dimensionality, noise and convoluted regulation in genomic data. Fortunately, the deep learning (DL) paradigm has come to the rescue as a transformative methodology, and evolutionary algorithms have offered potent tools for model fine-tuning, resulting in the recently-invented evolutionary deep learning (EDL). Nevertheless, despite recent interest, our understanding of how EDL approaches enable the interpretation of genomic data has been sketchy. To that end, this systematic review has aggregated and contributed to the EDL frameworks specific to genomic use cases, created from 2018 through 2025. The method complies with PRISMA recommendations and includes the literature obtained from IEEE Xplore, Scopus, Web of Science, ScienceDirect, and PubMed. The survey covers a range of architectures convolutional neural networks, recurrent neural networks, autoencoders, generative adversarial networks, graph neural networks, and transformers alongside evolutionary strategies such as genetic algorithms, particle swarm optimization, differential evolution, genetic programming, and neural architecture search. The results demonstrate large advances in motif discovery, variant effect prediction, single-cell expression analysis, genomic privacy protection and multi-omics integration. The reviewed evidence suggests that evolutionary optimization can improve selected DL workflows in genomic applications, particularly in architecture search, feature selection, hyperparameter optimization, and evolution-inspired data augmentation. However, the reported benefits vary across tasks, datasets, and evaluation settings. Nevertheless, challenges remain in computational complexity, the absence of meaningful benchmarking datasets, and low biological interpretability. In conclusion, this review highlights that EDL is an interesting and potential new direction for sophisticated genomic analysis, and outlooks for the multi-objective optimization, explainable artificial intelligence, and unified multimeric modeling are presented.
Abstract Autonomous scientific discovery requires the coordinated execution of hypothesis generation, evidence appraisal, formal reasoning, experimental selection and uncertainty-aware revision. Contemporary multi-agent systems have begun to connect literature interpretation with laboratory feedback, while neuro-symbolic methods can constrain machine-learned representations with interpretable logical rules. Quantum information processing introduces additional opportunities for structured search and experimental design, but also substantial constraints from device noise, access cost and limited empirical advantage. This paper develops Neuro-Symbolic Quantum Intelligence (NSQI), a proposed architecture that combines neural hypothesis generators, ontology-linked evidence graphs, a convergent symbolic validation layer, resource-bounded quantum or quantum-inspired search, and Bayesian experimental selection. Its mathematical formulation introduces a constrained candidate-scoring rule, density-operator evolution for quantum modules, risk-adjusted expected information gain and a provenance-weighted evidence update. An explicitly assumed numerical scenario demonstrates how the framework ranks three prospective experiment candidates and how a feasibility or safety gate can change the preferred action; these results are analytical consequences of declared inputs rather than observations from a deployed system. A prospective protocol specifies real-source evidence capture, independent verification, matched-compute baselines, reproducibility controls and human oversight. The contribution is a falsifiable integration model that separates logically valid proposals from empirically supported discoveries, permits classical fallback whenever quantum execution is unsuitable and identifies measurable criteria for assessing scientific usefulness. The study advances a testable research design rather than asserting established quantum advantage or laboratory discovery. Keywords: neuro-symbolic artificial intelligence; autonomous scientific discovery; quantum intelligence; ontology-grounded reasoning; Bayesian experimental design; quantum error correction; multi-agent systems; symbolic regression; epistemic uncertainty; active learning; formal verification; evidence provenance; constrained optimization.
The article presents the results of experimental studies of the process of obtaining organomineral fertilizers using an artificial neural network to control technological parameters. The purpose of the work is to improve the technical means of producing organomineral fertilizers by developing and justifying the parameters of an automation system based on an artificial neural network that ensures the production of a product with specified and stable quality characteristics. The laboratory studies were conducted using a setup that includes a reactor with a paddle mixer and a disintegrator. Lowland peat with a degree of decomposition of at least 25% and brown coal with a humic acid content of up to 40% were used as raw materials. Extraction was carried out using potassium hydroxide. The variables were the alkali application rate (10–50 g/kg of raw material) and the process duration (0.5–1.5 h). The content of humic and fulvic acids was determined using the spectrophotometric method. Based on the obtained data and literature sources, a database was created for training the neural network model. The optimal network architecture was determined using automated search using the AutoKeras library. The developed model was tested on a modernized production line. The research has shown that the most significant factor affecting the yield of target components is the amount of alkali applied. Additionally, the use of a trained artificial neural network has resulted in a 12.85% increase in the total content of humic and fulvic acids in the final product compared to the control. These findings highlight the potential for integrating neural network models into automation systems for the production of organic-mineral fertilizers to stabilize product quality and improve production efficiency.
This research investigates the automation of Recurrent Neural Network (RNN) architecture design through meta-learning. The proposed framework employs a predictive "mother" model that estimates the performance of candidate RNNs ("child" models) using their architectural representation. Each child model is parameterized as a single sparse square matrix, where regional variations in sparsity define its architectural structure. The computational dynamics of these models are defined by iteratively propagating outputs through the matrix for a fixed number of steps, with a portion of the final output serving as the prediction. A dataset of child architectures paired with their final test losses is used to train the mother model, enabling it to use architectural features to predict model performance. Once trained, the mother model enables gradient descent on child model architectures. The resulting process identifies architectures with minimal predicted test loss, thereby reducing the computational cost associated with exhaustive search or manual design. The mother-child optimization framework is applied to a variety of datasets in order to evaluate its effectiveness.
Most neural architecture search and dynamic-network approaches converge to an optimized (static) deployment configuration, whereas machine intelligence systems must operate under varying data distributions, computational requirements, energy constraint and safety requirements. This paper outlines an AI based self-adaptive deep neural learning architecture that completes the cycle of operational monitoring, architecture optimization, lifelong learning and controlled deployment. The framework features an elastic super network, drift and uncertainty tracking, a multi-objective adaptation controller, a hierarchical intervention, replay-enabled knowledge retention and shadow-mode safety validation. The controller does not perform an expensive full search of the global space after each change, but instead it picks out the most viable valid controller action, which is one of five: parameter tuning, subnetwork routing, structured compression, modular expansion, and global reconfiguration. A seeded prototype simulation with two concept-drift events and changing resource budgets is used to assess feasibility. With constraint-aware validation, unsafe adaptations were effectively rejected; the proposed configuration gained 94.6% final accuracy, 18.7-ms P95 latency at 300 requests/s, 0.82 J per 100 inferences and 93.1% knowledge retention. The outcomes demonstrate that closed-loop adaptation can enhance resilience and efficiency than the baseline static DNN, differentiable NAS, and elastic-super networks.
Terahertz (THz) communication is an essential component of sixth-generation (6G) wireless networks; however, its use is constrained by molecular absorption, ultrawideband beam squint, high- $\kappa $ Rician fading, and beam misalignment. The multiplicative fading penalty of passive RIS design can be avoided, and reliable coverage can be extended into the THz band using active, reconfigurable intelligent surfaces (A-RISs) that integrate per-element amplifiers. However, the simultaneous optimization of discrete phase shifts and continuous per-element amplification factors is a nonconvex, high-dimensional problem that is difficult to solve with conventional iterative solvers for the submillisecond coherence times of mobile THz users. A hardware-aware A-RIS-assisted THz communication framework is presented in this article, with an extensive channel model that accounts for frequency-selective molecular absorption, ultrawideband beam squint, high- $\kappa $ Rician fading ( $\kappa \in [{10,25}]$ dB), and stochastic beam misalignment. We present an unsupervised graph neural network-Transformer (GNN-Transformer) architecture to solve the resulting joint optimization problem in real time. Graph convolutional layers exploit local spatial dependencies among RIS elements, while Transformer attention layers capture global channel dependencies across the entire RIS panel. The network is trained end-to-end on channel realizations but, unlike optimal labeling, does not require optimal discrete-phase and continuous-amplification configurations, which are obtained in a single forward pass. The results of the simulations show that the proposed method, which performs an exhaustive search, achieves a higher SNR, meets the power budget without violating it, and reduces inference latency compared to successive convex approximation (SCA), thereby enabling large-scale deployments of THz A-RIS in real time.
Implementing a reliable framework to detect and classify electrical faults in photovoltaic arrays is essential for ensuring operational safety and maximising power generation. Machine learning methods, especially artificial neural networks, have been increasingly adopted in photovoltaic systems due to their ability to identify complex electrical signatures. Despite their proven performance, researchers traditionally rely on manual hyperparameter tuning for optimisation. This process is highly time-consuming and prone to human bias. To overcome this limitation, this study proposes an Automated Machine Learning (AutoML) based Artificial Neural Network (ANN) for the accurate detection and classification of normal operations, open-circuit, short-circuit, and degradation faults at different severities under varying irradiance levels. The framework processes raw current, voltage, and power curve data to eliminate manual feature extraction and employs Bayesian optimisation and neural architecture search to automate hyperparameter tuning and architecture selection. To ensure statistical robustness, the models were evaluated across 10 seeded runs. The results demonstrate that the AutoML approach outperformed the conventionally tuned ANN. The overall accuracy of AutoML-ANN was 93.93%, compared with 92.27% for ANN. The automated framework also reduced training time, converging in 27.65 seconds on average, compared to the 36.89 seconds in ANN. By automatically optimising hyperparameters and network architectures, this framework offers a highly reliable solution for photovoltaic fault diagnosis.
EraGene is a platform that compiles experimentally validated engineering targets and production data from engineered Escherichia coli strains for metabolite overproduction. By integrating published metabolic engineering studies, it includes recurrently modified reactions, enabling identification of successful engineering strategies for biochemically related metabolites. We demonstrate its utility through two applications: identifying conserved engineering targets for pathway reinforcement and reducing the search space for constraint-based metabolic model predictions. Leave-one-out analysis showed that, on average, 76 ± 29% of a metabolite’s engineered reactions were already annotated for other metabolites. This overlap is statistically indistinguishable from a degree-preserving null model (p = 0.135) and is therefore reported as a descriptive measure of the operational redundancy of the curated corpus rather than as evidence of a privileged set of transferable targets. Incorporating reactions with conserved modulation direction into QPAML simulations reduced the average search space from 180.21 to 43.51 candidate reactions per metabolite while recovering 81.88% of experimentally validated modifications in the prediction set, corresponding to a 3.41-fold enrichment that remains significant against a null model preserving reaction popularity (p < 0.001). Independently, the communities detected in the knowledge graph were enriched for KEGG pathways well beyond a size-matched permutation null (20 versus 2.0 ± 1.3 enriched groups; p < 0.001). At publication, EraGene contained 3,833 production measurements from 2,611 strains across 360 studies, comprising 342 native reactions involving 921 genes. The core analytical dataset was restricted specifically to 300 native reactions associated with 88 target metabolites produced in E. coli using D-glucose as the sole carbon source. All analyses are retrospective, and the corpus records only successful, published interventions, so the reported predictive performance should be read as an upper bound pending prospective validation.
Recent advances in AI are shifting the emphasis in software development from producing source code to other stages, such as design and validation. LLMs powering modern coding agents have been trained with code produced by humans and thus have inherited the trait to produce software bugs, a problem that is amplified by the faster development pace. In this paper, we focus on dynamic software debugging at the memory level, which is a relatively unexplored area, as most existing work focuses on static source code analysis. Specifically, we demonstrate how an LLM agent can dynamically inspect and modify a live C/C++ program to pave the way for novel AI-driven debugging solutions such as dynamic observability and live bug patching. These capabilities are implemented as low-level tools that enable the LLM agent to inspect and modify the memory of the running process. To facilitate the agent in using the tools, we provide it with access to the debug information of the compiled software, which maps source code statements to memory addresses, and to a code property graph for static source code navigation. We showcase the system with a prototype that locates, instruments, and patches a simple bug in a running, industry-grade C/C++ software from the telecommunications domain.
Personalized gamification promises to create serious games that better fit different motivational orientations, yet translating player-type models into concrete, playable game artifacts remains labor-intensive. We present HexaQuest, a constraint-based pipeline that operationalizes Marczewski’s Hexad user types as machine-readable generation constraints for adaptive serious-game micro-quests. HexaQuest maps a six-dimensional Hexad profile to weighted mechanics, compiles these mechanics into a typed QuestPlan representation, adds profile-specific choices, state variables, and optional interaction branches, validates and repairs the resulting graph, and exports playable Twine/Twee artifacts. We evaluate HexaQuest as an artifact-generation system, not as a human-subjects intervention. A paired template-only and local-LLM realization setup separates compiler-level personalization from text rewriting. In an automated sweep over 324 artifacts, conditioned generation preserved all structural and objective constraints and achieved higher intended-profile alignment than generic and shuffled mapping controls; a separate deterministic benchmark sustained 12.5–12.9 artifacts/s across balanced batches of 54–324 quests. The results suggest that Hexad-conditioned micro-quest generation is feasible and provides a reproducible foundation for later user studies on adaptive serious games.
Distributed-energy intelligent regulation workloads such as photovoltaic forecasting, multi-node load forecasting, and edge-side control policy inference are increasingly deployed on resource-constrained edge AI chips. These workloads are compute-and-memory-dense: irregular operator graphs, cross-layer tensor dependencies, and limited on-chip buffer capacity jointly create severe off-chip traffic and execution stalls. Existing edge deployment studies often emphasize model compression or accelerator hardware design, while giving insufficient attention to compilation-stage memory-access optimization for workload graphs with non-regular structures. This paper presents a topology-aware memory-access compiler framework for distributed-energy intelligent regulation workloads on edge AI chips. The framework integrates three tightly coupled mechanisms: memory-access compilation optimization, operator-depth-clustering-based layer fusion, and irregular subgraph scheduling under limited cache capacity. A lightweight data-management strategy further decouples memory and compute instructions and dynamically adapts dataflow paths and buffer partitions according to hardware topology. We conduct a compiler-level analytical evaluation on three application-derived regulation workload graphs under a common parameterized edge-AI hardware abstraction. The proposed framework reduces average off-chip traffic from 69.5 MB to 40.4 MB, improves onchip hit rate from 70.0% to 87.8%, cuts stall cycles by 46.0%, and achieves an average execution speedup of 1.42x over a layerwise baseline. These results establish memory-access-aware compilation as a first-class optimization problem for distributed-energy edge intelligence.
Accurate age estimates of salmon produced from scale growth patterns yield demographic information essential for population assessment and fisheries management. Conventionally, salmon ages are estimated from scales by trained personnel where demand often exceeds staffing capacity and financial resources. Convolutional neural network (CNN) models are a promising approach to augment manual age estimation. We compiled an extensive scale image repository from chum salmon (Oncorhynchus keta, Salmonidae) stocks throughout the North Pacific rim, evaluated two image preprocessing algorithms—Segment Anything Model (SAM) and You Only Look Once (YOLO)—to remove or reduce background noise, and trained an ensemble of three CNN models (EfficientNetV2L, NASNetLarge, and ConvNeXtLarge, weighted by their MSE) to infer chum salmon age. Agreement with expert estimates was higher for the SAM preprocessed images (93.2%) than YOLO (91.4%) across all major age classes (ages 3–5). Both preprocessing approaches showed reduced agreement for less common age classes because of inadequate representation in the model training and validation datasets. This approach may be integrated into fisheries management workflows to accelerate the age estimation process, complement manual reading, and bolster quality control protocols.
Deploying deep learning on microcontrollers (MCUs) requires fitting both the model weights and runtime activations within tens to hundreds of kilobytes of SRAM and Flash. Existing approaches either use an interpreter with a pre-written kernel library or combine a Python code generator with hand-written C kernels. This incurs dispatch overhead and requires framework dependencies. None produces self-contained, dependency-free C code specialized to a given model. In this paper, we present te-comp, the first complete MLIR-based compiler that takes a standard model in .tflite or .safetensors format and emits all C source files needed for deployment: the inference engine (kernels, weight arrays, memory layout) and the RTOS integration layer with zero runtime dependencies. On 10 MCUNet CNN benchmarks, te-comp outperforms MCUNet TinyEngine by using 38% less SRAM (up to 5.9$\times$) and 15% less Flash, while generating complete C code rather than linking a pre-written kernel library. To demonstrate, we use te-comp to compile an 8.3M-parameter GPT-Neo transformer (TinyStories-v2) and deploy it on a single-core Cortex-M7 at 160MHz. The full firmware uses 249KB of 512KB SRAM and 3.62MB of 4MB Flash and achieves approximately 1 token/s interactive text generation via Q4_0 weight quantization, a stateful int4 KV cache, and vocabulary pruning -- a capability no existing MCU inference framework supports.
Political conflicts among parties in economic and social domains have increasingly been investigated using data-driven deep learning (DL) approaches. However, existing data-driven DL models often fail to identify discriminative patterns within political narratives. These challenges are primarily attributed to the scarcity of annotated political datasets, the linguistic diversity of regional languages, and the inherent ambiguity of political discourse. To address these limitations, an attention-based DL framework is proposed for contextual classification in Indonesian political narratives using Bidirectional Encoder Representations from Transformers (BERT) variants. Political data were systematically collected and annotated through a human-in-the-loop process, using expert-defined labels for political conflict and mediation strategies. The proposed framework consists of two stages: an extractive summarization module followed by a contextual classification module. Multi-task DL models with customized shared-backbone architectures based on BERT, DistilBERT, and RoBERTa were developed and trained on the compiled dataset. Performance was evaluated using five-fold cross-validation and benchmarked against conventional machine learning and state-of-the-art DL approaches. Experimental results demonstrated that the customized DistilBERT model achieved the best performance on the Indonesian political corpus, attaining F1 scores of 0.972 for political conflict classification and 0.909 for mediation strategy classification. These findings demonstrate the effectiveness of the proposed framework and highlight its potential as a human-centric analytical tool for supporting political analysts, policymakers, and researchers in interpreting complex political discourse.
Predictive maintenance (PdM) is a key enabler of reliability and cost reduction in Industrial Internet of Things (IIoT) systems. However, conventional centralized learning approaches require raw data aggregation, raising serious concerns regarding data privacy, communication overhead, and single points of failure. Federated Learning (FL) addresses part of this challenge by enabling distributed model training, yet centralized FL architectures still rely on a central aggregator. This paper proposes a decentralized federated learning (DFL) framework based on a ring topology for multi-class predictive maintenance in IIoT environments. The proposed approach enables collaborative model training among distributed edge nodes without relying on a central server, thereby enhancing system robustness, scalability, and data privacy. A lightweight deep neural network (DNN) is employed to classify five machine operating states using multi-sensor data, including vibration, temperature, and current measurements. Extensive experiments are conducted using a real-world sensor dataset collected from three-phase AC motors under multiple fault conditions. The performance of the proposed DFL approach is systematically compared against a non-federated baseline and a centralized FL (CFL) scheme. Experimental results demonstrate that the decentralized framework achieves clas-sification performance comparable to centralized FL, with an accuracy of up to 99.83%, while eliminating the dependency on a central coordinator. The findings confirm the feasibility of decentralized federated learning for predictive maintenance applica-tions, particularly in distributed industrial settings where data privacy, fault tolerance, and communication efficiency are critical.
The physical random access channel (PRACH) enables initial access between user equipment and the base station, making reliable preamble detection essential for cellular networks. Growing device density in 5G and beyond-5G networks, massive machine-type communication, and the shift towards machine-learning-driven radio interfaces challenge conventional analytical detectors, particularly in complex radio environments and at low signal-to-noise ratios. This article formulates preamble detection as a 64-class classification problem and introduces two model-agnostic contributions: a robust adaptive threshold implementing the constant false alarm rate (CFAR) principle in the classifier decision-score domain (MAD-CFAR), and a hybrid receiver combining a learned detector with a conventional power-delay-profile CFAR detector through a logical OR rule. The learned branch is evaluated using two models trained on the same GPU-generated dataset: a support vector machine (SVM) with principal component analysis and a lightweight one-dimensional convolutional neural network (CNN). On a held-out test set, the SVM achieves 98.58% accuracy (macro-AUC 0.9991), while the CNN reaches 99.29% accuracy (macro-AUC 1.0000) at lower inference cost and memory footprint. MAD-CFAR reduces the learned branch's false alarm rate to near zero on the noise calibration campaign. The OR hybrid retains the analytical detector's low, standard-compliant false alarm rate while extending detection capability. Across six propagation channels and 258 conformance-style tests, the SVM-based hybrid improves low-SNR sensitivity by several decibels without increasing missed detections relative to the analytical detector. Under the same false alarm constraint, the CNN is competitive in some channels but currently trails the SVM in high-delay-spread multipath conditions.
This paper presents an overview of \emph{SciClaimEval}, a pilot task at NTCIR-19 on verifying scientific claims against the tables and figures of the papers in which they appear. Our dataset pairs authentic claims from 180 papers in machine learning, natural language processing, and medicine with contrastive versions of their tables and figures, one supporting the claim and one minimally modified so that it does not. The task comprises two subtasks: deciding whether a piece of evidence supports a claim, and selecting which of two evidence versions does. Twenty-two teams registered, the highest number among NTCIR-19 tasks, and 14 submitted 177 runs. Most systems built on multimodal large language models, relying on prompting and multi-stage inference rather than task-specific fine-tuning. On the first subtask, the best systems reached 98.1\% pair accuracy, but all six top-ranked systems relied on the contrastive pairs; the best system that judged each claim-evidence pair on its own scored 24.3 points lower. Open-weight models, embedded in carefully designed pipelines, matched proprietary ones. Yet verifying a single claim against its evidence, the setting that matters in practice, remains an open problem.
The convergence of robotics and the Internet of Things (IoT) has transformed multiple sectors through autonomous, interconnected systems. Despite significant scholarly output, the knowledge structure of Robotics-IoT research remains fragmented. This study provides an integrated scientometric and bibliometric mapping of global Robotics–IoT/IoRT research, combining performance analysis, collaboration-network mapping, thematic mapping, and LDA-based topic modelling to reveal the field’s intellectual structure, collaboration architecture, dominant clusters, emerging themes, and future research frontiers. A scientometric analysis was conducted on 21,733 unique documents retrieved from Scopus and Web of Science after screening and deduplication of 23,377 initial records, covering 1993–2025 across 2,284 sources and over 44,000 indexed author-name records/clusters, cleaned via the PRISMA protocol. Biblioshiny and VOSviewer were employed for publication trends, citation structures, and collaborative networks, while LDA was applied for topic modelling. Robotics-IoT research shows rapid growth post-2016. Using all-author affiliation-country document-level full counting, India, China, and the USA lead publication output. Among the ten most productive countries, Australia shows the highest average citations per document, followed by Canada, the USA, and the United Kingdom. Collaboration networks reflect extensive global partnerships, with emerging economies demonstrating substantial international participation. Thematic analysis identifies major research streams related to industrial digitalization, IoT–robotic systems, energy-efficient networking, machine-learning-based analytics, robotics and sensing, cybersecurity, and edge/cloud computing, alongside specialized applications in agriculture and healthcare. Observed distributions show varying degrees of descriptive alignment with classical scientometric laws (Lotka’s, Bradford’s, Price’s, Zipf’s, Matthew Effect, and Pareto Principle). The originality of this study lies in the domain-specific application and empirical synthesis of these approaches to the global Robotics–IoT/IoRT research landscape. Specifically, the study provides a longitudinal analysis covering 1993–2025, examines the intellectual, social, publication and thematic structures of the Robotics–IoT/IoRT domain, evaluates a comparatively broad set of classical scientometric regularities, and complements keyword-based analysis with abstract-based LDA topic modelling. The study consequently contributes new empirical evidence concerning the evolution, concentration, collaboration, intellectual structure and thematic development of Robotics–IoT/IoRT research.
The growing complexity of Wireless Sensor Networks (WSNs) within Internet of Things (IoT) enabled applications presents significant challenges related to fault detection, energy efficiency, and network resilience. Node failures can cause severe disruptions, leading to increased energy consumption, data loss, and reduced network lifespan. Conventional fault detection methods lack flexibility when dealing with node relocations, while machine learning-based techniques often overlook energy efficiency considerations. This research proposes the Adaptive Federated Learning for Fault Detection and Node Reuse (AFL-FDNR) framework, which combines federated learning with Deep Q-Network (DQN) to provide an adaptive, energy-efficient solution for fault detection and node recovery. Federated learning ensures scalable, privacy-preserving model training across sensor nodes, while DQN optimizes node recovery decisions based on real-time network conditions. Simulation results demonstrate that AFL-FDNR achieves 96.01% fault detection accuracy, reduces energy consumption to 1910 J, and extends network lifetime to 11,452 hours. These metrics demonstrate the framework’s ability to enhance the reliability, efficiency, and sustainability of large-scale WSNs, making it a robust solution for modern IoT applications.
To address challenges in Industrial IoT (IIoT) monitoring, this paper proposes a machine learning framework integrating sensor data classification and optimal sampling time prediction. First, a Linear Mixed Model (LMM) quantifies the correlation between signal features and device parameters, confirming significant positive correlations (p<0.001). Subsequently, an optimized strategy combining K-means clustering (K=5) and spline interpolation minimizes risk functions, establishing optimal sampling boundaries at 27.74 and 34.6 for environmental factors. For anomaly detection, a three-layer ensemble model identifies Decision Trees as the optimal classifier, achieving the lowest false negative rate (14.4%) for safety-critical applications. Monte Carlo simulations validate the model's robustness against ±5% measurement uncertainties. Experimental results demonstrate that the proposed framework achieves 83.1% classification accuracy while ensuring high reliability. This study provides a practical solution for balancing data quality and energy efficiency in large-scale industrial sensor networks
To address the insufficient classification accuracy and generalization ability of single machine learning models under extreme computational constraints in network intrusion detection, this paper proposes a lightweight network traffic classification model that combines weighted feature sampling with two-level heterogeneous stacking. The first-level base learner is an improved random forest with mutual-information-weighted feature sampling. Under controlled tree size and depth, it produces class-probability vectors via five-fold out-of-fold (OOF) prediction. The second-level meta-learner is LightGBM, which receives progressively enhanced meta-features composed of probability features, Top-10 key raw features, and confidence features. Experiments show that the enhanced stacking model under a constrained configuration achieves an F1-score of 0.9888 and an accuracy of 0.9897; the five-fold cross-validation F1 mean is 0.9889 with a standard deviation of only 0.0009. All metrics are substantially better than those of conventional single models under the same resource budget. The architecture achieves high classification performance and strong generalization at extremely low model complexity, and provides a feasible methodological reference for lightweight intrusion detection on resource constrained devices such as IoT gateways and embedded terminals.
The Internet of Things (IoT) connectssensors, smart appliances, mobile devices and industrial equipment so that they can exchange informationandrespond to events. This connectivity is useful, but it also creates security risks when devices have limitedcomputing resources and communicate through different protocols. A Man-in-the-Middle (MitM) attackisonesuch risk: an attacker places themselves between communicating devices and may observe oraltertheexchanged data. This paper presents a machine-learning- based workflow for identifying traffic associatedwithMitM attacks in IoT networks. The proposed workflow includes traffic preparation, preprocessing,featureselection, model training, classification and evaluation. Support Vector Machine (SVM), RandomForestandLong Short-Term Memory (LSTM) are considered because they offer different ways to classifynetworktraffic. The supplied synopsis reports results from the referenced 2025 study in which Random Forestachieved94% accuracy, LSTM 92%, and SVM 85.7% under that study’s experimental conditions. Thesefiguresare reported as prior-study results, not as results independently produced by this paper. Thediscussionalso identifies practical issues such as false alarms, limited device resources, dataset representativenessandthe need for testing in live IoT environments
Computer-aided traffic analysis is fundamental to defending enterprise networks against sophisticated ransomware attacks. This paper presents a novel AI-driven architecture combining dynamic graph neural networks with multi-head temporal convolution modules to identify and contain ransomware propagation in real time. Even in the presence of encryption or obfuscation, this approach can create adaptive propagation graphs to record temporal changes in spatial interactions and network flows between hosts. This enables precise modeling of ransomware lateral movement. A composite dataset, integrating enterprise-grade traffic and simulated attack traces, is used to evaluate the framework. The results show that the method has significantly improved in detection accuracy, recall rate and propagation containment rate compared with the traditional feature-based system and traditional machine learning classifier. In addition, the method guaranties low inference latency, making it suitable for enterprise deployment. To enhance the robustness of detection and the generalization ability in various attack scenarios, the architecture adopts advanced feature normalization, joint entropydriven edge modeling and nonlinear feature fusion. Comprehensive evaluation shows that the framework can adapt to complex and changeable network environment and effectively distinguish between benign and malicious propagation. These results demonstrate the importance of AI-enhanced traffic analysis in enterprise ransomware defense. It also shows that there are more effective methods than existing baselines.
Feed-forward 3D models use Transformers to predict camera poses, depth, and structure in one pass, avoiding iterative multi-view optimization. Long sequences and deep attention incur high compute and memory costs, motivating Post-Training Quantization (PTQ) for deployment. Existing PTQ methods mainly suppress channel-wise outliers by smoothing or orthogonal rotation. Geometry Transformers introduce a different challenge: heterogeneous token types share one activation quantization grid. A few special tokens may exert high leverage over global outputs and be updated by only a subset of blocks. Token-blind transforms, which apply the same transformation to every token, reduce overall quantization error but cannot remove differentiated degradation across token types. We propose QMapAnything, a structure-aware PTQ framework. We cast floating-point-preserving rotation and smoothing as a token-blind reparameterization group, and prove that they yield common-mode gains while leaving relative token SNR unchanged. We introduce a statistic $\Gamma$ that combines a special token's quantization disadvantage and output leverage to predict whether differentiated bias correction is needed beyond rotation. Experiments on geometric benchmarks show that W4A8 closely matches full-precision performance. Under W8A8, peak memory falls by $2.2$--$3.1\times$ versus FP32, with up to $3.1\times$ speedup on real hardware. These results provide a principled and efficient approach to deploying feed-forward 3D reconstruction models in resource-constrained edges.
A child who learns to add small numbers adds numbers of any length. LSTMs and Transformers trained on short sequences fail on longer ones. We trace the difference to what a learner reads when it decides. A position, a count or a time keeps taking new values as the input grows: it is a coordinate, and a learner that reads one meets in every long input values its experience never covered. Whether two things are the same has the same two answers at every size: it is a coincidence, and a learner that reads only coincidences finds in a long input nothing that short inputs lacked, so finite experience fixes its rule for every size. The memory carries the size of the input and the rule stays finite, and this determines the memory a generalizing brain can have: a memory that grows while what the learner reads of it stays finite has to look the same from every place, so it is made of cycles, and the past is reached by asking what happened when the cycles last stood where they stand now. The cycles are the grid modules, the cells that detect their return the place cells, the records the hippocampal index; a prefrontal controller moves the cycles and never sees them. Reading only coincidences costs no power: what finite experience fixes in such a learner is what Turing machines compute in step with their input, which is why a circuit that evolved to find the way home can learn arithmetic and grammar. In a network of neurons this machine of cycles is the dynamics seen through its symmetry, and symbols are its invariants. Trained on strings up to 40 symbols, the machine is at least 99% correct, on average over lengths 41 to 500, on nine of fifteen benchmark tasks, five of them tasks that no LSTM or Transformer solves; given new cycles after training it reverses strings of 15,015 symbols without error and fails beyond that on the curve the theory predicts; and its controllers run unchanged on neurons.
Optoelectronic synapses (OESs) with linear photocurrent generation and prolonged decay are critical for energy-efficient neuromorphic vision and in-memory computing. In solution-processed organic/inorganic hybrid heterostructures (HHs), however, the role of exciton dynamics in the organic layer in governing photocarrier relaxation behavior remains poorly understood. Here, we demonstrate that exciton delocalization in poly(3-hexylthiophene) [P3HT] can be systematically tuned via solvent-controlled molecular ordering in P3HT/Bi2O2Se HHs to tailor synaptic performance. By varying the casting solvent, we modulate P3HT crystallinity, conjugation length, and energetic disorder, which directly regulates exciton delocalization and interfacial charge separation, while Bi2O2Se provides a high-mobility transport channel. Chloroform-processed P3HT films with the highest aggregate fraction, longest conjugation length, and lowest disorder exhibit sublinear photocurrent generation and extended photomemory, ideal for temporal integration. In contrast, chlorobenzene-processed films show faster, more nonlinear responses with rapid decay due to suppressed exciton delocalization. The chloroform-processed OES operates at an ultralow voltage of 100 mV and emulates essential synaptic functions under optical stimuli. Leveraging experimentally extracted long-term potentiation characteristics, artificial neural networks achieve 92.0% accuracy on MNIST and 89.6% on ORL face recognition. A hardware-inspired U-Net further attains 85.5% validation accuracy on Cityscapes semantic segmentation. This work establishes solvent engineering as a general strategy to tailor exciton dynamics in organic/inorganic HHs, enabling high-performance OESs for energy-efficient neuromorphic vision and computing.
Server main memory is increasingly heterogeneous, incorporating technologies such as zswap, CXL memory expanders, and emerging phase-change memories (PCM). To decide where to place pages, operating systems have to choose between efficient but coarse-grained access bits, out-of-band profiling, or costly runtime instrumentation. This paper proposes memory fit, a hardware abstraction that provides precise and accurate page placement information to an operating system. Memory fit measures how well a page"fits"each type of memory in the system. Memory fit separates the mechanism of measuring affinity between pages and memories from the OS policy of choosing which pages to move when. The paper describes soft trace matching (STM), a computationally and space efficient algorithm to compute memory fit. The key technical insight behind STM is that it can estimate memory fit by comparing accesses to a page with a learned model of"good"traces for a memory. We have implemented memory fit and STM in a BOOMv3 RISC-V processor running Linux 6.6. On extended GUPS, memory fit placements recover 2.2-3.8 times as much of the performance gap to optimal placement as Linux TPP. For Redis with equal DRAM and PCM capacities, placements selected by memory fit recover 80% of the performance gap, versus 19% for TPP, in a placement replay. Memory fit requires only 64 bits of state per 4 KiB page and can be updated on every main-memory access with only a 1.2% overhead of OS kernel notification processing.
Multichannel room impulse responses (RIRs) in acoustics are often long and therefore the sample count becomes large if seen as finite impulse response (FIR) filters, making storage and real-time convolution costly. Pole-zero (PZ) models are an established way to represent RIRs compactly as infinite impulse response (IIR) transfer functions. We show that a PZ model implies a low-rank canonical polyadic decomposition (CPD) of tensorized RIRs: the early reflections form a sparse rank-1 sum, while the late reverberation admits a CPD with factors determined by the poles. This motivates higher-order-tensor-based compression of RIRs via the block term decomposition (BTD), a generalization of the CPD. We propose two multichannel compression schemes: (i) a stacked multichannel block term decomposition (SMCBTD) applied to a (D+1)-mode tensor of RIRs, and (ii) a joint multichannel block term decomposition (JMCBTD) that shares subsets of factors across channels discovered by spectral clustering on factor-subspace affinity graphs. Both are optimized by an alternating higher-order orthogonal iteration (HOOI) procedure. On simulated rooms as well as measured datasets, the proposed methods achieve consistently lower reconstruction error at equal compression ratios compared to separate single-channel decompositions and a generalized low-rank approximation of matrices (GLRAM) baseline. The largest gains occur with scalar cores (i.e., CPD), supporting the PZ-to-CPD modelling link. We also find D = 3 to be a practical sweet spot, offering most of the achievable gain at moderate compute/memory cost. The compressed models enable efficient downstream multichannel convolution.
Increased computing power and more efficient codes have allowed cosmological simulations of galaxy formation to grow massively in scale: the analysis of their outputs has consequently become a big data computational challenge. We present Octavius, a high-performance, hybrid-parallel simulation analysis toolkit. Octavius ingests a simulation snapshot with associated halo catalogue, locates galaxies with a 6D friends-of-friends algorithm, then produces an HDF5 catalogue of properties for haloes, subhaloes, and galaxies. The analysis is fast and memory efficient: we compare the performance against the widely used predecessor toolkit, Caesar, and show Octavius delivers order-of-magnitude improvements in speed with a factor of>5 reduction in peak memory. We demonstrate near-linear strong scaling of compute and linear peak memory scaling with the number of halo-assigned particles. Compute times scale with the superlinear computational load of larger haloes; we note the pipeline is presently I/O-limited. The code is written in Python and uses a declarative, plugin-based architecture so as to be easily extensible. Octavius can currently analyse SIMBA, SWIFT-KIARA, SWIFT-EAGLE, SWIFT-COLIBRE, and TNG snapshots; AHF, HBT-HERONS, and SUBFIND halo catalogues are supported. Properties computed include masses, galaxy photometry with dust attenuation, and kinematic, radial, morphological, and virial quantities. Output catalogues are easily accessed through a provided user API. The code is open sourced at https://github.com/jp-duminy/octavius
Resource management for distributed edge networks hosting different services has been handled efficiently using Reinforcement learning (RL). Reward generalization in complex workflows is still an open issue within the RL realm. Therefore, this work proposes RAGAM-AI, a hybrid framework that combines the capabilities of RL algorithms and Generative-AI (Gen-AI) for powering an Agentic AI workflow that optimizes resource allocation at the Multi-access Edge Computing (MEC) units. This framework further resolves the reward generalization issue by utilizing Large Language Model (LLM) for feedback. A Multi- Agent Soft-Actor Critic (MASAC) using Math{\Sigma}tral as a reward oracle is implemented for managing resource allocations at the MEC. RAGAM-AI interprets system state information and evaluates multi-objective trade-offs guiding the resource orchestration loop over distributed edge server agents. The complexity of the framework is further brought down by incorporating surrogate reward model. The overall performance is evaluated against the single agent SAC, REINFORCE and random policies for benchmarking. On using RAGAM-AI, lower allocation deviations for CPU, memory and bandwidth resources are reported while minimizing the latency and data rate costs. A scalable AI as a feedback (AIAF) solution optimizing network resources with significant performance improvement is thus presented.
Reinforcement learning post-training improves the capabilities of large language models, but generating training responses is a major computational bottleneck. In mixture-of-experts models, workers that finish their assigned responses may remain occupied because their experts are still needed by workers processing longer responses. Consequently, approximately one-third of occupied worker time in the baseline occurs after local response generation finishes. Here we present FlexFloor, a system that enables earlier resource release during synchronous response generation while preserving the expert computation required by ongoing responses. FlexFloor jointly adapts response scheduling and the active group of workers. It uses historical response lengths and memory constraints to guide batching and prompt placement, then progressively releases workers while maintaining expert availability and correct communication. Compared with the verl baseline, FlexFloor generates approximately 65.40% more response tokens per unit of occupied computing time and increases response generation throughput by 52.74%, with little observed variation in rollout reward. By making reinforcement learning post-training more efficient, FlexFloor supports faster development and iterative improvement of large language models and AI agents.
Hybrid language models interleave a few softmax-attention layers with many linear-attention or state-space layers, so every layer keeps a memory of earlier tokens: a key-value cache or a fixed-size recurrent state. Serving them spends more compute and high-bandwidth memory on these memories than the output requires. The prefill runs every layer for every prompt token, and every decode step loads and stores whole recurrent states, which on Qwen3.8-Flash-Next take as many bytes per request as the attention cache of thousands of tokens. We find two properties of trained hybrids that locate this excess. First, writes are easier than reads: writing a deep layer's memory of a token needs little of the deep computation that predicting the next token needs. Trained copies of the deep layers'write paths, which we call emitters, therefore write a prompt's deep memories from the residual stream of a shallower anchor layer, and all prompt tokens but the last run as few as 63% of the layers. Second, recurrent states have a state sink: most of their energy lies in a rank-one term along a fixed direction of each head, and the rest is a low-rank working memory. SEED builds on both: it makes the prefill shallow in depth and the decode state shallow in representation. Each recurrent state is stored as its sink plus low-rank factors, which the Gated DeltaNet, Kimi Delta Attention and Mamba-2 updates extend exactly; we prove that the open-loop error of the periodic compressions back to low rank is bounded independently of their number. SEED keeps the original weights frozen, compresses no attention cache, and runs all layers for every generated token. On six checkpoints of five production hybrids of 30B to over 300B parameters, it is nearly lossless on eleven official exams. Served with SGLang on NVIDIA GB300 GPUs on an agentic-coding trace, it raises the per-GPU throughput of Qwen3.8-Flash-Next 1.2-1.4x.
Every reconfigurable microwave device places a tuning element - a varactor, a diode, transistor, microelectromechanical or phase-change switch, a ferroelectric or liquid-crystal load - inside a fixed passive network. The network can make a small change of the element produce a large change of the response, and design practice often credits it with the device's tuning capability. This paper shows that a fixed, linear, passive network cannot add to that capability. Each tuning element carries a tuning budget, a dimensionless length computed from its own impedance trajectory, and a fixed passive network can transmit that budget or lose part of it, never enlarge it. The unit of the budget is the hyperbolic distance of the Smith chart: the distance from a match is the logarithm of the voltage standing-wave ratio, and the distance between two states sets their mutual mismatch and is the logarithm of their commutation quality factor. From one inequality we derive closed-form, ready-to-use limits for analog and digital phase shifters, true-time-delay lines, switches, attenuators, vector modulators, impedance tuners, tunable filters, reconfigurable antennas, metasurface cells and multiport networks; the two-state and attenuation limits are attained by explicit circuits, and the phase limit by a constant-loss trajectory. We introduce the $\Lambda$-chart, a Smith-chart companion on which budgets and requirements are read graphically, and apply it to four published microwave case studies.