oryx
Oryx (SRS Stack) is an AI-driven, all-in-one, out-of-the-box, and open-source video solution for creating online video services, including live streaming and WebRTC, in the cloud or through self-hosting.
9 repos · 19 papers · 48 scholar
Showing 76 of 76 items
Oryx (SRS Stack) is an AI-driven, all-in-one, out-of-the-box, and open-source video solution for creating online video services, including live streaming and WebRTC, in the cloud or through self-hosting.
Video codec based point cloud compression (V-PCC) Test Model (encoder and decoder)
Deterministic Visual Attention & Saliency Prediction Studio. In-browser neuromorphic computer vision, convex maximum-likelihood fusion, Bayesian AOI credible intervals, and simulated gaze scanpaths. 100% zero-cloud.
Neuromorphic Computing D7064E 27024 HT2026
Zilong Liang | Neuromorphic systems researcher at UNSW Sydney
Simulation study of a 6-input, 5-class time-race classifier using only p-type organic thin-film transistors (C10-DNTT) on flexible foil, with fitted device models, host protocol, and fully reproducible ngspice campaigns
Network-on-Chip (NoC) and Advanced Interconnects: From Silicon Physics to Chiplets and Neuromorphic Computing
We present a toolchain for development of ML models on AWS EC2 F2 cloud FPGA infrastructure from high-level language, implementing a complete Vitis HLS to HDK bridge to AFI workflow that has been absent from the academic literature despite eight years of AWS FPGA availability. We introduce a four-level methodology (FP32 PyTorch, bit-accurate quantized Python, HLS C-simulation, physical FPGA) to aid in localizing errors, including FPGA-specific circuit errors. Together these tools allow coding agents to effectively implement FPGA designs from user directions, testing and debugging as needed. We document the deployment workflow, timing closure strategies, and practical HLS engineering decisions that proved crucial. We validate using TinyBERT fine-tuned for address Named Entity Recognition, implementing the embedding layer and the first of four encoder layers on the FPGA hardware. Our implementation streams weights from host DRAM, yielding poor performance compared to CPU/GPU implementations; we discuss forward paths.
We show that a hardwired distributed processing unit (HDPU) can take a neural decision at the sensor and so remove the inference from the host. Built for zebra-crossing detection on a low-cost Artix-7 FPGA, the unit decides on every camera frame in a fixed 2,269,477 clock cycles (34.04 ms depends upon Fpga clk ), inside the 37.49-ms frame period, and on the board it returns the same scores and decisions as its bit-exact software model on all 639 test images. The whole detection path, from camera pixels to decision, takes 1,814 LUTs, with no processor, no operating system and no full-resolution frame store. When the unit answers instead of reporting "not sure", 95.7% of its answers on street photographs are correct, and a hardware contrast stage lifts accuracy on camera-like dashcam frames from 47.7% to 78.2%. The host is spared all of the inference: 8.13 ms of a core in every frame period, where a full-precision ResNet-18 would need more than a frame period even on 16 threads. It also receives 16 bytes per frame instead of 614,400, 38,400 times less. The unit is an 80x60-pixel pipeline feeding a four-layer binarized CNN with 93,408 one-bit weights, verified bit-exactly from a Python model through RTL simulation to the board. Earlier HDPUs did pre-processing (edges, corners, steering); this one shows that the decision itself can move to the sensor.
Gated-delta linear attention layers such as Gated DeltaNet (GDN), Kimi Delta Attention (KDA), and GDN-2 are tuned for GPUs, where small tensor-core tiles and fast vector registers make per-channel decay rescaling and sequential in-block solves cheap. On Google TPUs, which concentrate compute in wide $256 \times 256$ systolic matrix units backed by a slower vector unit, porting these layers directly leaves half of each matrix tile empty and bottlenecks on vector operations. We present \textbf{Systolic Linear Attention (SLA)}, which reshapes gated delta attention around the matrix unit by widening keys to $256$ with grouped value heads, grouping decay gates into four $64$-channel tiles, and computing the chunk solve entirely with block matrix products in a single fused kernel. On TPU v7x, a full \texttt{SLA-R4} layer runs $1.6\times$--$2.5\times$ faster than the open-source Tokamax KDA kernel ($1.4\times$--$2.4\times$ with exact chunk inversion), driven primarily by head sharing and kernel fusion. Across a $2{,}048$-chip TPU v7x pre-training study of sparse Mixture-of-Experts (MoE) models from $6.5\text{B}$ to $63.1\text{B}$ parameters, SLA trains $11\%$--$24\%$ faster than KDA at similar perplexity, and reaches slightly lower perplexity than KDA in smaller matched-seed runs.
Processing-in-Memory places small programmable cores on the memory device, where they execute beside the DRAM banks and beyond the control-flow defenses that protect host programs. The security of this in-memory code is an attack surface that remains largely unstudied. We present PIMROP, to our knowledge the first demonstrated code-reuse attack on a commercial Processing in-Memory ISA. We show that return-oriented programming is feasible on a real, commercially shipped PIM instruction set, the UPMEM DPU, whose DRAM Processing Units run C compiled by a dedicated LLVM toolchain. We build Galileo-for-DPU, a gadget scanner adapted to the DPU's fixed-width, 8-byte-aligned RISC encoding, and run it entirely on the UPMEM functional simulator in a reproducible Docker and VM environment. We find the runtime library that every kernel links already yields 517 non-trivial gadgets covering all five canonical ROP primitive classes, and all 13 PrIM benchmarks we scanned carry gadgets as well. We then demonstrate two verified end-to-end attacks, each reproducible on 10 of 10 runs. In the first, a stack buffer overflow in a DPU kernel overwrites the saved return address and diverts control into a routine that copies a secret into host-readable memory. In the second, we leak the same secret using a chain of reused gadgets alone, without diverting control into any such routine. We further show that two standard defenses--stack canaries and _FORTIFY_SOURCE-style bounds-check fortification--fail on the DPU in ways specific to the platform (misplaced canary, silently inactive wrapper), and we verify a bounds-checked DMA that stops the attack on 10 of 10 runs with no effect on benign requests. To enable further research on PIM security, we open-source Galileo-for-DPU and our testing harness at https://github.com/wtegge2/ROP-in-PIM.
Diffusion models have become the dominant paradigm for high-quality image generation, but their iterative inference across tens of timesteps incurs substantial computational overhead. Prior accelerators exploit temporal similarity in diffusion models via differential quantization or output sparsity, yet suffer from increased memory footprint, design complexity, and image degradation. In this work, we present Phantom, an algorithm-architecture co-designed accelerator built on temporal caching. We view temporal similarity as differential sparsity: between adjacent timesteps, only a small fraction of tokens (outliers) change significantly. Phantom recomputes only these outlier tokens and reuses the rest from a temporal cache. Phantom runs on standard systolic arrays with lightweight additions: a Phantom Engine at the global buffer that detects outliers and updates the cache as activations stream through, and a condensed pipeline that maps the resulting sparse updates onto dense systolic computation for FC, convolution, and attention layers. Evaluations on modern diffusion models show that Phantom achieves an average 4.0x speedup and 5.5x energy saving over GPUs and prior accelerators while preserving generation quality.
Automatic modulation classification (AMC) underpins spectrum sensing and non-cooperative signal identification in cognitive radio, yet the two standard CPU-feasible approaches to it, a compact convolutional neural network (CNN) on raw in-phase and quadrature (I/Q) samples and a classical higher-order-cumulant feature classifier, are rarely compared against each other conditioned on signal-to-noise ratio (SNR). We show that they are, in fact, complementary: the CNN is markedly more robust at low SNR, where hand-crafted statistics collapse toward chance, while the cumulant classifier is more accurate at high SNR, where higher-order statistics cleanly separate dense constellations. We exploit this complementarity with FuseAMC, a lightweight decision-fusion model that trains a logistic-regression meta-combiner over the concatenated class-posteriors of an approximately 198k-parameter residual CNN and a gradient-boosted-tree cumulant classifier, on a validation split disjoint from test. On RadioML 2016.10a (5 seeds, 44k-frame held-out test set), FuseAMC reaches 55.8% overall accuracy, improving on the best cumulant-feature baseline by 5.1 points and on the CNN alone by 3.2 points (paired-bootstrap 95% confidence-interval lower bounds of 5.0 and 3.1 points, both positive), while matching the best classical baseline in its own strong high-SNR regime and the CNN in its strong low-SNR regime. With about 0.2 million parameters, a model footprint of a few megabytes, and roughly 3 MFLOPs per frame, the whole pipeline trains and runs on one CPU core, targeting general-purpose and embedded processors without hardware acceleration.
Hierarchical text embedding models offer variable-length representations, while quantization reduces numerical precision; both aim to improve computational efficiency. This work is an empirical benchmark study of how these two orthogonal, already-established techniques interact. We evaluate 38 tasks from the Massive Text Embedding Benchmark (MTEB) using four transformer-based embedding models, covering 1D and 2D hierarchical representations combined with 32-bit, 8-bit, and binary quantization, for 168 configurations per task. We report each efficiency dimension separately: theoretical per-vector storage and end-to-end evaluation time. Under a Friedman test with Nemenyi post-hoc analysis, int8 quantization is statistically indistinguishable from float32 for pair classification and re-ranking and costs at most 1.5 percentage points overall, whereas binary quantization is significantly worse across every task category. Halving the embedding dimension once is not statistically detectable above $m=16$; two successive halvings always are. Combining an 8x dimensionality reduction with int8 yields a 32x reduction in per-vector storage at 94.4% of baseline effectiveness. Because evaluation times were measured once, we treat them as indicative: we find no evidence that quantization speeds up evaluation, whereas sampling from intermediate layers shortens it substantially. Retrieval degrades far faster than every other category, so we report findings per task category.
Recursive self-improvement requires AI systems to improve the models and infrastructure used by subsequent systems. LLM compression offers a concrete efficiency objective: retain a pretrained model's behavior within a smaller model storage budget. We study whether coding agents can carry out this step autonomously. We formulate compression as a terminal task in which an agent receives a source checkpoint, storage limit, inference environment, and time budget, then constructs compressed weights and executable inference code. The agent controls its algorithm and experiments; perplexity feedback guides search, and independent verification checks the final model. Across six agents and six model--size settings, 29 of 36 submissions pass verification. Selecting the lowest-perplexity agent design per task yields an average 3.4% perplexity increase over BF16 and 77.33% downstream accuracy, compared with 79.56% for BF16. Recorded experiments reveal how agents achieve these outcomes and where improvement breaks down. Agents lower PPL in 21 of 24 revisions at nearly fixed storage; complete-model feedback reverses 17 of 79 prior local rankings; and generation and inference tests expose failures that likelihood alone misses. The results identify representation design, experimental judgment, and capability validation as central to agent-driven model compression, a measurable component of efficiency-oriented self-improvement.
We study which training and calibration data help three open models solve ARC-AGI-2 tasks when the solver runs offline on four L4 GPUs for at most 12 hours. We compare data offline on public tasks and report 189 runs on the 120 hidden semi-private tasks. For the NVARC Qwen3-4B grid model, we built a corpus of 12,000 executable rules in a typed language and 95,993 tasks; a rule is admitted only if its behaviour on a fixed probe set differs from that of every admitted rule, and a task only if its pairs refute a shortlist of nearest competing rules and defeat 12 shortcut solvers. Continued fine-tuning on 1,920 real task records with revised four-GPU scheduling raised the semi-private score from 28.96 (mean of the released model) to 33.06 (mean of five runs, best 34.86); no checkpoint trained with synthetic tasks, even all 95,993 of them, improved on the real-task model. For the Qwen3.8-27B reasoning model, we fine-tuned on 1,036 of its own reasoning traces with correct final answers and compared sets that add traces of other models and synthetic tasks. With an inference procedure built around the model (a thinking cap enforced in the serving engine, a Python program checked on the demonstrations after each answer, a long first attempt per input), the score rose from 26.81 to 44.03, the level the ARC Prize Foundation measured for the base model without a time limit (42.36); most of the gain came from the inference procedure. For the Qwen3.8-Flash-Next mixture-of-experts model (125B parameters, 6B active), GPU memory limited concurrency. Pruning from 512 to 384 experts per layer by router-weighted activation saliency grew the key-value cache from 0.55M to 1.39M tokens; a 296-expert list adapted with ARC routing statistics, keeping experts for program tokens, plus an FP8 indexer cache, reached 1.96M. With a tool that checks programs, the model scored 56.67, our best result.
Wi-Fi~7 and 5G NR use constellations up to $4096$-QAM and $1024$-QAM at multi-gigabit peak rates. The MIMO detector must sustain these rates at a low cost. Linear MMSE detection meets these rates at low cost but loses several dB in SNR. $K$-Best detection recovers this loss, but its per-layer sorting grows with the modulation order and needs large sorting networks or deep pipelines at these rates. Box Decoding is a sort-free tree-search detector whose complexity does not depend on the QAM order, and its deterministic pruning keeps the cost fixed as the MIMO size grows. Single-step candidate pruning (SCP) is fully parallel but less accurate, while iterative candidate pruning (ICP) is more accurate, but its selection loop costs either throughput or area. This paper proposes Max-Min-Exchange Candidate Pruning (MMECP), which replaces the selection loop of ICP with a single max-min exchange without a feedback path and keeps \mbox{near-$K$-Best} performance. The paper also presents a VLSI architecture that represents each candidate by its integer constellation index, so that candidates and their distances are obtained by bit selection instead of operations on symbol values. A $4\times4$ detector synthesized in TSMC $16$\,nm detects one vector per cycle at $1$\,GHz for $64$- to $1024$-QAM and produces up to $40$\,Gb/s, while its gate count grows by $63\%$ over this range. Its normalized hardware efficiency is $2.3\times$, $12\times$ and $14\times$ that of the best compared hard-output detector at $64$-, $256$- and $1024$-QAM.
Large language models (LLMs) undergo safety alignment to prevent harmful generations, yet they remain vulnerable to jailbreak attacks. Recent work reveals a deeper structural weakness: safety alignment can be highly sparse, such that pruning a small number of safety-critical neurons or ablating a single refusal direction can substantially undermine aligned behavior. In this work, we take the first steps toward strengthening safety alignment such that it is less dependent on sparse model components. We propose Addition by Subtraction, which gradually removes identified safety-critical neurons during finetuning and uses preference optimization to re-establish safety-aligned behavior through other parts of the model. Our results show that the resulting models are substantially more robust to pruning-based attacks. Remarkably, these benefits generalize beyond the white-box attacks motivating our design: Addition by Subtraction improves robustness against other sparsity-based interventions, including refusal-direction ablation, as well as a broad range of black-box jailbreak attacks. The finetuned models also exhibit improved truthfulness while largely preserving general capabilities. Comprehensive experiments across four models from three vendors demonstrate consistent improvements in safety robustness, reducing attack success rates by up to 90% for some attacks.
Training high-performing neural networks often involves evaluating populations of candidate models with varying hyperparameters, architectures, or random seeds. Training every candidate to completion is computationally expensive, yet stopping seemingly weaker runs early risks discarding models that would ultimately perform better, as learning curves can cross. To address this challenge, we investigate whether intermediate training checkpoints can predict future model accuracy and enable more informed pruning decisions. Our nowcasters predict test accuracy five epochs ahead or at the end of training. They are trained with accuracy and loss histories, weight statistics, and SANE representations as inputs. SANE is a transformer autoencoder that encodes a checkpoint as a sequence of latent tokens. We then use the nowcasters in Survival-of-the-Fittest (SOTF) training. SOTF repeatedly ranks runs during training by predicted accuracy and stops the lowest-ranked ones. Our results show that weight statistics work best for five-epoch forecasts, and SANE representations work best for final accuracy prediction. For evaluation, we compare SOTF with pruning the population by current accuracy. In homogeneous model zoos such as the CNN zoos, pruning by current accuracy already retains a final model only slightly worse than the best. Model zoos with a larger spread in performance, such as the ResNet-18 and ViT zoos, are more difficult to prune by current accuracy alone. In these zoos, SOTF with nowcasted final accuracy almost closes the gap to the best model.
Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substantial context overhead and dilute the agent's attention to task-relevant information. Existing token pruning and summarization methods remain constrained by discrete text representations, limiting the achievable compression ratios. In this work, we challenge the prevailing assumption that agents must consume tool observations as explicit text tokens. To this end, we propose Coca-Cola, a continuous latent context compression model for LLM agents. Coca-Cola compresses tool observations into latent tokens and embeds them into the input embedding space of the agent. We further introduce Action-Preserving Training, which aligns the agent's next-action distributions under full and compressed observations. This trains the compression model to learn action-sufficient latent tokens that retain decision-critical information. Experiments with web-search agents on four QA benchmarks and two model families show that Coca-Cola reduces the agent's token consumption by 70.01%-89.24% and end-to-end runtime by 24.45%-55.97%, while consistently improving task performance. These results suggest that Coca-Cola offers a promising approach for context compression in agents. Our code is available at https://github.com/THU-Agent/Coca-cola.
Offline retrieval-augmented generation (RAG) on edge devices is updated through signed knowledge packs that carry text and precomputed embedding vectors. A signature proves who published a pack, not that its vectors encode its text. We study this text-vector integrity gap. A verifier that recomputes vectors and accepts them within a cosine tolerance concedes a predictable share of queries to an attacker who controls the vectors: the hijack rate equals the distribution of angular retrieval margins at the tolerance (the margin law). Margins are small in five corpora and two embedding models, so an idealised tolerance at the level of floating-point noise leaves up to 5.5% of queries exposed (at most 1.2% in any corpus once exact duplicate documents are excluded), and a threshold calibrated from our CPU INT8-FP32 comparison of a small embedding model (at the empirical 0.1% quantile of honest cosines) leaves 98-100% exposed; on a Raspberry Pi 5 running a local language model, such an attack lowers correct answers from 0.75 to 0.39. Integrity audits of compressed vector stores fare no better: on a FAISS product-quantisation index over 57,638 documents, tampered vectors that all pass the audit put the attacker's document first for 25% of targets. Verifying quantisation codes instead, with ambiguity decided by the verifier, a calibrated guard band and canonicalisation, accepted every honest vector in our samples on three platforms and left the publisher no choice. We also show that common floating-point and dynamic-INT8 inference pipelines are not bit-reproducible across batch sizes or architectures, and report pre-registered negative results on abstention, context pruning and semantic screening. We release code and data.
Target-conditioned diffusion drafters predict candidate tokens in parallel, yet each new block must wait for updated target-model hidden states. This forces drafting and verification to run sequentially, limiting speedup even though each block is generated in one pass. We introduce D3D, a standalone diffusion drafter that removes this inter-round dependency. D3D combines blockwise denoising with auxiliary autoregressive supervision to achieve long accepted prefixes without target hidden states. This independence lets D3D draft successor blocks during verification in one packed pass, enabling asynchronous speculative decoding. We progressively prune layers according to their effect on acceptance length to reduce the cost of packed drafting. Removing five of 36 layers retains over 99\% of mean acceptance length across three benchmarks with Qwen3-8B as the target. Across seven benchmarks, D3D averages 5.00$\times$ and 5.71$\times$ decoding speedup with Qwen3-8B and Qwen3-32B targets, compared with 4.03$\times$ and 3.67$\times$ for DFlash. On GSM8K with Qwen3-32B, speedup reaches 7.48$\times$. Code is publicly available at https://github.com/fudan-generative-vision/D3D.
Open-weight language models that fit on a single consumer GPU are becoming capable on a growing range of tasks, and they matter because they can run locally without sending the code to a hosted service. Vulnerability reproduction, turning a vulnerability description into an input that crashes the vulnerable build, is a natural target for them, since it is an automated analysis task closer to test generation than to code comprehension and it runs against a build that must be executed rather than read. In order to assess whether current models can perform it, we evaluate twelve configurations of local models, spanning three generations, two architectures and three quantization levels, on ten CyberGym vulnerability-reproduction tasks, three runs each under one harness, and release the full artefact. We show that every configuration, down to the 4B model, solves most of the six easier tasks, and configurations separate only on the remaining four. For the harder tasks, the 27B model is required. Quantization did not change what the models solved in any matched pair, but it changed how fast they got there, and the mixture of experts reached the same scores in less wall clock. We also examine the agent trajectories. Almost every run we examined, including those that fail, names the vulnerable function and states the correct hypothesis, and what separates a solved task from an unsolved one is whether that hypothesis becomes a working input.
Achieving performance portability is challenging for irregular scientific applications built as repeated multi-kernel GPU pipelines, because throughput depends on out-of-kernel decisions as much as on kernel efficiency. Using molecular docking as a representative case study, we propose an out-of-kernel tuning methodology that treats batching policy, worker concurrency, and per-worker memory allocation as explicit optimization parameters, replacing backend-dependent occupancy-driven choices with a more portable execution strategy based on explicit resource bounds. We implement this methodology in the muDock mini-app under a manager/worker execution model where multiple host workers share a GPU through independent streams. We then evaluate the methodology on NVIDIA, AMD, and Intel GPUs across CUDA, HIP, and SYCL implementations. Across all evaluated platform/backend pairs, the memory-bounded formulation exceeds the peak throughput of the occupancy-guided baseline. The fully tuned methodology improves the end-to-end throughput on all tested configurations, achieving an average speedup of 1.53x over the baseline.
Molecular dynamics predicts molecular motion from atomic forces and trajectory integration. Density functional theory (DFT) provides first-principles forces, but its $O(N^3)$ cost limits simulations to thousands of atoms; empirical force fields scale to millions but miss bond breaking and charge transfer. Equivariant message-passing Transformers learn the ab initio potential at linear cost. Stacked layers propagate information beyond a single cutoff, giving four layers a receptive field up to 23Å. At this frontier, scale is limited by communication: layers exchange boundary features across devices, force backpropagation repeats the exchange, and moving atoms change neighbor graphs and tensor shapes. We present Tessera, a distributed runtime that fixes workload shapes between rebuilds. It sorts atoms along a space-filling curve into equal-count shards, ensuring identical atom counts and tensor shapes per rank. Peer sets scale as $K^{0.5}$, not $K$. A conservative neighbor skin preserves topology across updates, enabling each rebuild interval to replay as one CUDA Graph. Fused kernels, three-level neighbor search, and overlapped All-to-All complete it. On 512 GPUs, Tessera sustains 30.7M atoms at 89.7% weak-scaling efficiency (77--90% across three per-GPU loads); per-step time rises 1.11--1.29x as machine and problem grow 16x. Graph reuse cuts full-step time $1.87\times$. Across an $8\times$ device increase, neighbor-search cost grows $1.7\times$, versus $3.0\times$ for Morton order. In a live $10.24$M-atom salt solution, NVE energy drift is $0.42\,μ$eV/atom/ps over $10$ ps, and structure matches a serial reference. A 10.06M-atom Norovirus capsid sustains 100 ps of NVT at $\approx 8$ ps/day on 256 A100 GPUs. Over 200,000 steps, it has no broken bonds or anomalously close pairs and uses three compiled variants, with no graph breaks or eager fallbacks.
Many neuromorphic systems possess rich internal computational structures shaped by their physical dynamics. However, existing metrics characterize only global computation, leaving these local structures unresolved despite their importance for system interpretation, optimization and miniaturization. Here, we introduce local information processing capacity (L-IPC), which represents individual dynamical states as vectors in an infinite-dimensional Hilbert space and decomposes their computational properties through subspace projections onto nonlinearity orders and memory steps, with the residual quantifying computational irrelevance. Applied to three systems with comparable global capacities (an echo state network, an ion-gating reservoir and a spin-wave reservoir), L-IPC reveals markedly different distributions of computation, ranging from homogeneous to spatially, temporally and spectrally structured patterns. These local profiles further guide task-specific pruning, outperforming standard approaches at the same size in the structured systems. These results establish L-IPC as a general framework for anatomizing computation and guiding task-adaptive design in neuromorphic systems.
The efficiency of a modified genetic algorithm (MGA) with threshold selection and two-stage fitness evaluation is studied for designing computer device cores by substituting library elements. A library of 44 parameterised VHDL modules (111 variants in nine categories) was developed; its metadata were measured on the open GHDL, Yosys and nextpnr flow for Lattice ECP5 FPGAs. For the ALU-8 core, all 512 implementation variants were evaluated exhaustively, which provides the true optimum for comparison. With a median of 9 full syntheses, the MGA finds solutions from the top 5.7 % of the design space (Fmax rises from 45.3 to 103.9 MHz), i.e. it needs 4.4–11.8 times fewer syntheses than the classical GA, with a statistically insignificant difference in quality. For the FIR-16 core (|DS| ≈ 5.4·10⁸), the MGA with a median of 21 syntheses outperforms the classical GA with a budget of 40 syntheses in all 25 pairs of runs. Neither algorithm finds the solution that requires simultaneous replacement of eight identical parallel units, and a simple calibration of the fast estimate reduces the area error from 24–89 % to 3–11 %.
The relentless pursuit of higher integration densities in the semiconductor industry, marked by the transition to sub-5nm process nodes, has pushed traditional Design for Testability (DFT) methodologies to their breaking point. As System-on-Chip (SoC) designs scale to encompass billions of transistors and hundreds of heterogeneous Intellectual Property (IP) blocks, the conventional gate-level DFT insertion and verification flow has become a primary source of schedule unpredictability and late-stage design closure failures. This report presents a comprehensive analysis of the "Shift-Left" paradigm as applied to DFT, proposing a hierarchical, Register-Transfer Level (RTL)-based verification framework. By moving DFT architectural decisions, insertion, and design-rule checking (DRC) to the RTL stage, designers can concurrently optimize for power, performance, and area (PPA) while drastically reducing the time required for Automatic Test Pattern Generation (ATPG) and physical synthesis. The research explores the mechanisms of hierarchical partitioning, core-wrapping via IEEE 1500 standards, and the integration of functional fault grading to bridge the coverage gaps inherent in structural testing. Central to this framework is the concept of silicon correlation ensuring that pre-silicon RTL-based metrics accurately reflect post-silicon yield and diagnostic results. Through an examination of industrial case studies and recent advancements in Electronic Design Automation (EDA), this report demonstrates that a hierarchical RTL-based approach provides a scalable and predictable path to achieving zero-defect quality in modern high-performance computing and automotive applications.
Unlike conventional architectures that treat memory as static pattern storage, the hippocampus organizes experience as ordered sequences and retrieves them through temporally compressed replay spanning prior, current, and anticipated states. This capacity of context-appropriate sequence prediction remains absent from conventional deep learning accelerators and existing neuromorphic fabrics. We present HippocamFLEX, a fully synthesizable digital implementation of these hippocampal computational principles, realized with SystemVerilog RTL on the AMD Zynq UltraScale+,XCZU9EG. The design maps the hippocampal trisynaptic pathway onto three pipeline stages: a Dentate Gyrus / Entorhinal Cortex (DG/EC) stage for hypervector proximity matching, a CA3 stage for context-gated recurrent sequence prediction with multi-phase temporal tagging, and a CA1 stage that integrates past, present, and predicted future events into a theta-sequence inspired output. While much brain-inspired hardware relies on analog mixed-signal design for energy efficiency, this work targets FPGA prototyping as a deliberate architectural choice: FPGAs offer rapid design-space exploration without tape-out cost, bit-exact determinism that simplifies verification. The design supports 64 events across 16 contexts and is verified on physical silicon across context-isolation, temporal-ordering, and mid-sequence sliding-window test cases and is inherently scalable to larger event and context sets.
In the implementation of the emerging embedded computer-vision systems the real-time edge detection is becoming an essential preprocessing operation. The fixed-threshold Sobel realizations have advantages because they have simple arithmetic circuits with low complexity, but it is difficult for a single threshold to explicitly differentiate between strong responses, uncertain weak responses, and its background. In this paper a confidence-aware adaptive Sobel edge-detection architecture is presented and realized on FPGA with timing-closed performance and physically validated FPGA realization on a Zynq-7000 XC7Z020 Zed-Board. The architecture basically used a method that combines a conventional 3×3 Sobel gradient operator with three-level image recognition metric as confidence classification and surrounding neighborhood supported approval of intermediate-gradient responses. The recommended hardware architecture is based on utilization of synchronous dual-port frame buffers, a pipelined processing or display path, and a mixed-mode clock manager (MMCM)-derived VGA pixel clock. The behavioral verification experiment is executed for 1,679,998 pixel-clock samples, 2,099 raster lines, and four frame starts with zero assertion errors and zero monitor warnings; the calculated pixel-clock frequency was 25.174825 MHz. The post-synthesis timing shows WNS = −0.443 ns and TNS = −2.069 ns, whereas the routed implementation achieved improvement in WNS = +0.022 ns, TNS = 0 ns, and WHS = +0.064 ns. The implemented design utilizes 1,563 LUTs, 1,284 flip-flops, one RAMB36 and two RAMB18 blocks, and one MMCM. The total post-route analysis estimates the on-chip power of 0.254 W. The results obtained shows the practical hardware implementation feasibility and timing closure with the generated bitstream. The target hardware was programmed on the Zed-Board and shows validation using an externally connected VGA display. The expected three-panel display—input in grayscale, conventional Sobel as baseline, and confidence-aware adaptive as output—was reported, and a multiple SW [7:0] switch configurations was exercised and produced repeatable threshold-dependent changes in the detected-edge maps. The observed results establish the simulation correctness, timing closure, bitstream generation, and physical on-board VGA functionality test for the desired system.
Level-1 Triggers at the LHC classify collisions within a few microseconds on FPGAs, where a network is a fully unrolled circuit whose cost is area rather than time. We present SCRAMJet, a transformer encoder that replaces multi-head attention with a low-rank projection, trains under per-parameter learned fixed-point quantization, and compiles to Verilog with distributed arithmetic optimization. On the ten-class JetClass dataset it reaches $71.4 \pm 0.6\%$ accuracy with 0.157M multiply-accumulates, 6.1 accuracy points above the strongest published resource-constrained model at an eighth of the operations, and meets timing on a Virtex UltraScale+ VU13P at 147 ns using 25.7% of the logic. We also show that resource-aware training silently removes attention channels: one whose learned bit-width reaches zero still carries signal during training but is absent from the compiled circuit, so the loss never sees it go. We call these virtual channels, and a floor on the attention datalane that prevents them is worth 3.1 accuracy points at matched budget. Code and trained checkpoints are available at https://github.com/Aaddy-1/hgq-jetformer.
Designs for collective digital systems require efficient access to memory. Examples of digital systems include: processor, embedded systems, and systems-on-chip (SoCs). Standard methods of designing RAM modules typically involve writing and reading a single complete data word. However, this often leads to the overwriting of currently existing data already stored in the RAM module. Therefore, these systems can consume significant amounts of energy and inefficiently utilize the digital system’s available bandwidth. All digital systems are therefore designed with this form of data loss issue in mind. So, for instance, while designing a CPU one must pay particular heed to the RAM module for new and modified data being written; to ensure no data in the RAM module will be lost because of the newly written data. In the body of this work I will describe a technique for designing RAM that can be written and read in byte mode opposed to word mode. This will be accomplished using only a single instance of a 32-bit data word, and designing this uses Xilinx Vivado.
The paper gives the implementation and full study of video encoder system based on FPGA by the use of Xilinx Zynq-7000 series device (xc7z020clg484-1). The design shows good use of hardware resources as the slice logic is optimized to reach the 4.15 percent LUT utilization and 5.18 percent register utilization. The video encoder was implemented with Vivado 2023.2 design suite with real-time video processing application. Strategic analysis of resources indicates the complex data path optimization using F7 and F8 multiplexers, and low power consumption due to minimum DSP and memory block utilization. The performance indicators show that it is an efficient design that completes synthesis successfully and achieves timing closure, thus, can be applied to embedded video processing systems.
Approximate computing has become a prominent design paradigm for energy-constrained edge AI hardware, where relaxed arithmetic precision enables significant gains in speed, area, and power. This paper presents a novel $n$-bit Hybrid OR-MUX Lower-Part Approximate Adder (HOMAxA-1) that divides operands into an approximate lower part (bits $[k-1: 0]$) and an exact upper part (bits $[n-1: k]$). In the 8-bit instance $(n=8, k=4)$, the lower nibble employs bitwise OR with generate-based compensation for sum bits [1:0], governed by OR for bit [2], and MUX-controlled carry-select for bit [3]; the upper nibble part computes an exact addition using an AND-generated carry. Exhaustive evaluation over all 65,536 input combinations under a carry-preserving 9-bit output convention yields ER=57.81%, MED=1.5625, NMED=0.00306, and WCE=7. Benchmarked against five conventional, well-established approximate adders, HOMAxA-1 attains the lowest NMED and lowest WCE of the compared set. The proposed architecture was implemented using synthesizable Verilog HDL and validated on a Xilinx Artix-7 FPGA using the Xilinx Vivado.
Thermal transients can evolve faster than software and firmware control loops, leading to conservative throttling. Prolepsis is an ISA-independent thermal management unit (TMU) that samples telemetry every 4096 cycles $(41.80 \mu \mathrm{s}$ at 98 MHz), executes one-step Q16.16 Kalman prediction in RTL, and selects phase-aware duty throttling, QoS steering, cache warming, or firmware-assisted migration. A 5-core RISC-V SoC implementation targets an AMD Xilinx UltraScale+ XCU250 FPGA. FPGA synthesis validates the RTL, while calibrated compact-model replay evaluates thermal behavior. The TMU uses 18,015 LUTs (1.04% of device capacity and 12.55% of SoC logic) and reaches an $F_{\max}$ of 101.16 MHz. Prediction achieves an RMSE of 0.087° C and MAE of 0.059° C. At an 85.0° C limit, Prolepsis maintains normal-speed execution for 96.3% of cycles versus 68.7% reactively, while cache warming reduces 90%-IPC recovery from 446 to 118 cycles. Thus, RTL-resident prediction enables faster, ISA-independent thermal management.
Resistive RAM (RRAM) crossbars enable inmemory analog matrix–vector multiplication (MVM). However, practical devices exhibit strongly nonlinear current–voltage $(I-V)$ characteristics that degrade the accuracy of conventional voltage-amplitude encoding. This work evaluates a special case of Stream Computing called Asynchronous Stream Computing (ASC) as a circuit-level approach to mitigate this nonlinearity in RRAM-based compute-in-memory (CiM). ASC encodes inputs in the time domain as the duty cycle of constant-amplitude 1-bit $\Sigma \Delta$ streams, so each RRAM cell operates at a fixed Read bias, $V_{\text {ref }}$, and device nonlinearity no longer directly distorts the output current. Circuit-level simulations using the Stanford Verilog-A RRAM model in Cadence Virtuoso (65 nm PDK) compare conventional analog CiM and ASC-CiM under identical device parameters. ASC reduces the maximum relative error for single cells from 25.41% to 0.94% in high-resistance state (HRS) and from 25.57% to 3.96% in low-resistance state (LRS). For 4 × 1, $8 \times 1$, and $16 \times 1$ crossbars, ASC lowers the mean full-scale normalized MVM error from $\mathbf{5 . 7 6 \% - 1 0 . 1 \%}$ to $\mathbf{0 . 0 4 0 5 \% - 0 . 0 2 6 0 \%}$, corresponding to up to 387× improvement. These results indicate that ASC substantially reduces nonlinearity-induced MVM error in the evaluated RRAM CiM configurations and offers a tunable accuracy–latency tradeoff via the accumulation window $T_{\text {acc }}$.
A key challenge in colon capsule endoscopy (CCE) analysis is accurately identifying whether multiple images depict the same polyp, which is critical to avoid over-counting and unnecessary clinical interventions. Additionally, this prevents prolonged follow-up colonoscopies where endoscopists search for duplicate polyps without precise localization. This study presents a novel contrastive learning-based approach to polyp matching using a Siamese neural network architecture. The model was trained on a dataset from the Danish CareForColon2015 trial. The dataset comprises 5,103 image pairs of confirmed same polyp and 5,103 pairs of different polyps, annotated by experienced clinicians. The proposed model employs ResNet variants (ResNet18, ResNet34, ResNet50), DenseNet121 and EfficientNet B0/B3 as backbones and utilizes contrastive learning with cosine vector distance metric for classification. Experimental results demonstrate high classification accuracies of up to 92% and AUC values close to 0.98. These results suggest that automated polyp matching has the potential to reduce polyp over-counting, improve patient risk stratification, and reduce unnecessary follow-up colonoscopies.
Refinery decisions often require evidence that is distributed across plant historians, maintenance systems, work permits, incident records, management-of-change registers, audit actions, shift logs, engineering documents, and workforce-compliance records. Conventional search or dashboard solutions generally address one source type at a time and leave engineers to assemble the complete operational story manually. This paper presents the Autonomous Refinery Intelligence Architecture (ARIA), an on-premise, evidence-grounded architecture developed in-house at a public-sector refinery to unify structured plant records and unstructured operational knowledge while maintaining data sovereignty. The architecture is designed to connect 15 enterprise systems through governed tools and a Model Context Protocol integration layer. A refinery-wide source inventory contained 423,614 structured records and 6,898 shift-log records across 12 operational domains. The detailed case study evaluates eight operational source classes using controlled subsets from crude and vacuum distillation and hydroprocessing units. The method combines canonical equipment identity, deterministic source-quality checks, structured and chronological candidate generation, retrieval-augmented language-model synthesis, a refinery-specific qualification gate, and human review. Three domain patterns are supported: safety investigation, reliability investigation, and contractor-workforce compliance. The language model acts as a controlled reasoning and summarization layer; it is not permitted to authorize plant or workforce decisions. The evaluated crude-distillation pipeline assembled 52,664 evidence rows and reduced 124 candidate packs to 15 final cards, of which only three met the strongest machine qualification. The hydroprocessing pipeline assembled 81,313 rows and reduced 132 candidate packs to 15 cards, again with three receiving the strongest qualification. The deliberate rejection and review rates show that conservative evidence gating is more important than maximizing the number of generated insights. The evaluation also exposed material source-data limitations, including missing work-permit tags, invalid notification dates, and incomplete closure fields. These limitations were carried forward as confidence constraints rather than silently repaired by the model. An initial internal workflow benchmark for a representative incident investigation indicated approximately 80% less evidence-assembly time than the prior manual approach. This is a preliminary operational benchmark, not a controlled statistical trial. A 15-20% reduction in unplanned downtime remains a projected benefit to be tested through longer-term asset performance monitoring. The principal conclusion is that agentic artificial intelligence can support refinery operations only when identity, provenance, qualification rules, human accountability, and on-premise governance are designed as first-class system components.
The relevance of the research is determined by the rapid digitalization of cultural heritage, the growing volume of digital images, textual descriptions, metadata, and the increasing need for accurate, adaptive, scalable, and explainable automated identification and classification of cultural heritage objects. The problem addressed in the study is the limited effectiveness of traditional and partially integrated approaches based on k-NN, manually designed descriptors, convolutional neural networks, and static knowledge graph representations when processing heterogeneous, incomplete, and semantically complex cultural heritage data. The purpose of the research is to analyze modern approaches to intelligent decision support systems for cultural heritage identification and to substantiate an improved multimodal architecture capable of integrating visual, textual, and semantic information. The research methods include comparative analysis of existing approaches, analysis and generalization of previous research results, conceptual modeling, and evaluation of modern deep learning, multimodal learning, and graph-based methods. Particular attention is given to Graph Neural Networks, Graph Attention Networks (GAT), Relational Graph Convolutional Networks (R-GCN), Vision Transformers, BERT-based language models, self-supervised learning, and Explainable AI. The results demonstrate that replacing static graph representation methods with GAT and R-GCN enables more effective consideration of relationship types and weights, contextual information, and incomplete or noisy data. A multimodal architecture integrating Vision Transformer for image processing, a BERT-based model for textual information, and GNNs for Knowledge Graph representation is proposed. Their integration into a shared embedding space enables cross-modal analysis, semantic search, and zero-shot and few-shot classification. The study concludes that the transition from isolated classification models to integrated multimodal intelligent decision support systems can significantly improve the accuracy, robustness, adaptability, and interpretability of cultural heritage identification. Further research should focus on adaptive decision-making, self-supervised learning, Explainable AI, automatic Knowledge Graph enrichment, integration with national and international cultural heritage databases, human-in-the-loop approaches, and federated learning.
Evidence-aware ranking reduces rating manipulation and commercial bias in local service directories by conditioning vendor order on verified reviews, physical distance, and identity trust stored in a shared database index. Deploying hyperlocal discovery as a multi-role civic product requires durable relational storage, role and booking isolation, and a ranking path that does not force a separate paid-bidding or machine learning service beside the application. This paper presents a PostgreSQL-centred architecture for that setting. All vendors, service categories, and user roles share one PostgreSQL database with GIN full-text and GiST spatial indexes. Isolation and rating integrity are obtained by tagging every booking and review with customer and vendor identifiers and applying a completed-booking check before any rating updates the vendor score. Sparse ratings are stabilised by Dynamic Bayesian Shrinkage, distance is modelled by continuous Haversine decay, and identity trust is captured by a Multi-Factor Verification Trust Score (MFVTS). The design is realised in Smart Vendor Hub, an Angular 19 and FastAPI platform that connects citizens with local tradespeople, answers searches with Explainable Information Retrieval (XIR) score breakdowns, and exposes the same ranking function to a citizen discovery portal and to a live Researcher simulation lab. We describe the index schema, the ranking and review-gate algorithms, the isolation invariant, and an empirical evaluation on a deployed instance. The result is a practical blueprint for using PostgreSQL as a shared ranking store in hyperlocal service discovery without pay-to-play ad auctions or an external neural ranker.
Background: Early diagnosis of Alzheimer's disease (AD) and mild cognitive impairment (MCI) represents a critical unmet need. Artificial intelligence (AI) and Machine Learning (ML) models may leverage speech features to detect AD/MCI before overt clinical presentation. Despite growing interest in AI applied to speech and language analysis as non-invasive cognitive screening biomarkers, available evidence remains fragmented and heterogeneous regarding diagnostic accuracy across populations, languages, and architectures. Objective: To systematically evaluate the diagnostic accuracy of AI/ML models based on speech and language analysis for early detection of AD and MCI in comparison to healthy controls. Methods: A systematic review was conducted following PRISMA 2020 guidelines. PubMed and Cochrane databases were searched (last 5 years) using MeSH and free-text terms combining "Alzheimer Disease," "Cognitive Dysfunction," "Machine Learning," "Artificial Intelligence," "Speech," and "acoustic features." Studies applying supervised AI/ML classification models trained on speech or language features from AD/MCI populations and reporting accuracy, area under the curve (AUC), sensitivity, or specificity were included. Reviews and studies without extractable performance metrics were excluded. Three independent reviewers performed selection and data extraction via Rayyan. Results were synthesized qualitatively. Results: From 175 records identified, 23 studies were included (2021–2025), covering populations from 9 countries across 5 languages. Eighteen studies addressed binary AD vs. healthy control (HC) classification and five employed multiclass (HC/MCI/AD) approaches. Models employed acoustic features (Mel-frequency Cepstral Coefficients, Fundamental Frequency, speech pauses), linguistic/Natural Language Processing features (Bidirectional Encoder Representations from Transformers embeddings, Part-of-Speech tags, bag-of-n-grams), and multimodal fusion. For binary classification, accuracy ranged from 78.0% to 99.8% and AUC from 0.743 to 0.990. For MCI detection, AUC ranged from 0.721 to 0.988. Transformer-based and Large Language Models (LLM) consistently outperformed traditional classifiers. Kuroda et al. (2025) achieved AUC=0.990, ACC=95.0%, Sp=100% (Bidirectional Long Short-Term Memory + Fully Connected Neural Network); Wang et al. (2025) reported AUC=0.988 for MCI (Deep Learning+LLM); Chen et al. (2025) obtained AUC=0.950, F1=0.816 in multiclass (n=1,098). Limitations included heterogeneous tasks, small samples, and limited external validation. Conclusion: AI-based speech analysis demonstrates high diagnostic potential for early AD and MCI detection, with AUC consistently exceeding 0.80 across most studies. Multimodal and transformer-based approaches represent the state-of-the-art, particularly for non-invasive MCI screening. Methodological heterogeneity limits clinical applicability, reinforcing the need for prospective multicenter studies with standardized protocols and external validation.
Background: Alzheimer’s disease is the leading cause of dementia worldwide, affecting approximately 50 million individuals. Early diagnosis remains a major challenge, as conventional neuroimaging methods often detect structural changes only after significant neurodegeneration has occurred. Advances in artificial intelligence have enabled computational models capable of identifying subtle neuroimaging patterns associated with early disease stages. Machine learning and deep learning techniques have increasingly been applied to neuroimaging data to detect early structural and metabolic alterations, potentially improving diagnostic accuracy and enabling earlier clinical intervention. Objective: To evaluate the diagnostic performance of artificial intelligence models applied to neuroimaging for the detection and early diagnosis of Alzheimer’s disease. Methods: A systematic review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines using a search strategy structured according to the Population, Intervention, Comparison and Outcome framework. The protocol was prospectively registered in the International Prospective Register of Systematic Reviews. Studies investigating artificial intelligence algorithms, including machine learning and deep learning approaches applied to neuroimaging for diagnosing Alzheimer’s disease or mild cognitive impairment, were included. Searches were performed in PubMed, Cochrane Library and LILACS for studies published between 2021 and 2026. A total of 2,085 records were identified. After filtering, 49 studies were screened using the Rayyan platform. Following duplicate removal and title and abstract screening, 12 articles underwent full-text evaluation and 8 met the eligibility criteria for qualitative synthesis. Results: The included studies applied several artificial intelligence techniques, including convolutional neural networks, generative adversarial networks, random forest models and other deep learning architectures. Neuroimaging modalities included structural magnetic resonance imaging, functional magnetic resonance imaging and positron emission tomography, including amyloid and tau imaging techniques. Diagnostic performance was high, with accuracy ranging from approximately 85% to 94% and area under the receiver operating characteristic curve between 0.87 and 0.94. Multimodal approaches integrating different imaging modalities or biomarkers showed superior performance compared with single-modality models. The hippocampus, medial temporal cortex and temporoparietal regions were the brain areas most frequently associated with diagnostic prediction. Conclusion: Artificial intelligence applied to neuroimaging shows strong potential to improve the early diagnosis of Alzheimer’s disease, demonstrating high diagnostic accuracy and the ability to identify relevant neurodegenerative patterns that may support clinical decision-making in neurodegenerative disorders.
Neural architecture search (NAS) aims at automatically searching for high-performance architectures and has achieved considerable progress. Existing NAS methods learn architectures by minimizing average-case validation losses. As a result, the searched architectures are less capable of making correct predictions under worst-case scenarios. To address this problem, we propose a framework that leverages a deep generative model to generate adversarial validation examples to measure the worst-case validation performance of an architecture and improves the architecture by minimizing the loss on the generated adversarial validation data. Our framework is based on multilevel optimization, which performs multiple learning stages end-to-end. Experiments on a variety of datasets demonstrate the effectiveness of our method.
Graph Neural Networks (GNNs) have been shown to be vulnerable to backdoor attacks, where attackers implant hidden backdoors into model weights of GNNs. However, backdoors hidden in GNN weights are non-persistent and easily erased by retraining models on clean data. To this end, we investigate a new threat of implanting backdoors in GNNs neural architectures that is independent of the model weights of GNNs. A new backdoored graph neural architecture search model (BGNAS for short) is proposed to use graph neural architecture search algorithms to discover new GNN neural architectures with backdoors. Unlike previous works that rely on manually designed backdoored neural architecture and triggers, BGNAS simultaneously designs backdoored GNN neural architectures associated with their trigger generators, where the generated triggers are indistinguishable within the overall graph data to ensure stealthiness. Theoretical analysis demonstrates that backdoors embedded in graph neural architectures remain effective even after weight re-initialization and retraining on clean data. Experimental results further show that BGNAS achieves high attack effectiveness while maintaining competitive prediction performance on benign data.
Decisions about the direction of research in power engineering determine which technologies mature and, through them, shape infrastructure that lasts decades, yet their consequences cannot be observed before they are taken, and existing forecasts of research fields extrapolate a single trajectory. Here we introduce research world models, which treat a research field as a dynamical system that can be acted upon, and establish one for power engineering from 1.1 million papers. A distilled language model reads each paper into role-typed concepts, two sophisticated, high-fidelity graph-temporal neural network cores learn how the resulting concept network evolves, and operators learned from natural experiments in the publication record estimate the effects of interventions. In temporally held-out tests, the model forecast years ahead which concepts would grow and which method-target pairs would be joined for the first time, outperforming tuned machine-learning baselines in most settings, and its estimated intervention effects tracked those realized in natural experiments it had not seen. Its training scheme and three of its operators replicated in two fields withheld from development. Questions posed in natural language are compiled by a domain-specialized 27-billion-parameter language model into executable actions and answered as paired simulations with and without the intervention. The answers were rated 4.1 of 5 for faithfulness by an independent language-model judge. By revealing the latent dynamics of a research field and its response to interventions, research world models offer a new way to simulate and test research decisions before they are made.
As deep learning (DL) models become increasingly complex and hardware architectures grow more heterogeneous, DL compilers play a critical role in optimizing computational efficiency and enabling scalable deployment. This paper presents a systematic analysis of DL compilers, covering their architectural evolution, compilation workflows, optimization techniques, and emerging research directions. We organize DL compiler architectures into frontend, middle-end, and backend stages, emphasizing the role of the middle-end in connecting high-level model representations with hardware-specific optimizations. We further examine representative DL compilers and recent optimization techniques across different compilation stages. Beyond architectural and technical analysis, we conduct an empirical evaluation of compiler behaviors to investigate the effectiveness of graph optimization, compilation overhead, and dynamic-shape handling. Experiments using PyTorch/TorchInductor on an NVIDIA RTX A6000 demonstrate that compiler benefits depend strongly on workload characteristics, with speedups of up to 8.21\times for fusion-friendly computation graphs, while other workloads exhibit limited improvements. Our results further reveal that compilation overhead can require thousands of executions to amortize, and dynamic-shape compilation reduces cumulative shape-transition costs by approximately 70\% at the expense of modest steady-state performance degradation. By combining systematic architectural analysis with quantitative experimental evidence, this work provides insights into the design, optimization, and practical deployment of modern DL compilers.
Oil and gas field development planning remains a slow, expert-dependent workflow, with multi-year plans scattered across subsidiaries in heterogeneous formats and tacit knowledge locked in individual experts. This paper presents an enterprise scientific research agent that compiles development plans into a unified knowledge base and enables autonomous scheme generation, simulation and continuous optimisation at enterprise scale. The architecture fuses a 240-indicator standardised ontology, multi-modal LLM knowledge compilation, a lifecycle-aware knowledge graph (facts, states, events), a RAG-augmented three-tier memory, long-horizon multi-agent orchestration (Build and Optimize modes with frozen-evidence acceptance), a unified MCP tool bus, and a git-style model relation tree that makes every run traceable and reversible. The platform has compiled 2,924 development plans across 14 subsidiaries of a national oil company. In a black-oil case, the agent autonomously generated water-flooding well-pattern schemes in two hours without human intervention, raising the ten-year recovery factor from 18.9% to 34.6% against a 30% target, while planning-cycle latency was substantially reduced. The results confirm that progressing from multi-agent research models to unified enterprise intelligent agents is a scalable path toward knowledge-driven decision-making in large, multi-asset petroleum portfolios.
Approximate Nearest Neighbors (ANN) search over neural embeddings powers modern retrieval systems such as Retrieval-Augmented Generation. Embeddings come in three representations (dense, sparse and multi-vector), and efficient search combines an index structure, a similarity function and a quantization scheme. Existing libraries fix one combination, hard-code each one separately, or factor only one side of the problem, and none covers all three representations. We present Vectorium and kANNolo, two open-source Rust libraries that split this design space into a data-management layer and an indexing layer. Vectorium abstracts the vector, the distance function, the quantizer and the dataset independently, and provides 11 encoders, from 32-bit floats to sub-bit codes. On top of it, kANNolo implements Hierarchical Navigable Small-World (HNSW) graph and Inverted Vector File (IVF) indexes that are generic over representation, encoder and metric, and a two-stage index that composes any first-stage index with any rescoring dataset. Since every combination is specialized at compile time, this modularity adds no run-time overhead, and the resulting indexes reach state-of-the-art performance: our Product Quantization matches Faiss in accuracy at competitive speed, our RaBitQ and HNSW match or outperform their reference implementations, and our IVF outperforms Faiss. Against six dense and five sparse systems on seven public collections, our indexes give the best memory-throughput trade-off on dense data in 9 of 12 collection-accuracy pairs and perform well on out-of-distribution queries. On sparse data, the systems built on Vectorium lead on all three collections, with Seismic and SeismicWave the fastest and kANNolo's indexes the most compact.
Machine learning supports materials design, atomistic simulation and experimental characterization, yet models and workflows remain fragmented. We present PaddleMaterials, an open-source framework integrating heterogeneous materials-learning tasks while preserving task-specific representations and scientific semantics. It unifies data processing, training, evaluation, pretrained-model management and workflows across six task classes: property prediction, structure generation, machine-learning interatomic potentials, electronic-structure fields, spectrum elucidation and microscopy image processing. Task interfaces record physical quantities, units, aggregation conventions and references, making cross-task composition verifiable. A shared compiler backend enables switching from dynamic-graph to compiled execution through one setting, using graph optimization, operator fusion and hardware-targeted code generation without changing model definitions. Across 65 model instances in 11 groups, compiled execution accelerated every instance in eight groups, with group median speedups of 11%-224% and group maximum speedups of 11%-336%. Case studies on LiMnO2 and toluene link structure generation or spectrum elucidation, interatomic-potential relaxation, property prediction and electron-density prediction, using same-geometry density functional theory (DFT) calculations as references. For LiMnO2, predicted formation energies for generated and relaxed structures differed from same-geometry PBE+U references by at most 7 meV per atom, while the computational cost of model steps was less than one-twentieth of a DFT single-point calculation. PaddleMaterials extends unified materials workflows beyond conventional property, potential and structure-generation tasks while supporting verifiable cross-task operations and compiled execution.
Accelerating modern foundational models has pushed domain-specific compilers toward increasingly complex, multi-tiered optimization pipelines. However, this heavy optimization focus introduces delicate tradeoffs between peak execution performance and correctness. Subtle codegen defects can trigger numerical degradation and memory or buffer corruption, invalidating multi-million-dollar training runs and degrading live serving quality without raising obvious compiler errors. In this paper, we present a battle-tested methodology for verifying and triaging Google's XLA:TPU compiler at hyperscale. We leverage fuzzing techniques, combining synthetic program generation with real production traces, while orchestrating AI agent swarms to autonomously generate stress tests, reproduce failures, and draft fixes. To pinpoint root causes in full workloads, we pair this with a hierarchical differential pipeline that bisects large graphs down to individual failing operations inside fused kernels. Deployed in production, our framework has already driven the resolution of over 142 compiler bugs (with 103 additional findings under review), from which we distill pragmatic lessons learned.
Production support teams spend a disproportionate amount of time on repetitive, crossservice triage: correlating order state, provisioning records, logs, and ticket history to answer why a given transaction is broken. Existing runbook automation is typically hardcoded, owned by a single engineering team, and brittle to new failure modes. We present the Workflow Intent Graph System, an architecture that lets any team declare a triage or operational workflow in YAML, compiles that declarative definition into a LangGraph state graph, and uses a large language model (LLM) to match natural-language queries directly to the graph node representing the correct entry point, then navigates the graph at run time to reach a diagnosis or an automated fix. The system supports sub-workflow reuse as edges between independently authored subgraphs, parallel step execution, and dynamic multi-workflow chaining, without requiring workflow authors to write orchestration code. We describe the architecture, the YAML-to-LangGraph compilation step, the node-level intent-resolution pipeline, and the resulting extensibility, auditability, and performance characteristics.
Video diffusion transformers (DiTs) process tens of thousands of spatio-temporal tokens per denoising step, so low-latency inference and training commonly spread them across several accelerators. Auto-parallelization systems price a sharding on its logical graph, but its cost depends on the program the compiler generates and, on direct networks such as TPU sub-slices, on which chip runs each logical rank; both are visible only after compilation, while the chip assignment must be fixed before it. We present AoiZora, a compiler-agnostic auto-parallelization framework built on two IRs, one for the program a partitioner shards and one for the program a compiler emits; it chooses the logical mesh, the layout of every tensor and operation, and the device assignment of an unsharded model by pricing its compiled programs, without executing any candidate. At the core of AoiZora is evidence, a compiler-neutral record of a compiled program that leaves its placement open, so that one compilation can be priced under many placements by a calibrated model of the accelerator network and the placement can determine which sharding is selected. On TPU v5e, AoiZora reduces the denoising latency of Wan 2.1 and HunyuanVideo by 1.23x and 1.32x on average over expert configurations and Alpa, finding each plan within 10 minutes; it also speeds up training on TPU v5p by 1.11x on 4 chips and 1.21x on 8 chips on average and parallelizes PyTorch programs on GPUs with the same core.
Human health does not present itself through a single modality [5]. Disease, wellbeing, affective state, cognitive status, and behavioural change leave traces across a broad spectrum of observable and latent signals: speech and language, body sounds, facial and gesturing behaviour, movement, physiology, medical imaging, laboratory measurements, electronic health records, and patterns of everyday human–technology interaction, social media, and IoT sensor data. Yet, much of contemporary artificial intelligence for health still operates on individual modalities, isolated biomarkers, or narrowly defined prediction tasks. Even nominally multimodal systems often treat multimodality as a late-stage engineering problem—the combination of separately extracted representations—rather than as a defining property of the patient and of health itself [1, 3]. This perspective argues for a conceptual and technical shift: the patient should be treated as a dynamic multimodal signal. The relevant computational object is not merely an image, waveform, clinical note, or physiological measurement, but a continuously evolving constellation of signals generated by one person in a particular biological, behavioural, social, and environmental context. Such signals differ radically in sampling rate, semantic level, reliability, observability, and clinical specificity. Some are acquired episodically in the clinic; others can be sensed longitudinally in daily life. Some directly index physiology, while others express health indirectly, for example through behaviour, affect, voice, language, or interaction. Their value frequently lies not in any one channel but in their temporal relationships, complementarities, contradictions, and deviations from a person’s own baseline. Foundation models offer a plausible route toward representing this complexity. The emerging paradigm of generalist medical artificial intelligence envisions reusable models that can operate across heterogeneous data and tasks with limited task-specific supervision [7, 8]. Recent work has similarly highlighted the potential of multimodal biomedical AI to integrate information from rich sources [1]. At the same time, foundation models are reshaping neighbouring fields such as affective computing, where large pretrained models increasingly replace isolated pipelines for recognising and modelling human states [9]. The opportunity for health AI is therefore not simply to construct larger modality-specific models, but to develop patient-level multimodal foundation models: models that learn integrated representations of individuals across modalities, time, tasks, and contexts. I propose three transitions as a compact research agenda for this development. First, health AI should move from biomarkers to patient representations. Biomarkers remain essential, but a collection of independent biomarkers is not equivalent to a computational representation of a patient. Patient representations should encode relationships among biomedical, behavioural, affective, and contextual evidence and should evolve as new observations become available. This reframes multimodality from the accumulation of inputs into the learning of a coherent latent representation of the person generating them. Second, the field should move from multimodal integration to cross-modal reasoning. Classical multimodal machine learning has established the importance of representation, alignment, fusion, translation, and co-learning across modalities [3]. Health applications, however, increasingly require models to reason about why modalities agree, when they conflict, what information is missing, and how the evidential value of one modality depends on another. Missingness is particularly important: in real clinical settings modalities are rarely complete, synchronised, or missing at random. A laboratory test may be absent because it was considered unnecessary; a wearable stream may disappear because a device was removed; a patient’s speech may change because of both pathology and context. A useful patient model must therefore distinguish an unobserved modality from a genuinely negative clinical finding, quantify uncertainty, and infer when cross-modal compensation is justified rather than simply imputing or masking missing inputs. Third, health AI should move from episodic prediction to longitudinal modelling. Human health unfolds over time. In many settings, the most informative observation is not an absolute measurement but a departure from an individual’s previous state. A patient-level model should therefore represent trajectories [2] and changing cross-modal relationships over hours, days, months, and years. Such longitudinal representations could support earlier detection [4, 6], personalised baselines, adaptive monitoring, treatment-response assessment, and ultimately carefully constrained forms of patient digital twins. Importantly, the model should not merely store more history; it should learn which temporal changes are clinically meaningful and how the significance of one signal depends on the evolving state of others. This changes how we should evaluate multimodal health foundation models. Progress should not be measured merely by the number of modalities accepted by an architecture or by improvements on a fixed downstream benchmark. More consequential questions are whether a model forms stable and transferable patient representations; whether it can exploit a modality when available without becoming dependent on it; whether it handles asynchronous and systematically missing information; whether it adapts to individuals without amplifying spurious correlations; whether it recognises when available evidence is insufficient; and whether its representations remain useful under shifts in population, device, environment, and clinical workflow. Robustness, uncertainty, fairness, privacy, explainability, and prospective validation are therefore not peripheral “responsible AI” additions but constitutive properties of any model claiming to represent human health. The convergence of multimodal learning, foundation models, digital health, and related and supporting disciplines such as affective computing creates an opportunity to rethink the basic unit of modelling in medical AI. The central thesis of this discussion is deliberately simple: the patient is the multimodal signal. Realising this thesis requires a move from biomarkers to patient representations, from multimodal integration to cross-modal reasoning, and from episodic prediction to longitudinal modelling. If these transitions can be achieved, future AI systems may integrate what can be seen, heard, sensed, measured, retrieved, and observed over time—moving beyond specialised prediction systems toward computational models of human health that are more holistic, adaptive, personalised, genuinely patient-centred, and ultimately more deserving of our trust.
Pesticide contamination in agricultural water systems threatens ecosystem health, water security, and sustainable agriculture. Conventional laboratory analyses are highly accurate but remain expensive, labor-intensive, and unsuitable for continuous field monitoring. This review summarizes recent advances in artificial intelligence (AI)-enabled sensing technologies for the real-time detection of pesticides in irrigation water, rivers, reservoirs, drainage canals, and groundwater. It covers electrochemical sensors, biosensors, hyperspectral imaging, fluorescence spectroscopy, optical sensing, and microfluidic platforms integrated with Internet of Things (IoT) networks, edge computing, and cloud analytics. Particular emphasis is placed on machine learning, deep learning, explainable AI, federated learning, graph neural networks, and multimodal data fusion for contaminant detection, prediction, and risk assessment. The distinctive contribution of this review is providing a pesticide-specific source-to-decision framework that links hydrological transport and chemical transformation to matrix-conditioned sensor responses, AI-assisted signal interpretation, validation maturity, and field deployment, thereby identifying where intelligent analytics can provide a meaningful advantage for pesticide surveillance in agricultural water systems.
Underground mines are high-risk working environments where hazardous gases, oxygen deficiency, elevated temperature, physical fatigue, and limited access to workers can lead to severe safety incidents. Delays in detecting dangerous conditions and initiating emergency responses can further increase the consequences of such incidents. This study proposes a conceptual intelligent Internet of Things (IoT)-based system for real-time monitoring of both workers' physiological conditions and their surrounding environmental conditions in underground mines. The proposed architecture consists of three main layers: a smart vest, a tunnel communication layer, and a central control station. The smart vest integrates physiological and environmental sensors to collect data such as heart rate, blood oxygen saturation, body temperature, motion status, methane, carbon monoxide, ambient oxygen concentration, temperature, and humidity. An ESP32-based processing unit is considered for local data acquisition and preliminary processing, while LoRa-based communication nodes are proposed to transmit data through underground tunnels. A central control station based on a Raspberry Pi is designed to receive, store, visualize, and manage monitoring data and emergency alerts. A threshold-based decision mechanism is initially considered for identifying potentially hazardous conditions, while the architecture is designed to support future integration of machine learning methods for personalized anomaly detection and predictive safety monitoring. The study provides a technical evaluation of the proposed architecture, its components, communication structure, operational scenarios, and implementation limitations. As a conceptual design, the proposed system does not claim field-tested performance; rather, it provides a structured framework for future prototype development, laboratory evaluation, and deployment in underground mining environments.
Agriculture is increasingly expected to produce more food while minimizing the usage of water, chemical-based fertilizers, energy, chemicals, and labor. Agricultural decision-making is further complicated by climate unpredictability, soil degradation, shortages of water, crop illnesses, pests, weeds, and yield uncertainties. An IEEE-style review and implementation methodology for AI in agriculture is presented in this research. Examined are a number of important applications, including as crop recommendations, crop-yield estimation, disease and pest recognition, weed identification, soil and nutrient administration, smart irrigating, weather forecasting, livestock monitoring, agricultural robots, and post-harvest management. A multilayer AI-enabled farming system is proposed in which a conventional sensor, image processing, and historical information are acquired through IoT, UAV, satellites, and field devices; processed information is provided to machine-learning(ML)/deep-learning models; and the ensuing decisions are transmitted to farmers or automated actuators. Conventional models such as Decision Tree(DT), Random Forest, K-Nearest Neighbor(KNN), Support Vector Machine(SVM), and Artificial Neural Network(ANN) are contrasted conceptually with CNN, LSTM, and Transformer-based techniques.Substantial research gaps in dataset variety, model generalization and explanation, efficiency of computation, cybersecurity, field validation,, and cost for small-scale farmers are also identified in the report. For sustainable intelligent agriculture, a future research path based on explainable AI, edge intelligence, hybrid AI, and multimodal data fusion is suggested.
Fiber sensors encompass optical-fiber platforms and electrically functional fibers, yarns, and textiles that produce spectral, temporal, and spatial data across wearable, biomedical, robotic, infrastructure, and environmental applications. Their responses are shaped by nonlinear transduction, cross-sensitivity, conformal contact, manufacturing variation, drift, and limited calibration coverage. This review examines how machine learning (ML) can be used across the fiber-sensing pipeline, from algorithm-assisted physical design and fabrication-aware optimization to denoising, demodulation, representation learning, multimodal inference, and application-level decision making. Direct demonstrations in optical and electrically functional fiber systems are distinguished from transferable methods developed in adjacent photonic, additive manufacturing, tactile, and soft-sensor platforms. The available evidence is strong for calibration, signal recovery, and task inference within a defined device and operating domain. By contrast, experimental evidence for ML-guided physical co-design, cross-device and cross-batch transfer, long-term stability, and hardware-constrained deployment remains limited. Accordingly, ML is most promising as part of a coupled sensor–model–hardware system that preserves task-relevant information and reports when an input lies outside its validated operating range. This review therefore emphasizes validation boundaries over headline accuracy alone and identifies uncertainty quantification, physics-guided learning, fabrication metadata, and edge-aware co-design as priorities.
Automobile sales analysis provides an important basis for production planning, procurement, supply-chain coordination, inventory management, and dealership resource allocation. However, dealership sales are jointly influenced by operational practices, market conditions, and regional policies. Conventional machine-learning models may have difficulty capturing spatial heterogeneity, maintaining predictive performance under regional sample imbalance, and providing interpretable evidence for managerial analysis. To address these issues, this study constructs a multi-source automotive dealership dataset based on automaker marketing records and proposes CDMEP, a multi-paradigm framework integrating spatial diagnosis, ensemble transfer learning, predictive interpretation, and assumption-conditional structural diagnosis. The framework comprises a Spatial Heterogeneity Decoupling Diagnostic Model (SHDM) for examining spatial clustering and regional differences; a Regional Adaptive Stacked Learner (RASL) combining balanced cross-regional transfer, TPE-tuned base learners, and constrained ensemble weighting; and a Causally Scrutable Optimization Policy (CSOP) incorporating SHAP-, GAM-, and PDP-based predictive interpretation, temporally ordered LiNGAM diagnosis, edge-level stability and sensitivity analysis. The results indicate the following: (1) dealership sales exhibit spatial clustering and regional heterogeneity in variables related to lead generation, customer conversion, and dealership service capability; (2) RASL achieves an R2 of 0.875 on the independent test set, corresponding to an absolute improvement of 0.051 over the second-best benchmark in the present dataset; and (3) under the specified preprocessing procedures and LiNGAM assumptions, lead volume, customer traffic, and conversion rate repeatedly appear in candidate directional structures associated with next-month sales, although their estimated coefficients and diagnostic stability vary across regions. These findings provide preliminary evidence that CDMEP may serve as a useful analytical framework for regionally differentiated sales prediction and assumption-aware dealership analysis.
Assisted history matching is bottlenecked by the cost of the forward reservoir simulator: every candidate parameter vector requires a full flow simulation, and ensemble-based calibration needs hundreds to thousands of runs. The prevailing machine-learning proxy — gradient-boosted regression on a fixed vector of static uncertainty multipliers — captures smooth, monotonic responses such as field gas rate while falling short on the path-dependent water-breakthrough dynamics that drive history-matching decisions. We show this failure is diagnostic: on a held-out set of 160 unseen realizations, a tuned GBDT proxy attains a per-realization median R2 of 0.90 on field gas rate but −0.58 on field water rate — worse than predicting each realization's own mean — a shortfall hidden by aggregate metrics. We introduce a cell-resolution Graph Neural Network (GNN) proxy that represents every active simulator cell as a node carrying its petrophysical attributes and connects cells through physics-typed, transmissibility-weighted edges — Cartesian neighbors, explicit non-neighbor connections (NNCs), and fault-aware links — with learned edge-type embeddings and a transmissibility-gated multi-hop message-passing scheme. Cells are aggregated to wells through completion-aware hierarchical pooling, and an autoregressive cumulative-state head closes the temporal feedback loop. On the same held-out realizations the GNN recovers per-realization water skill to R2 = 0.68 (field-aggregated) and 0.74 (well-level) — an improvement of more than one R2 unit over the baseline on the quantity that actually gates calibration — while remaining competitive on gas. An autoregressive variant, evaluated closed-loop over the full trajectory, additionally captures cumulative water (WWPT) at R2 = 0.76 and oil/gas rates at R2 ~ 0.89 with no teacher forcing at inference. Training uses a teacher-forcing to scheduled-sampling curriculum with partial-horizon supervision that restricts learning to the early, information-rich production window and reduces the simulator runtime needed to assemble calibration ensembles. A leakage audit confirms the surrogate consumes no simulator output at inference — only the reservoir parameters under calibration, the prescribed well schedule, and a trivial field-at-rest initial condition — so it is a genuine in-the-loop replacement rather than a post-hoc fit. The result is a deployment-oriented surrogate that reduces simulator dependency exactly where tree-based proxies fail, and that lifts the parameter- space ceiling beyond a fixed set of static descriptors and, in future work, toward cross-reservoir transfer.
Substantial progress has been made in artificial emotion-recognition systems; however, reproducing neurotransmitter-mediated emotional processing remains a major challenge for neuromorphic computing. In this study, we report bimodally modulated artificial synapses based on a van der Waals-stacked graphene quantum dot (GQD)/MoS2 heterostructure that integrates dopamine and optical stimuli for emotion-inspired neuromorphic processing. Incorporation of GQDs increases the ON-state current and dynamic range and improves relearning behavior and synaptic weight modulation. The GQD/MoS2 heterostructure exhibits dopamine-dependent electrical responses at concentrations of 1-5 nM and stable synaptic plasticity with retention maintained over the 1000 s measurement period. Optical stimulation in the presence of dopamine produces postsynaptic current (PSC) increases that exceed the additive expectation calculated from the individual optical-only and dopamine-only responses, supporting synergistic enhancement under the tested conditions. Device-informed facial emotion recognition simulations achieved an accuracy of 73.4% on the FER-2013 dataset using a convolutional neural network (CNN). Separately, classification of PSC-derived temporal signals using electroencephalography (EEG)-inspired processing achieved an accuracy of 76.6% with a support vector machine (SVM). These findings highlight the potential of this approach for affective computing and human-machine interaction.
Recurrent memory requires learning useful updates, but attention added to assist learning can also compromise length generalization. We study bounded delta attention: causal attention over uncorrected update proposals whose bounded correction enters the recurrent update. Its implementation in TraceRelay combines additive phase recurrence with attention computed before a parallel scan. On prefix-supervised modulo-5, trained at lengths 1--40, three findings characterize its behavior at length 80. First, upper delta attention attains $70.81\pm8.26$% accuracy across seven seeds, versus $69.21\pm22.42$% for the local-attention recurrent baseline (mean $\pm$ sample SD): a 63.2% reduction in SD with a similar mean. Upper state attention also reduces variation, reaching $74.75\pm6.24$%, so this benefit is not unique to update-level attention. Disabling the upper delta correction retains $70.62\pm8.60$%. Second, blocking gradients at the delta-attention inputs changes accuracy to $65.25\pm12.40$% in the matched seven-seed comparison, increasing SD by 50.2%. It also delays a 99% validation threshold on average among the same six seeds that reach it; the 95% threshold is reached earlier on average with input-gradient blocking across all seven seeds. Third, lower delta attention exhibits substantially less extrapolation degradation than lower state attention: 62.60% versus 19.40%, against a 71.06% baseline. Removing its correction reduces accuracy to 22.56%; its advantage over lower state attention also persists under input-gradient blocking. These observations motivate bounded update-level attention as a recurrent learning scaffold, combining reduced seed variation with less severe degradation in a placement where state attention fails.
Visualizing multivariate scientific simulation data is essential for understanding complex physical phenomena, yet the massive scale of such data imposes a severe storage bottleneck in high-performance computing environments. Implicit neural representations (INRs) have emerged as a promising continuous alternative to traditional lossy compressors, but existing methods struggle to explicitly capture cross-variable interactions and to adapt to the diverse spectral characteristics of physical variables. To address these limitations, we propose SWAP-INR, an architecture for high-fidelity compression of multivariate, time-varying simulation data. SWAP-INR introduces two key components: (1) The switchable attention path (SWAP) module explicitly models cross-variable interactions through a cross-variable attention layer, then routes each variable to an attention or non-attention path by a deterministic, loss-driven criterion, and prunes the inactive layers for structural compression. (2) The adaptability amplification SIREN (A2-SIREN) replaces rigid frequency scaling with channel-wise learnable modulation, adapting to heterogeneous physical features with negligible overhead. Combined with post-training 16-bit quantization, SWAP-INR achieves state-of-the-art performance on small modular reactor (SMR) datasets and remains competitive on public physics simulation datasets, while maintaining a constrained memory footprint. Our code is https://github.com/hvcl/SWAP-INR.
Multi-agent LLM systems assign different responsibilities to different agents: one plans, another retrieves evidence, another verifies, another executes. This division of labor is only meaningful if each agent stays within the functions it is authorized to perform. Over many turns, however, an agent can produce a locally reasonable message that exercises authority belonging to another role: a verifier that begins planning, a retriever that asserts unsupported conclusions, an executor that changes state no one approved. We formalize this as role drift, a turn-level breach of an agent's role contract. We introduce Agent Trajectory Control (ATC), which makes each role's authority explicit as a contract and enforces it at the point where a message would enter the shared conversation. ATC scores a proposed output for out-of-role deviation, leakage into the functions of other roles, impermissible speech acts, and logical contract breaches, then computes a leakage-regularized target that constrains regeneration; the revised message is re-checked before admission. Enforcement acts on the message rather than the generator, so ATC governs agents whose internal representations are unavailable. Across homogeneous and heterogeneous teams, ATC improves task performance and role consistency while lowering role-drift energy relative to strong baselines spanning prompting, memory, workflow control, and activation editing, with the largest gains under explicit role-confusion stress tests and in larger, more diverse teams.
The ability of mechanical systems to perform basic computations has gained traction in recent years, providing a robust supplement to digital computing in off-grid, low-power, and extreme environments where electronics are inaccessible or risk failure. However, prior work has focused on Boolean logic applied via bistable structures, snap-through mechanisms, or origami lattices. Digital signal processing (DSP) underpins a wide range of computational tasks, providing the foundation for information encoding, manipulation, and analysis. In contrast to conventional DSP approaches, we present a new paradigm for analog information processing, where an elastic substrate houses a spatial wave-based platform for mechanical signal manipulation in the temporal domain. The proposed metasurface-based system exploits temperature-dependent phase transitions of shape memory alloys (SMAs) to realize key signal-processing functionalities, including gating, retention, and programmable delay of impinging mechanical loads. In doing so, we address a new set of challenges by manipulating transient mechanical waves through dispersive and lossy media. Building on this capability, we introduce the concept of mechanical wave-based memory, in which a locally resonant metamaterial's dispersion characteristics are actively reconfigured to achieve programmable delays while preserving spatial waveform integrity, and demonstrate its utility in the context of input shaping and frequency demultiplexing of superimposed waves. More broadly, the modular and scalable nature of the demonstrated framework paves the way for advanced mechanical computing architectures, enabling higher-order operations such as signal classification, parallel processing, and reservoir computing through the controlled sequencing and temporal alignment of mechanical signals.
A team serving a 9B-20B open-weight language model on vLLM or SGLang must choose a quantization recipe, and needs two numbers: the speedup over BF16 and the accuracy given up. We assemble every public measurement of these numbers that we could verify and find that the record supplies them only in pieces. Only two checkpoints, both INT4 builds of Gemma 3 12B, have both a vLLM timing and an accuracy score, and that score covers a single MMLU subject, one that the AWQ build was calibrated on. The batch-1 timings predate current engine releases, AMD and Intel Gaudi have no timing in this band at all, and FP4 has none against BF16. The timings that do exist follow one rule. Decoding a small batch is limited by memory bandwidth, so 4-bit weights roughly double speed (2.12x for Qwen2.5-14B on an A100), but the gain vanishes once a large batch makes the GPU compute-bound (1.02x for Gemma 3 12B at batch size 256), where FP8 is the recipe that can still help. Many published accuracy deltas, including every HumanEval delta, lie within benchmark sampling noise. We close with a seven-item reporting checklist for future measurements.
Post-training quantization (PTQ) for vision-language-action (VLA) models promises smaller and faster VLA deployments. However, when analyzing existing VLA quantization works, we realised that it is hard to reason about the performance of the suggested quantization method compared to existing, similar, or in many cases, simpler PTQ methods. We present LoopQ, a closed-loop benchmark that compares multiple quantization policies on the same checkpoint, layers and engine, varying only the method and the kernels. We evaluate RTN, AWQ, GPTQ, ActQuant and QuantVLA on four VLAs with more than 180,000 episodes of LIBERO and LIBERO-plus, compare against QVLA's reported results on the same layers, and benchmark whether the four public VLA-quantization releases can be reproduced. We benchmark the methods on an NVIDIA H100, an L4, and a Jetson AGX Thor. Our results show that while quantization reliably saves memory, 2-3x for real low-bit models, VLA-specific PTQ methods underperform compared to simpler generic baselines. At 4 bits, RTN, AWQ and GPTQ match or exceed ActQuant and QuantVLA, plain RTN matches the calibrated methods, and plain AWQ matches the success QVLA reports for itself. Standard LIBERO hides the losses that do exist: configurations with no significant LIBERO loss lose 1.5-4.3 points under paired LIBERO-plus perturbations. Speed is a weaker story. At batch 1 on the H100, no quantized pi0.5 or OpenVLA-OFT configuration is faster than its 16-bit model (1.06-1.58x slower on library kernels); only the L4 with int8 kernels gains (4-14%), the Thor is slower on every 4-bit configuration, and the engine alone changes latency by 2.1x. Profiling explains why: pi0.5's quantized backbone is compute-bound, while its memory-bound action expert, 63% of GPU time and 13% of the parameters, stays in 16 bits.