On-Device Inference

Embedded devices from wildlife monitoring stations to clinical wearables require local AI inference due to latency, communication, or privacy constraints. Optimizing models for heterogeneous microcontrollers (MCUs) requires simultaneously satisfying hard physical constraints on memory, power, and temperature while preserving accuracy, a multidimensional optimization that is today performed manually by experts. We ask whether an LLM agent can autonomously navigate this complex, multi-turn pipeline guided by real hardware feedback, and introduce a hardware-in-the-loop agent arena in which the agent iteratively refines both model and firmware -- compiling, flashing, and measuring on real hardware -- to enable closed-loop optimization. Frontier models, including Claude Opus 4.7 and Gemini 3.1 Pro, fail entirely without hardware feedback (0% deployment success), whereas our hardware-in-the-loop formulation achieves the first successful deployment within three iterations and can surpass human expert results within seven. This agentic co-optimization achieves 250x compression for vision models with <3.3% accuracy loss and 400x for audio with <6% Feature Error Rate loss, enabling battery-free operation on a commercial MCU via solar harvesting. We demonstrate practical impact in two real-world systems: an elk-detection camera trap (96.7% accuracy) and a phonetic-transcription wearable (8.44% FER) for child development research.

Ultrasound imaging is the most widely adopted medical modality globally due to its low cost and portability, yet artificial intelligence (AI) deployment remains constrained by reliance on GPU-accelerated models, creating a structural paradox where the cost of "intelligence" exceeds that of the imaging device itself. Here, we present the systematic adaptation and extensive evaluation of UltraSeg, an ultra-lightweight architecture originally developed for colonoscopic polyp segmentation, now engineered for point-of-care ultrasound (POCUS) across ten public datasets spanning six anatomical sites (breast, thyroid, kidney, carotid, fetal, and small-animal tumor). We systematically validate both variants in ultrasound domains: UltraSeg-130K (0.13M parameters) achieves 89.7 FPS on single-core CPUs and 34.8 FPS on a refurbished mobile device, while UltraSeg-500K (0.5M parameters) delivers 44.6 FPS on CPU and 16.1 FPS on mobile device. UltraSeg-500K matches or exceeds the Dice performance of the 31M-parameter UNet and approaches 105M-parameter TransUNet in average performance, with superior zero-shot cross-dataset generalization on external validation sets (UDIAT, DDTI). By enabling clinical-grade segmentation without GPU dependency, this work brings AI costs in line with ultrasound accessibility, making advanced diagnostics available in resource-limited settings.

Deep learning has become the dominant paradigm in Wearable Human Activity Recognition (WHAR), yet progress is obscured by a comparability crisis. Results are often reported using inconsistent datasets, custom data processing, and varying evaluation protocols, making state-of-the-art claims fragile. We address this with a large-scale, open-source benchmark that integrates 30 diverse datasets under standardized processing, unified model interfaces, and a shared cross-subject evaluation protocol. Evaluating 17 representative architectures across 4760 training runs, we jointly measure predictive performance alongside on-device latency, peak memory, and model size on an Android reference device. Our results reveal that the WHAR state of the art is distributed rather than dominated by a single architecture. While CNN-HAR achieves the highest mean macro-F1, top-performing models cluster tightly, indicating contemporary architectures have converged near a predictive performance ceiling. When accounting for deployment efficiency, compact neural models, such as TinierHAR, and classical Random Forests define the practically relevant Pareto frontier, whereas larger recurrent and hybrid models incur high hardware costs without corresponding performance gains. Consequently, while predictive performance has plateaued, substantial potential for future progress remains in optimizing deployment efficiency and improving adaptation to domain shifts. We release our full framework to support transparent reuse and extension.

Background and Objective: Falls among elderly people can cause serious injury and reduce quality of life. Timely prediction and detection are essential to prevent harm and support well-being. We propose a portable, low-power, battery-operated, vision-based fall prediction and detection system using HPE on an AMD Kria K26 System-on-Module (SOM). The objective is a non-intrusive, privacy-preserving system for real-time fall detection. Methods: The system uses an Intel RealSense D455 range-sensing camera connected to the K26 SOM by USB. It captures synchronized RGB and depth frames, 640 x 480 x 3 and 640 x 480 pixels, at 60 FPS. The SOM runs a three-stage pipeline with quantized YOLOX, Anchor-to-Joint (A2J), and fall-detection models. YOLOX identifies human bounding boxes from RGB frames, then discards the RGB frames to preserve privacy. A2J uses depth frames to estimate 15 joint keypoints per person. A CNN uses selected joint coordinates (x, y, z) to classify fall activity. YOLOX was trained on CrowdHuman; A2J on ITOP, MP-3DHP, UR Fall Detection, and a custom SDSU PSG dataset; and the CNN on UR Fall Detection and SDSU PSG. The design used a single-core DPU with a serial pipeline and a dual-core DPU running YOLOX and A2J with multiple threads. Results: Quantized accuracy was evaluated using IoU >= 50% for YOLOX, mAP with a 10-cm rule for A2J, and classification accuracy, (TP + TN)/(TP + TN + FP + FN), for the CNN. Accuracies were 74%, 84.13%, and 75.85%. Throughput improved from 2.5 FPS for the single-threaded pipeline to 4.5 FPS for the multi-threaded version. Conclusion: Results demonstrate the feasibility of privacy-preserving fall detection on an AMD Kria K26 edge device. On-device HPE and fall classification runs without cloud dependency, supporting elderly monitoring and assistive healthcare. Future work will improve model accuracy and speed.

As AI systems move into operating systems, privacy no longer turns only on whether a model runs locally. A local assistant may assemble email, calendar entries, files, screenshots, notifications, and app intents; retain embeddings or summaries; invoke tools; emit telemetry; or route difficult requests to cloud infrastructure. Local inference reduces some exposure, but it answers only one question: where computation occurs. It does not answer who may assemble context, what derived state persists, which actions are authorized, or how updates change the system's authority. We develop an OS-centered privacy framework for on-device AI that treats privacy as an institutional accountability problem rather than a deployment attribute. The framework specifies a threat model, a six-part privacy risk taxonomy, privacy-by-architecture controls, and a four-level audit rubric. We demonstrate the rubric through a documentation-bounded comparison of Apple Intelligence/Foundation Models, Android AICore/Gemini Nano, and Microsoft Recall. Meaningful privacy in on-device AI depends on constrained information flow, bounded authority, visible user control, and auditable governance across the operating-system lifecycle.

Speech applications such as meeting transcription and voice agents would benefit from on-device speaker diarization, but practical adoption is limited by inference cost. We study how far a Pyannote 3.1-based pipeline can be accelerated on consumer hardware (an RTX 5070 Ti GPU and an Apple M4 laptop) while preserving diarization error rate (DER). A simple recipe: coarser segmentation stride and per-chunk embedding, yields multi-fold speedups and is DER-neutral on AMI, but degrades sharply on in-the-wild data: on VoxConverse, DER rises from 0.075 to 0.113. We trace the failure to speaker under-counting in the clustering stage, caused by a fixed minimum cluster size interacting with the reduced number of embeddings per speaker. We propose a relative minimum cluster size, mcs = round(f * n) with f = 0.01, which adapts to the embedding budget per recording. A single value of f recovers VoxConverse DER to 0.079 (about 89% of the lost accuracy) while keeping AMI flat, and the accelerated pipeline reaches up to 12.2x speedup on AMI (MPS) over our CAM++ baseline.

Smart eyewear enables unobtrusive, context-aware interaction through multimodal sensors and on-device intelligence, but is severely limited by power, memory, and compute constraints in a compact form factor. Open-hardware platforms supporting event-based vision and embedded ML at this scale are rare. This work introduces an open-source smart glasses platform for rapid prototyping of novel sensors and algorithms. Its modular design uses a flexible FPC interposer to support both event-based and frame-based cameras without full PCB redesign. A hardware-software co-designed power management system combines a configurable PMIC with event-driven wake-up via an nRF5340 coordinator, keeping the GAP9 RISC-V SoC powered down between inferences. The prototype achieves up to 11.5 hours of continuous on-device ML from a 200 mAh battery. As a demonstration, an egocentric hand gesture recognition pipeline was evaluated on the LynX dataset using polarity-separated event histograms from a Prophesee GENX320 camera. R(2+1)D achieved the best cross-subject accuracy of 83.94% (macro F1 = 0.781) under leave-two-subjects-out validation, with 78.3 ms end-to-end inference latency on the GAP9. Temporal augmentation and removal of ambiguous classes provided the largest gains (+8.9 pp). All hardware designs, firmware, and models are released open source.

Blind face restoration on consumer devices has to balance image quality against speed and memory. Strong methods such as GFPGAN and CodeFormer give good perceptual quality, but they rely on large pretrained generative priors and on operators such as attention, codebook lookup, and style modulation that are hard to compile and quantize on the small neural processing units (NPUs) used in consumer hardware. Small convolutional restorers run fast enough, but they tend to over-smooth and to leave artifacts around the eyes, nose, and mouth. We present CFRNet, a 2.0,M-parameter ResNet-style restorer for on-device use at $256\times256$, the common face-crop size on consumer NPUs. The main idea is Cycle-Consistent Fixed-Point Training (CCFP). Instead of training the network for one pass and then running it several times by hand, we train it to act as a fixed-point operator, so that applying it again to a restored face does not change the face. CCFP uses three training losses, namely progressive multi-cycle supervision, an idempotence loss, and a re-degradation cycle loss, and it adds no cost at inference. To compare fairly under our deployment limits, we retrain all baselines from scratch at the same $256\times256$ resolution. On a 300-image test set, CFRNet reaches the best perceptual score (LPIPS 0.250 at three cycles, which is 31% lower than one cycle) and also the best PSNR and SSIM at two cycles. It runs in about 23,ms per cycle in INT8 on a HiSilicon Hi3402 NPU, while the same baselines cannot be compiled to that chip. The cycle count $k$ acts as a simple quality knob that needs no retraining: PSNR is best at $k\!=\!2$ and LPIPS keeps improving up to $k\!=\!3$. We further show that the same idea works with a plain CNN that is even easier to deploy, and we run the model in real time on an in-car driver-monitoring board.

Autonomous spacecraft require rapid, lightweight, and reliable onboard detection of cyber-RF threats. Using the SPARTA attack model, we analyze the latency-accuracy trade-offs of TinyML-compatible classical models -- Random Forest, Logistic Regression, SVM, and MLP -- for detecting uplink jamming, Fake-NR spoofing, payload manipulation, ground-segment compromise, and unauthorized command injection. We present a physics-informed theoretical analysis of each model's computational complexity, VC dimension, Lipschitz continuity, and latency scaling, supported by empirical measurements on adversarial RF spectrograms generated via BandErasure, FakeNR, and NoiseBurst corruption modes. Results show that Logistic Regression achieves microsecond-level inference with only a 1% accuracy drop relative to Random Forest, making it an effective TinyML baseline for onboard autonomy. The study also identifies opportunities for advancing spacecraft cybersecurity through richer feature encoders and multi-timescale learning architectures, building on recent progress in edge intelligence and trustworthy AI.

Sensor-based Human Activity Recognition (HAR) models often degrade on unseen users due to domain shifts caused by individual movement patterns and sensor placement. Practical wearable HAR systems therefore require personalization methods that are lightweight, applicable across diverse calibration scenarios, and robust under limited calibration data. We present BayaHAR, a gradient-free framework that repurposes pretrained HAR classifiers as Prototypical Networks using prior prototypes that preserve zero-shot performance while regularizing adaptation. For labeled calibration data, we introduce closed-form Bayesian prototype estimation and extend the same principle to weakly labeled data, requiring only knowledge of which activities were performed. With only 3 seconds of calibration (one shot) per activity, supervised adaptation improves test macro-F1 on unseen users by +2.76 to +33.44 percentage points across four datasets, while weakly supervised adaptation improves by +0.56 to +32.13 points. Since adaptation requires only closed-form prototype updates, the framework enables efficient and robust on-device personalization of preexisting HAR classifiers.

Hearing aids impose strict latency and power constraints that current DNN-based speech enhancement systems struggle to meet on embedded hardware. We characterize this gap by deploying both speech separation and denoising using the lightweight SuDoRM-RF++ architecture on the AMD-Xilinx Kria KV260, evaluated at FP32 and 16-bit fixed-point precision for each task. Across these configurations, first-sample latency tracks with on-chip parameter caching rather than arithmetic throughput, identifying data movement as the primary bottleneck. Precision reduction halves the model memory footprint without compromising objective speech quality. The fixed-point denoising accelerator achieves a first-sample latency of 9.7ms, meeting the 10ms clinical threshold, while speech separation reaches 16.0~ms. These measurements establish concrete resource requirements for embedded DNN-based speech enhancement and quantify the remaining gap to hearing aid deployment.

Our work presents a method for ECG segmentation and arrhythmia detection using Tiny Machine Learning (TinyML) models for real-time, on-device inference on resource-constrained embedded systems. We develop INT8 quantized autoencoder-based TinyML models with minimal layers and parameters for embedded deployment. These models are evaluated using a custom dataset derived from the MIT-BIH Arrhythmia Database and validated in both PC-based simulations and on-device environments. For the evaluations, over 95,000 ECG segments are processed on an ESP32-S3 microcontroller running the TensorFlow Lite Micro runtime. Post-evaluation, detailed analysis, including annotation-wise and record-wise failure analysis, is conducted to characterize model behavior across diverse ECG morphologies and rhythm patterns and to explain missed detections. In several cases, apparent misclassifications may correspond to early or subtle anomaly patterns labeled as normal in the reference annotations, highlighting the model's sensitivity. A refined evaluation by filtering out ambiguous cases in the dataset shows that the best-performing DNN-based autoencoder achieves a recall of 84%, an F1-score of 79%, a model size of approximately 180 KB, and an inference latency of 9 ms on-device. These results demonstrate the feasibility of low-power, privacy-preserving embedded wearable systems capable of performing accurate arrhythmia detection entirely on-device.

The rapid growth of the satellite industry has driven a significant increase in geospatial data acquisition, highlighting a critical bottleneck: the severe disparity between the volume of collected sensor data and the limited downlink bandwidth available to ground stations. While On-Board Computing (OBC) has helped address this by pre-processing data in orbit, this article further advances the paradigm by introducing an in-sensor computing framework. We present an optimized end-to-end Earth Observation (EO) pipeline tailored for strict computational constraints by integrating TinyML techniques with the Sony IMX500 Intelligent Vision Sensor. Specifically, our approach shifts processing directly to the sensor level, offloading the computation from the primary embedded device, and effectively mitigating the downlink transmission of noisy or irrelevant data. We evaluated several efficient Convolutional Neural Networks (ConvNets), i.e., SqueezeNet, ShuffleNetV2, and MCUNetV1, on the EuroSAT dataset. Experimental results show that, despite the optimizations required for deployment on the IMX500 platform, our models maintain a competitive 96.68% accuracy while operating within its 8 MB constraints. Specifically, the models reach an average processing throughput of 17.40 FPS with a latency of 27.43 ms. Furthermore, our system profile exhibits high energy efficiency, with a low energy footprint of 14.19 mJ per inference and an efficiency rating of 42.26 GMAC/J, demonstrating its viability for in-sensor deployment.

Reasoning-based robotic policies using large language and vision-language models achieve strong semantic planning capabilities but mostly suffer from a high inference latency that limits practical real-time deployment. In this work, we observe that robotic reasoning workloads contain substantial temporal redundancy, where consecutive observations frequently produce identical actions and subgoals. Based on this insight, we present REIS, a human cognition inspired robotic decision-making framework that minimizes unnecessary reasoning while preserving semantic adaptability. REIS combines lightweight scene gating, KV-steered affordance routing, and deliberative reasoning to accelerate robotic control under embodied constraints. Experiments on ALFRED, and real-world robotic tasks demonstrate that REIS significantly suppresses reasoning overhead while maintaining competitive task performance.

Trajectory prediction is a fundamental task for autonomous systems, requiring complex reasoning about multi-agent interactions and intents. Large language models (LLMs) have recently been adopted for this task, as they provide strong contextual reasoning and interpretable, language-based trajectory representations. However, these LLM-based predictors are extremely memory- and compute-intensive, making them difficult to deploy on resource-constrained edge devices such as on-board computers in autonomous robots. To bridge this gap, we propose BitTP, which converts an LLM-based trajectory predictor into a lightweight bitlinear architecture. We demonstrate that weight-only quantization to 1.58-bit (BitTP-Weight) is optimal. Crucially, activations must remain in full precision, as quantizing them leads to severe degradation and instability in spatio-temporal reasoning. Empirically, BitTP-Weight not only preserves but improves prediction quality over the full-precision (BF16) LLM baseline, reducing ADE by 14.29% and FDE by 20.97% on average, while simultaneously reducing memory usage and inference latency relative to other quantization methods. These results demonstrate that carefully designed quantization acts as an effective regularizer, enabling the practical deployment of sophisticated LLM-based reasoning on edge devices. Code is available at: https://github.com/MintCat98/BitTP.

Evaluations of dental vision-language models remain fragmented across datasets, task definitions and metrics, and often ignore their computational cost. This limits their widespread deployment for dental screening outside specialist centres, where timely inference, limited hardware, and local handling of patient images are vital for practical, privacy-preserving clinical prescreening. Here we present Pocket-Dentist, an efficiency-aware benchmark for dental multimodal question answering that brings together three datasets spanning approximately 1,159 patients from BRAR and MetaDent, five task types and seven metrics. Across 14 typical VLMs, our results reveal an interesting observation: compact VLMs, such as 2B-parameter models, become competitive with much larger VLMs on most metrics after lightweight adaptation while requiring substantially lower computational costs in dental image understanding. Deployed locally on an iPhone 17 Pro, our finetuned compact VLM Pocket-Dentist-2B processed each sample in 4.31 s, reducing latency by 4.9x and memory use by 2.3x compared with a 7B baseline. Our project page is available at https://2026-icml.github.io/pocket-dentist-icml.

The remarkable generation quality of modern diffusion models often comes at the cost of massive parameter counts, which necessitate server-side inference with significant computational costs and potential privacy risks. Consequently, there is growing momentum toward developing efficient on-device alternatives. While recent efforts have optimized text-to-image models for mobile hardware, they remain relatively bulky, typically ranging from 0.5B to 1B parameters. We present BlazeEdit, a highly efficient, generalist image-to-image diffusion model tailored for on-device deployment. By identifying that many practical image editing tasks do not require text-based guidance, we eliminate the text-conditioning components and develop a multi-task architecture that consolidates object removal, outpainting, tone correction, relighting, and sticker generation into a single, compact model of only 195M parameters. BlazeEdit achieves a substantial reduction in download size and memory overhead while maintaining competitive generation quality. It completes a full inference pass in just 290ms on a Pixel 10, delivering a seamless, privacy-preserving, and lightning-fast experience for generalist image editing on the edge.

Mobile graphical user interface (GUI) agents enable AI models to autonomously operate smartphones on behalf of users. However, most existing systems focus primarily on optimizing task accuracy and rely on cloud-hosted models for inference, which introduces privacy concerns and network-dependent latency. As a result, fully on-device deployment of mobile GUI agents remains underexplored. We propose MobileExplorer, a new framework that accelerates on-device inference for vision-based mobile GUI agents via online exploration. The key idea is to exploit the long per-step reasoning time of vision-language models (VLMs) by performing lightweight, parallel exploration of UI elements. During model inference, the agent proactively probes semantically relevant UI elements and records these exploration traces as structured memory. To ensure reliable execution in live mobile environments, we design a two-level rollback mechanism that robustly restores the initial UI state when a fast but naive backtracking strategy fails. The collected exploration traces are then summarized into concise contextual hints and injected into the prompt to enhance the subsequent reasoning step. We evaluate MobileExplorer on multiple off-the-shelf devices using the AndroidWorld benchmark, as well as newly designed, more complex tasks and dynamic on-device environments. MobileExplorer reduces the average number of reasoning steps and end-to-end latency by 23%, while maintaining or improving task success rates by up to 5%. A video demonstration of MobileExplorer performance in the real world is available at https://youtu.be/thK7MJmdlvM .

Motion capture is the gold standard for measuring human movement, but clinical use remains limited by cost, technical complexity, and privacy concerns. AIGaitor is a privacy-preserving, cloud-free motion analysis system that runs markerless monocular motion-capture pipelines and downstream deep-learning analysis entirely on a consumer smartphone using on-device neural accelerators. To motivate its design, we surveyed 74 rehabilitation clinicians: 92 percent said they would adopt an accurate, cost-effective, easy-to-use AI gait analysis tool, while 79.7 percent cited operating cost, 68.9 percent insufficient training, and 64.9 percent privacy concerns as leading barriers. We then optimized and benchmarked mobile iOS implementations of current monocular pipeline components, including 2D and 3D pose estimation, pose optimization, skeleton-based deep-learning analysis, and a vision-language model. A Time-Priority end-to-end on-device pipeline processes a 10 s 4K 60 fps video clip in 77 s on an iPhone 14, matching or beating the same pipeline on a high-end NVIDIA H200 cloud server when network transfer is included: 94 s at global mobile-average uplink and 66 s at developed-world Wi-Fi. Lightweight models such as ViTPose-s achieve real-time keypoint extraction, and skeleton-based action-recognition models provide sub-millisecond gait classification on the same clip. To our knowledge, AIGaitor is the first monocular system to demonstrate end-to-end on-device motion capture and downstream deep-learning analysis, supporting clinically applicable movement analysis that is low-cost, private, and accessible to smartphone users.

Deploying large deep neural networks on memory-constrained mobile devices is a central challenge in edge ML. While compression, pruning, and quantization reduce per-parameter cost, transformer-based models remain too large for the 3.3-7.4 GB RAM envelope of commodity Android handsets. We present the DNN pipeline scheduling subsystem of CROWDio, which achieves practical ONNX inference across resource-constrained Android workers without model modification, by distributing memory pressure across devices via five mechanisms: JIT deferred partition loading, a single-partition-resident constraint, a 4-tier affinity scheduler, a zlib-compressed tensor transport, and a streaming 1:1 dependency model. Evaluated on DistilBERT (Sanh et al., 2019) (approximately 67 M parameters, SST-2) across five Android handsets over ten runs, our system holds peak per-device RSS to 43+-2 MB and limits battery draw to 50+-3 mAh per run, while streaming concurrency cuts batch latency 34% below barrier synchronisation.

Deep learning models detect pneumonia from chest X-rays with high accuracy, but the performance declines under domain shifts caused by differences in devices, patients, or institutions. We present PneumoNet, a domain-incremental learning method for point-of-care pneumonia diagnosis in resource-limited settings. PneumoNet combines a lightweight CNN for on-device prediction, a dual-stage balanced buffer for class-balanced replay, and a dynamic class-weighted loss to correct training-batch imbalances. Evaluated on a domain-shifted PneumoniaMNIST dataset simulating five realistic domain change scenarios, PneumoNet achieves 86.6% accuracy with 1.4% forgetting while being smaller and faster than existing baselines. These results highlight PneumoNet's potential to enable adaptive, privacy-preserving diagnostic AI directly on point-of-care medical devices in real-world and pandemic-ready healthcare.

Deploying adaptive intelligence at the edge remains challenging due to the high computational and energy cost of training neural models. Spiking Neural Networks (SNNs) offer a promising alternative, but enabling on-device learning requires hardware-algorithm co-design. This paper presents SPIKER-LL, an FPGA-based SNN accelerator that extends the open-source Spiker+ inference architecture with efficient support for the STSF local learning rule. Through targeted microarchitectural extensions, SPIKER-LL performs inference and online learning with minimal overhead. Across MNIST, F-MNIST, and DIGITS, it achieves up to 93% accuracy, sub-millisecond latency, and less than 0.1 mJ per inference, while remaining DSP-free and highly scalable for edge-FPGA deployments.

Transformer models are rapidly becoming a cornerstone of modern Internet of Things (IoT) applications, yet their computational and memory demands far exceed the capabilities of a single typical ultra-low-power IoT device. We present CATS, a framework for distributed transformer inference on ultra-low-power wireless devices, enabling multiple devices to collaboratively execute models far larger than what a single device can sustain. At its core, CATS is a communication-aware distributed transformer inference scheme co-designed across transformer partitioning, wireless communication and training. It employs SomeGather, a new pruned communication primitive that selectively broadcasts activation columns to reduce communication bandwidth and RAM usage without sacrificing model accuracy. Building on SomeGather, we design a partitioning method that exploits this primitive for efficient model parallelism. To cope with unreliable wireless communication, CATS employs message-dropout during training, which mimics packet losses and yields models that are robust to message loss during inference. In real-world experiments, we show that CATS brings distributed transformer inference to ultra-low-power wireless devices for the first time, with deployments on up to 16 devices that collaboratively execute transformer models up to 14 times larger than what a single device can run.

The Diffusion Transformer (DiT) architecture is the state-of-the-art paradigm for high-fidelity image generation, underpinning models like Stable Diffusion-3 and FLUX.1. However, deploying these models on resource-constrained mobile devices entails prohibitive computational and memory overhead. While efficiency-driven approaches like Linear-DiT and static pruning alleviate bottlenecks, they often incur quality degradation. Unlike cloud environments, mobile constraints require a single-model paradigm that dynamically balances fidelity and latency. We introduce ElasticDiT, which achieves this dynamic trade-off by adjusting spatial compression ratios and DiT block depths. By integrating Shift Sparse Block Attention (SSBA) and a Tiny DWT-Distilled VAE (T-DVAE), ElasticDiT reduces inference latency and memory footprint while maintaining image quality. Experiments confirm that ElasticDiT effectively covers a wide range of fidelity-latency trade-offs within a single set of parameters. By jointly adjusting compression and depth, a single ElasticDiT model can be reconfigured on-the-fly to outperform task-specific baselines. Specifically, our flex lite variant achieves an HPS of 32.87, surpassing the Flux model, while maintaining competitive quality at 84.16 percent average sparsity through SSBA. Furthermore, the plug-and-play T-DVAE provides SD3-level reconstruction with only 1/8x the computational cost of standard VAEs, and Flow-GRPO boosts semantic alignment (GenEval: 66.93 to 73.62). These results demonstrate that ElasticDiT offers a versatile, hardware-adaptive solution that eliminates the need for multiple specialized models, providing a promising path for future high-resolution image generation on mobile devices.

Diabetic Retinopathy (DR) is one of the leading causes of preventable blindness, yet rural regions often lack the specialists and infrastructure needed for early detection. Although cloud-based deep learning systems offer high accuracy, they face significant challenges in these settings due to high latency, limited bandwidth, and high data transmission costs. To address these challenges, we propose a two-tier edge-cloud cascade on the public APTOS 2019 Blindness Detection dataset. Tier 1 runs a lightweight MobileNetV3-small model on a local clinic device to perform a binary triage between Referable DR (Classes 2-4) and Non-referable DR (Classes 0-1). Tier 2 runs a RETFoundDINOv2 model in the cloud for ordinal severity grading, but only on the subset of images flagged as referable by Tier 1. On a stratified APTOS test split of 733 images, Tier 1 reaches 98.99% sensitivity and 84.37% specificity at a validation-tuned high-sensitivity threshold. The default cascade forwards 49.52% of test images to Tier 2, reducing cloud calls by 50.48% relative to using a cloud-based model for all images. In the deployed 4-class output space (Class 0-1 / Class 2 / Class 3 / Class 4), the cascade obtains 80.49% accuracy and 0.8167 quadratic weighted kappa; the cloud-only baseline obtains 80.76% accuracy and 0.8184 quadratic weighted kappa. On APTOS, the cascade cuts cloud use by about half with a modest drop in grading performance. Index Terms: Diabetic Retinopathy, Edge-Cloud Cascade, MobileNetV3-small, RETFound-DINOv2, Retinal Screening, tele-ophthalmology

Contextual bandits (CB) are online sequential decision-making problems under partial feedback that underpin many adaptive services. There is a growing demand to deploy CB agents directly on-device, under strict constraints on memory, compute, and energy. However, standard linear CB algorithms are often impractical for resource-constrained devices with their unfavorable scaling in computational and memory costs. Recently, HD-CB, a CB approach based on hyperdimensional computing principles, has been proposed to model and solve CB problems by moving into high-dimensional spaces. HD-CB offers faster convergence, favorable scalability, and improves memory efficiency compared to linear CB algorithms. However, its learning rule is accumulation-based: the values of action vectors grow over time, requiring high precision. While periodic binarization can prevent overflow in low-precision components, it may discard important information about magnitudes and degrade decision quality. This paper introduces probabilistic HD-CB, a low-precision variant that replaces deterministic accumulation with a probabilistic update rule. At each step, only a random subset of vector components is updated, with a time-decaying update probability, and component values are constrained to a predefined range [-k,+k]. This approach enables low-precision components, prevents overflow without periodic binarization, and reduces the expected update cost in proportion to the fraction of updated components. Off-policy evaluation on standardized synthetic CB benchmarks using the Open Bandit Pipeline shows that probabilistic HD-CB consistently outperforms binarized HD-CB at equal precision, while approaching the performance of HD-CB with as few as 3 bits per component.

While Mixture-of-Experts (MoE) scales model capacity without proportionally increasing computation, its massive total parameter footprint creates significant storage and memory-access bottlenecks, which hinder efficient end-side deployment that simultaneously requires high performance, low computational cost, and small storage overhead. To achieve these properties, we present DECO, a sparse MoE architecture designed to match the performance of dense Transformers under identical total parameter budgets and training tokens. DECO utilizes the differentiable and flexible ReLU-based routing enhanced by learnable expert-wise scaling, which adaptively balances the contributions of routed and shared experts. Furthermore, we introduce NormSiLU, an activation function that normalizes inputs prior to SiLU operators, producing a more stable trend of routed-expert activation ratio and a higher intrinsic sparsity level. We also identify an empirical advantage in using non-gated MLP experts with ReLU-based routing, indicating the possibility of MoE architecture simplification. Experiments demonstrate that DECO, activating only 20% of routed experts, matches dense performance and outperforms established MoE baselines. Our specialized acceleration kernel delivers a 2.93$\times$ speedup on Jetson AGX Orin compared with dense inference. Code and checkpoints are available at https://github.com/thunlp/DECO.

Running deep neural networks on microcontroller units (MCUs) is severely constrained by limited memory resources. While TinyML techniques reduce model size and computation, they often fail in practice due to excessive peak Random Access Memory (RAM) usage during inference, dominated by intermediate activations. As a result, many models remain infeasible on standalone MCUs. In this work, we present a fine-grained split inference system for networked MCUs that enables collaborative inference of Convolutional Neural Networks (CNN) models across multiple devices. Our key insight is that breaking the memory bottleneck requires splitting inference at sub-layer granularity rather than at layer boundaries. We reinterpret pre-trained models to enable kernel-wise and neuron-wise partitioning, and distribute both model parameters and intermediate activations across multiple MCUs. A lightweight, resource-aware coordinator orchestrates the inference across MCU devices with heterogeneous resources. We implement the proposed system on a real testbed and evaluate it on up to 8 MCUs using MobileNetV2, a representative CNN model. Our experimental results show that CNN models infeasible on a single MCU can be executed across networked MCUs, reducing the per-MCU peak RAM usage while maintaining the practical end-to-end inference latency. All the source code of this work can be found here: https://github.com/shashsuresh/split-inference-on-MCUs.

Mobile devices face diverse resource constraints and non-IID data class distributions, requiring fast on-device inference for local in-distribution (ID) classes and on-demand remote support for client-specific out-of-distribution (OOD) classes. Hybrid split federated learning (Hybrid SFL) couples personalized client-side front ends (supporting early exit) with a generalized server-side backend for fallback inference, balancing accuracy and cost. However, under client architectural heterogeneity, the existing hybrid SFL suffers from representation skew, where features from customized extractors fail to align in the shared space, leading to a sharp degradation in the server model responsible for OOD prediction. We propose HARMONY, the first hybrid SFL framework to support heterogeneous client architectures. HARMONY modifies meta-learning to simulate diverse extractors across parameters and architectures, and to learn to personalize. To mitigate representation skew, HARMONY conducts server-side contrastive learning to align extracted features, neither sacrificing clients' personalization nor sharing raw labels. Compared to the state of the art across multiple datasets and model families, HARMONY improves test accuracy by up to 43.0%/28.3% without/with OOD, respectively, while maintaining acceptable latency.

We present a deployed system for on-orbit wildfire detection aboard a nine-satellite commercial thermal infrared constellation, operating under demanding joint constraints: sub-megabyte model footprint, sub-150 ms per-batch TensorRT FP16 inference on an NVIDIA Jetson Xavier NX, and an end-to-end alert pipeline targeting under 10 minutes from satellite overpass to fire event communication. The system operates on uncalibrated mid-wave infrared (MWIR) single-band imagery at 200 m ground sampling distance, where fires frequently appear as sub-pixel or single-pixel thermal anomalies under extreme class imbalance -- challenges not addressed by the contextual thermal-thresholding pipelines (MODIS, VIIRS) that currently dominate operational fire monitoring. We present an empirical study of lightweight dense representation learning for this regime using a proprietary nine-satellite MWIR dataset. We compare dense masked autoencoding (DenseMAE) and a hybrid DenseMAE+EMA (exponential moving average) distillation variant, and evaluate representations via linear probing and full-distribution pixel-level average precision (AP) under extreme class imbalance. DenseMAE pretraining enables compact downstream models on the latency-accuracy Pareto frontier: our fastest SSL-pretrained model achieves 0.640 test AP and 0.69 event-level Fire-F1 with 65.34 ms latency per batch and a 0.52 MB engine, without pruning or compression. The best configuration reaches 0.699 AP and 0.744 Fire-F1 below 1 MB, outperforming a supervised baseline (0.650 AP) under comparable constraints.