edge device developers

Abstract: Edge applications increasingly combine CNN-based perception with edge-deployable Transformer-based language models, but deploying such heterogeneous workloads on resource-constrained devices is challenging because inference latency varies with both input difficulty and runtime hardware state. Early-exit inference reduces computation by terminating at intermediate stages, yet existing methods are often model-specific and rely on static latency profiles that become stale under dynamic edge conditions. This paper presents UniExit, a unified latency-aware early-exit runtime for heterogeneous edge AI. UniExit abstracts CNN blocks and Transformer layers as ordered stages and applies a common latency-aware routing policy while retaining model-specific quality mechanisms. It predicts the cost of reaching the next exit from runtime state and inference progress to make budget-aware continue-or-exit decisions. UniExit combines runtime-state-aware latency prediction, cache-based prediction reuse, trend-aware fallback, and CNN-specific exit stabilization to provide low-overhead and robust online control. We implement UniExit on multiple edge and edge-cloud platforms and evaluate it with representative CNN and Transformer-based language-model workloads. Across single-workload, concurrent heterogeneous, and edge-cloud settings, UniExit consistently improves deadline satisfaction and latency-accuracy trade-offs over existing progressive inference baselines, while reducing redundant exit-module computation and task-level energy consumption.

Abstract: We perform a novel study of CPU micro-architectural characteristics on how they affect agentic AI CPU side latency. Small enterprises and individual users looking to reduce costs and seeking data privacy benefit from running agents locally, motivating an investigation into the optimal performance characteristics of an edge CPU to improve user experience. We investigate the relationship between CPU core frequency, CPU uncore frequency, active physical cores, and symmetric multi threading with agentic task latency in single and multi agent configurations. Across eight experiments on an Intel edge CPU, we find that P-core frequency produces the largest median latency improvement among the evaluated parameters. Increasing P-core frequency from 2.5 to 5.5 GHz yields a 1.154$\times$ speedup for a single agent and a 1.341$\times$ speedup for eight concurrent agents. We are the first to characterize hardware micro-architectural relationships on single and multi agent configurations with TerminalBench; we also provide an open source framework and dataset for repeatable CPU side latency agentic benchmarking.

Abstract: Large language models (LLMs) are increasingly recognized for their remarkable capabilities; federated learning (FL) offers a promising framework for collaboratively fine-tuning LLMs on edge devices while achieving data privacy. However, edge devices are inherently heterogeneous: clients may exhibit diverse data distributions and resource constraints, including limited memory and computation budgets. Recent federated Mixture-of-Experts (MoE) approaches address these challenges by using sparse expert activation or partial expert loading. However, existing partial-expert methods rely on heuristic full-pool observations or prioritize system-wide load balance, without guarantees for recovering client-specific relevant experts from partial observations. We propose $S^3$-MoE, a server-side sparse expert assignment framework that estimates client-specific expert relevance from partial observations and refines assignments through an exploit--explore mechanism. We also establish relevance-estimation error and exact Top-$K$ recovery guarantees. Experiments using a CNN-based MoE on CIFAR-100 show that $S^3$-MoE recovers task-dependent client--expert preferences and reduces conflicts caused by correlated expert selection across clients. Compared with Full-20 FedAvg, which gives each client access to all 20 experts for local Top-5 routing, $S^3$-MoE achieves 1.31 percentage points higher mean final accuracy while loading only $5$ experts per client.

Abstract: Embodied and edge systems increasingly rely on vision-language models (VLMs) to support heterogeneous skills, yet deploying a separate task adapter for each skill while retaining a high-precision backbone creates substantial memory and management costs. We present UniCompress, a deployment-oriented framework that converts multi-task expertise into one compact, adapter-free model. Starting from a consolidated BF16 Qwen3-VL-4B teacher, UniCompress constructs an asymmetric INT2 student using GPTQ initialization and teacher-guided quantization recovery. The approach jointly quantizes and recovers the projection weights in the vision encoder and language backbone, enabling aggressive compression of the complete vision-language model while preserving its multi-task behavior. We evaluate UniCompress on four heterogeneous embodied tasks: posture recognition, household-safety understanding, multimodal dialogue routing, and mobile-GUI action prediction. UniCompress reduces the approximate weight storage from 8.9 GB in BF16 to 2.1 GB in INT2, corresponding to a reduction of approximately 76%. It improves the INT2 macro task score by 34.3 percentage points over GPTQ-INT2 while remaining within 0.9 points of the consolidated BF16 teacher. These results demonstrate the potential of teacher-guided joint vision-language INT2 recovery for deploying a single multi-skill VLM on resource-constrained embodied platforms.

Abstract: Looped language models apply one block of layers many times per token, so a serving engine pays every per-layer overhead once per loop; we measure serving for Ouro and Huginn-0125 in an open-source looped-model engine and grade our optimizations from bitwise-identical changes to those that require an accuracy check. At batch 1, fusing all loop submissions leaves Ouro throughput at 0.99x; the 5689 small kernels per token are the real cost, and fusing them behind rounding fences raises throughput 1.11-1.59x while 1014/1014 requests remain identical. An approximate fusion is no faster and misses a one-point GSM8K rule on 87 questions (61 and 62 correct versus 64), so we do not ship it. Huginn's released defaults avoid dead KV slots, reuse prefix-cache pages, and serve 1.77-3.16x faster than the base engine; Ouro reaches 1.65-2.15x and agent sessions 1.91-3.21x, while Huginn's noise and prefix-reuse changes stay within one point on all 1319 Huginn test questions; Huginn's packed prefill, -0.61 points on the same questions, meets that rule only on its point estimate. Position-keyed noise makes CPU cache reuse exact but still leaves 20 of 128 GPU full hits different, while loop fusion delays prefill admission and a 64-graph cache prevents eager fallbacks. Under prefill-decode disaggregation ExactLoop reaches up to 3.3x the base throughput with peer error handling disabled, and 3.1x in two of three runs with it enabled. Its transfer budget, sized to hold one request's KV in flight, cuts the commit tail from about one second under the earlier budget to 16 ms (base: 2.8 s), removing the polling rounds that the engine's interval would otherwise impose.

Abstract: Online continual learning aims to continuously train neural networks on a non-stationary data stream in a single pass. Rehearsal-based methods alleviate catastrophic forgetting by replaying stored samples, but the imbalance between current and previous data often leads to predictions biased toward current classes. Many approaches have been proposed to overcome the forgetting problem by correcting the bias; however, they still need to be improved in an online fashion. By decomposing the dot-product logits into an angle factor and a norm factor, our analysis reveals that the angle factor efficiently captures novel knowledge but is particularly susceptible to prediction bias, whereas the norm factor helps preserve historical knowledge. Based on these observations, we propose Dual-Logit Experience Replay (DLER), which exploits dual-logits for both model optimization and memory management. Specifically, DLER learns current samples using cosine logits and incorporates additional dot-product logits when replaying previous samples to mitigate prediction bias. To strengthen dual-logit learning under limited replay memory, an auxiliary supervised contrastive objective promotes angular compactness within each class, extending corrections learned from limited replay samples to other samples of the same class. Meanwhile, we introduce a memory update strategy that preferentially retains samples for which dot-product prediction assigns a higher probability to the correct class than cosine prediction, aligning sample selection with dual-logit training. Extensive experiments on three datasets demonstrate the effectiveness of DLER, particularly under limited memory budgets.

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

Abstract: Linear recurrences are becoming an increasingly common building block in large language models, replacing the growing KV cache of softmax attention with a fixed-size recurrent state that enables efficient decoding. However, the quantization stack has largely been developed around softmax Transformers, whose numerical properties and inference state differ from those of recurrent models. Our goal is to understand these differences and study the design space of quantization methods for recurrent models. In our analysis, we differentiate between two targets. First, extra-state quantization concerns weights and activations outside the recurrent state, whose distributions are shaped indirectly by the sequence mixer during training. In controlled 1.3B parameter models, we find that decayed recurrences develop weaker residual outlier channels than softmax attention and are therefore easier to quantize without incoherence processing. Second, intra-state quantization concerns the recurrent state itself during decode, where repeated updates create a different numerical problem from KV cache quantization. We instantiate our findings as LoReS, a state quantization pipeline combining value-axis scale grouping to match output channel magnitude, decay-based mixed-precision to protect slowly decaying heads or channels, and stochastic rounding to limit error accumulation under repeated requantization. Across Qwen-3.5 9B, Nemotron-H 8B, and Kimi-Linear 48B-A3B, LoReS supports robust 8-bit state quantization and lower effective precision on shorter decode horizons. We implement LoReS in vLLM and demonstrate end-to-end decode speedups.

Abstract: Adam keeps two FP32 moments per parameter, twice the bytes of the FP32 weights, yet compressing them is risky: the update divides by the square root of the second moment, so small reconstruction errors can change the restored computation substantially. We study guarded moment compression, in which a coordinate's decoded moments are accepted only if they reproduce the native FP32 Adam update to within a tolerance $τ$, in units of the learning rate, for every gradient in a clipped domain; all other coordinates are stored exactly. The natural certificate, which ranges each update separately over gradient intervals, fails: on one million trained OPT-125M coordinates it escapes 98.9% of coordinates and the payload grows to 126.6% of raw FP32. Our key observation is that both updates consume the same gradient and parameter. Bounding their difference directly, with FP32 rounding-error allowances and a fixed mixed linear-logarithmic gradient partition, cuts escapes to 0.53% and the payload to 28.3% at $τ=0.01$. Across 120 complete-checkpoint restores of OPT-125M and ResNet-18, with four trained seeds and three training fractions per model, guarded 8-bit log codes shrink moment payloads by 69.5% and 71.9% relative to FP32, 40.8% and 45.4% relative to equally guarded BF16, and 29.3% and 37.6% relative to equally guarded blockwise 8-bit, and every restored first update stays within tolerance. Encoding and restoration together take 2.3-2.6$\times$ as long as guarded BF16 but less time than guarded blockwise 8-bit. Searching both moment codes jointly saves under 0.001% more bytes at 6% more codec time, so per-moment rounding suffices. The guarantee covers the first restored update under a pinned Torch operation model; we measure subsequent training empirically.

Abstract: Quantization of Large Language Models (LLMs) has be- coming increasingly important for reducing the computa- tional costs and resource requirements for practical deploy- ment.Growinginterestinthisareahasmotivatedthecommu- nity to move beyond 8-bit quantization towards 4-bit preci- sion, leading to the development of customized 4-bit formats such as MXFP4 and NVFP4. Recently, HiFloat4 (HiF4) has emerged as a competitive alternative which adopts a Block Floating-Point(BFP)formatwithathree-levelscalinghierar- chy, thereby achieving a significantly larger dynamic range. While quantization can substantially reduce memory usage and accelerate inference, it frequently introduces accuracy degradation. Calibration is therefore essential for narrowing theperformancegapbetweenlow-precisionquantizedmodels and their full-precision counterparts. Nevertheless, existing calibration methods are mainly designed and validated us- ing integer-precision formats, such as INT4, lacking support for HiF4 due to its hierarchical BFP structure. In this paper, weinvestigatethreecalibrationapproaches,namelyHiGPTQ, HiAutoRound, and HiFlatQuant, to further enhance model quantizationbasedontheHiF4format.Extensiveexperiments demonstratethatallthreeproposedapproachesimproveinfer- ence accuracy compared with direct HiF4 conversion, reduc- ing the gap between HiF4 and BF16 to less than 1% and out- performing the state-of-the-art NVFP4 format across various models and downstream tasks.

Abstract: Crowd-counting models deployed on edge devices face two shifts: inputs are degraded along the capture and transmission pipeline, while models are quantized to satisfy resource constraints. Yet corruption robustness and deployment quantization are typically evaluated separately. We introduce CrowdCC, a controlled corruption benchmark that jointly evaluates these effects using 23 deployment-motivated corruptions at three severity levels, four crowd-counting datasets, and 26 models using A100 TensorRT engines at FP32, FP16, and post-training INT8 precision, paired with throughput and energy measurements from corresponding Jetson Orin NX configurations. The corruptions span six families, with head annotations adjusted when geometry or visibility changes. Our evaluation shows that capture-chain corruptions consistently produce the largest degradation across the evaluated models, while crowd density affects absolute and relative error differently. Clean accuracy does not fully predict corrupted performance. INT8 alters corruption robustness substantially more than FP16, with no systematic direction, and does not consistently improve latency or energy. INT8 inference efficiency is strongly associated with reformat overhead in the compiled engine, while slowdowns are also observed under limited INT8 coverage. Precision changes corrupted-error rankings and robustness-cost frontiers, showing why model selection must consider corruption robustness and measured deployment cost together. The code is available at https://github.com/Tulane-Computer-Architecture-Systems/CrowdCC-Release.

Abstract: Deploying large vision-language models (LVLMs) on resource-constrained embodied agents for 3D visual question answering necessitates cloud-edge collaboration. However, offloading massive visual data to the cloud over bandwidth-limited and noisy channels introduces severe latency and transmission distortion. To address this, we propose TP-JSCC, an efficient token pruning and joint source-channel coding method for LVLMs-based 3D-VQA in Cloud-Edge Collaboration. First, we develop a Local and Global Semantics-based Hybrid Token Pruning (LGS-HTP) module to significantly reduce the visual token count. By decoupling foreground-background tokens and jointly selecting local and global semantics, LGS-HTP effectively reduces redundancy while preserving reasoning-critical tokens. Second, we introduce a lightweight Token Joint Source-Channel Coding (T-JSCC) module to further compress the pruned tokens and ensure their robust transmission over error-prone channels. By employing lightweight cascaded linear layers and Signal-to-Noise Ratio (SNR)-adaptive mechanism, this module adapts to fluctuating channel conditions while maintaining minimal computational overhead. Extensive experiments on ScanQA and SQA3D datasets indicate the exceptional efficiency and robustness of our approach. Specifically, on ScanQA validation set, the LGS-HTP module could achieve 90% pruning ratio while incurring less than a 1% CIDEr loss. Building on this, the T-JSCC further compresses the tokens by 18 times, resulting in an overall 180 times data volume reduction for robust transmission, while saving 90% in computational complexity for LVLM deployment. Our code is available at https://github.com/yuanhui0325/TP-JSCC.

Abstract: Distributed Cyber-Physical Systems (CPS) and Edge AI networks increasingly depend on public-key authentication and key-establishment mechanisms that are vulnerable to cryptographically relevant Quantum Computers (QC). Shor's algorithm threatens widely deployed asymmetric primitives, while Grover's algorithm reduces the effective security margins of symmetric and hash-based mechanisms. Although Post-Quantum Cryptography (PQC) provides a migration path, its larger public keys, ciphertexts, and signatures introduce additional computation, memory, communication, storage, and packet-fragmentation costs that can constrain dynamic multi-agent systems. In this article, we introduce rotational secrecy, an architectural objective that limits temporal key exposure through authenticated transitions between cryptographic epochs. We present Private Authentication for Rotational Secrecy (PARS), a reference mechanism that coordinates policy-triggered key establishment, rotation, and reconciliation among authorized Physical AI agents. For deployments that need distributed agreement on security state, we further develop the PARS-Enabled Edge Security Ledger (PARS-ESL), an optional permissioned architecture maintained by authorized validators or capable edge gateways. PARS-ESL does not store high-volume operational transactions. Instead, it records authenticated state for key epochs, group membership, revocation, algorithm-suite policies, challenge validation, and recovery checkpoints. As a result, resource-constrained agents can take part as lightweight clients, and low-latency operational traffic stays outside the ledger consensus path, protected by the currently active symmetric epoch keys. Our results show that ledger-assisted post-quantum key-epoch management is feasible for selected Edge AI and CPS security-control workloads under the stated trust, compromise, connectivity, and resource assumptions.

Abstract: Hardware-aware neural architecture search (NAS) requires jointly optimizing predictive accuracy and device-specific latency over discrete search spaces. While multi-objective swarm optimizers show promise, continuous formulations such as the Multi-Objective White Shark Optimizer (MOWSO) can suffer from quantization noise when rounded to integer-encoded architecture vectors. To address this limitation, we propose MOSHO (Multi-Objective Shark Hunting Optimization), a native discrete metaheuristic that replaces continuous motion equations with discrete architecture-editing operators. MOSHO integrates adaptive operator scheduling, energy-regulated restarts, and deferred crowding-distance truncation to enhance exploration and search efficiency. On the merged NAS-Bench-201/HW-NAS-Bench benchmark across 18 contexts, MOSHO achieves a median normalized hypervolume of 0.9987 and IGD+ of 0.0064, yielding Pareto fronts competitive with NSGA-II while reducing search runtime by 96.4% (0.64 s vs. 17.72 s). These results demonstrate that tailored discrete search mechanisms can enable swarm optimizers to achieve high-quality multi-objective solution quality with substantially lower search cost in hardware-aware NAS.

Abstract: Split inference partitions a Vision Transformer between a resource-constrained edge device and a server. However, transmitting intermediate tokens across the communication channel is costly. Deploying fine-tuned backbones to every endpoint becomes expensive as well. We study token reduction for split inference under a frozen backbone, where only a small adapter is shipped. We observe that the CLS token plays a dual role: it acts as the global summary the receiver needs for classification, and it serves as a signal on which patches to keep. We propose Consommé, a method utilizing both characteristics. There, an information bottleneck compresses the CLS token into a compact latent which is transmitted in place of the full CLS token and conditions a learned patch gate merging patches into exactly $k$ tokens via soft assignment. It reaches any budget in a single step, thus escaping the per-layer cap of bipartite matching (BPM). On DeiT-S/T, it leads all methods at aggressive token budgets and remains operable at shallow splits where BPM cannot operate, while matching fine-tuned methods at gentle budgets at a fraction of their deployment cost.

Abstract: For efficient deployment of large language models (LLMs), methods of ultra-low-bit backbones reduce weight storage, memory traffic, and linear-layer execution cost, while multi-token prediction (MTP) methods reduce sequential decoding steps by predicting multiple future tokens in parallel. Although complementary, the direct combination of these methods cannot accelerate decoding at comparable quality: masked drafting lacks intermediate token input supplied in next-token training, while wider computation and causal key-value (KV) updates can offset the savings from fewer steps. To address these challenges, we propose BitMTP, an algorithm-system co-design framework for efficient MTP on ultra-low-bit LLM backbones. A joint next-token prediction and MTP objective equips a single model with both autoregressive (AR) generation and block-causal drafting. Confidence-only decoding, speculative-only decoding, and hybrid decoding provide different quality-throughput operating points without using separate draft and verifier models. Accepted-prefix alignment keeps the committed prefix, materialized causal KV boundary, and CUDA Graph replay state consistent for continued generation. BitMTP uses this contract to select width-specialized low-bit execution and reuse captured graphs over shared KV storage. On NVIDIA H20, BitMTP with hybrid decoding achieves five-task mean speedups of 2.2x on BitNet b1.58-2B and 2.1-2.4x on 1.7B-4B Bonsai models at comparable task quality. On RTX 3090 and Jetson Orin, it further achieves up to 2.0x in five-task mean throughput and 3.3x on MBPP+, respectively. These results demonstrate that BitMTP accelerates ultra-low-bit LLM generation across backbones, model sizes, and GPU types. The source code is released at https://github.com/PolyU-VCLab/BitMTP

Abstract: Analog in-memory computing delivers efficient neural network inference at the edge, but systems deployed in non-stationary environments may need to adapt in situ to locally acquired data. Backpropagation faces two structural obstacles here. Its global error transport lets stochastic write errors compound across layers, and its backward pass demands bidirectional, high-precision peripheral circuitry absent from inference-oriented crossbar arrays. We introduce Channel-wise Spatial Gradient learning (CwSG), a strictly local, forward-only rule that trains each convolutional layer by rewarding spatial variation in its feature maps. In simulations replaying the measured trajectories of 1,189 phase-change memory and 1,268 filamentary memristor devices, CwSG learns with single, unverified, sub-1 V pulses. It matches backpropagation on phase-change memory with 59-fold fewer pulses and outperforms it by 53 percentage points on highly variable filamentary memristors, with 8-bit forward converters and without transposed crossbar reads or network-wide activation storage. Analog training thus becomes a problem of credit assignment, in which algorithmic design rather than hardware compensation alone overcomes severe programming noise.

Abstract: Resource-constrained robots and autonomous vehicles need scene memories that localize a camera and render nearby views without a desktop GPU. Scene-coordinate regression (SCR) and 3D Gaussian splatting (3DGS) fit because they avoid an SfM stage, but ACEsplat-style pipelines are GPU-centered, and moving them to an edge FPGA is hard because SCR is pose-sensitive and 3DGS rendering is irregular and capacity-sensitive. We present LENS, an edge scene-memory acceleration system on the AMD KR260 that brings an ACEsplat-style workload to a low-power FPGA/ARM platform. Rather than proposing a new scene model, LENS co-designs the two dominant kernels with the board data path: an INT16 SCR accelerator that preserves pose accuracy, and a bit-exact fixed-point ordered 3DGS renderer that rewrites alpha transmittance as a log-sum-exp recurrence and removes provably useless streamed Gaussians. All LENS latency, power, and quality results are board-measured. The SCR engine preserves DSAC* localization accuracy at 25.42 ms, the on-board renderer reaches 34.11 dB on Cambridge against ACEsplat's 33.20, novel-view synthesis on RealEstate10K reaches 27.84 dB against 26.11, and all ten Wayspots scenes reach the ACEsplat rendering target. LENS wins performance per watt over a Jetson Orin NX on every deployment task, by 1.9x to 5.7x at about a quarter of the power, and is faster in absolute latency on localization and Wayspots rendering. On-device training reaches ACEsplat-level quality but not the GPU's time or energy, marking the remaining boundary for a future training accelerator.

Abstract: Recent works show that LLMs can generate kernels with strong performance on server GPUs. Edge deployment presents distinct challenges because the hardware is heterogeneous, the software stacks are fragmented, and the resource budgets are tight. LLM kernel generation across diverse edge backends has yet to be studied systematically. We introduce EdgeKernelBench, which combines a unified IR deployment framework with a hierarchical benchmark for edge workloads. The unified IR captures reusable computation, scheduling, and memory decisions and connects to each target backend through a compiler interface. The benchmark spans progressively larger program scopes, from individual operators to compact model graphs, across representative edge application domains. We instantiate the design with TVM as a reference compiler and validate it across heterogeneous edge backends. With a single generation attempt, the best model achieves 64.5% Correct@1 but only 25.5% Fast@1.0. This gap shows that functional correctness does not consistently translate into performance gains. We therefore introduce the Layered Edge Kernel Agent (LEKA), which organizes optimization through a layered IR. On the agent evaluation subset, LEKA achieves 43.5% Fast@1.0, compared with 13.0% for direct generation and 26.1% for Claude Code. LEKA and Claude Code use the same base model and optimization budget. Together, EdgeKernelBench and LEKA support unified evaluation and layered optimization of kernels generated by LLMs across edge backends.

Abstract: Vision-Language-Action (VLA) models offer a promising paradigm for autonomous driving. However, the massive parameter scale of VLAs strains onboard computational resources during real-time inference. This directly increases energy consumption and limits vehicle battery range, making it critical to minimize computational overhead without compromising driving capability and explainability. Current research on efficient VLA-based driving focuses on sequential adaptation (\textit{e.g.}, Chain-of-Thought and token adjustment), leaving structural adaptation underexplored despite parameter redundancy across dynamic driving scenes. To address this gap, we propose Scene-Adaptive VLA, an efficient framework leveraging dynamic layer routing to adjust active parameters based on real-time scenes. Inspired by the temporal continuity of scenes and model predictive uncertainty, our approach characterizes model-perceived scene complexity via frame-level temporal scene variations and decision deviations. To uncover latent scene-to-parameter correspondences, we encode this characterization into scene-aware tokens, alongside two specialized queries. A budget-conditioned layer routing mechanism evaluates these refined queries to allocate a budget that determines the active parameter ratio and subsequently selects appropriate layer combinations. The entire framework is optimized end-to-end. Extensive evaluations on the Bench2Drive benchmark show that when adapted to a state-of-the-art VLA model, our approach significantly reduces inference-time computational overhead by 53.5\% with negligible degradation in driving and language capabilities. Code is available at https:github.com/yang-luvu/Scene-Adaptive-VLA.

Abstract: LLMs and LVLMs are increasingly deployed across heterogeneous hardware, ranging from data-center GPUs to resource-constrained edge devices. Their fixed-size architectures limit the ability to reuse a single model across deployment settings with different compute, latency, and model-size targets. We introduce HydraLLM, an elastic MQA-based architecture. Its number of active attention heads determines the effective widths of the embeddings, attention projections, feed-forward layers, and the output head, forming a nested family of subnetworks within a single parameter set. We introduce a static-to-elastic conversion pipeline that, starting from a pre-trained model checkpoint, includes importance-driven weight reordering and stochastic subnetwork training with knowledge distillation. We extend this principle to multimodal inference through an elastic projection module whose output width follows the active text width. HydraLLM covers a wider range of active parameter configurations than FFN-only elastic architectures. With 75% of heads active, HydraLLM achieves 2.2 percentage points higher TriviaQA accuracy than a state-of-the-art nested baseline configuration which has 17.25% more parameters. With 50% of heads active, our multimodal model achieves a CLIPScore comparable to that of the 7$\times$ larger Gemma 3 4B model. With only 25% of heads active, HydraLLM increases token-generation throughput by 75.3% over the full-width configuration.

Abstract: Diffusion-based LLMs (dLLMs) have recently emerged as a promising alternative to autoregressive (AR) LLMs by enabling bidirectional parallel refinement, alleviating the sequential decoding bottleneck of AR generation. However, their parallel iterative refinement mismatches AR accelerators optimized for sequential decoding and their discrete token generation differs from DiT accelerators designed for continuous denoising. Recent dLLM accelerators have explored workload-specific optimizations to reduce vocabulary processing overhead and redundant computation across denoising iterations. However, these approaches retain all tokens in parallel execution, despite varying token refinement utility and execution requirements. This paper presents DynaTE, a hardware--software co-design architecture that dynamically adapts accelerator execution to evolving token states during dLLM decoding. DynaTE first enables adaptive token execution by skipping low-utility token computation, while a dimension-reconfigurable PE array maintains high utilization under varying active-token patterns. Second, DynaTE exploits dynamic token dependencies through FLDD to refine a small number of locally dependent tokens within the current iteration, reducing the overall number of denoising iterations, while a Merge--Split--Merge dataflow hides the resulting serial overhead. Third, a streaming vocabulary engine interleaves multiple token streams from the LM head to accommodate irregular output variations caused by selective token computation and uneven vocabulary-selection demands. Evaluated on two representative dLLMs, DynaTE achieves 2.05--2.78$\times$ speedup and 2.99--3.93$\times$ higher energy efficiency over state-of-the-art dLLM accelerators, while delivering 2.55$\times$ speedup and 6.07$\times$ higher energy efficiency over Jetson AGX Orin.

Abstract: SVD-based pruning and quantization have recently emerged as a promising strategy for the ultra-efficient compression of large language models. In these methods, compression is performed in two stages: components are first truncated, and the remaining ones are subsequently quantized. Although this decoupled pipeline benefits from both pruning and quantization, it requires separate optimization for each stage and fails to fully exploit their balance, which can lead to suboptimal performance under aggressive compression. To address this limitation, we propose a new LLM compression method that co-optimizes pruning and quantization in a unified framework. Our key idea is a differentiable method for learning component-wise bit-widths, allowing less important components to be assigned 0-bit precision and pruned away. Notably, our method performs favorably against two-stage baselines, even when subjected to extreme quantization settings ($1.61$ bits) designed for ultra-efficiency. Code: https://github.com/MMAI-Laboratory/DBW.

Abstract: Split Computing (SC) enables efficient deployment of Deep Neural Networks (DNNs) by partitioning inference between edge devices and cloud servers. However, intermediate feature representations are simultaneously exposed to hardware faults and adversarial attacks, which are traditionally evaluated independently. This paper presents a unified framework for the joint assessment of reliability and security in Split Computing. First, reliability is characterized through neuron-level fault injection using the Mean Relative Accuracy Degradation (MRAD) while security through feature-map-aware adversarial attacks simulations using the Attack Success Rate (ASR). Based on these complementary analyses, the Joint Vulnerability Score (JVS) is introduced, along with a confidence-aware extension that jointly captures prediction errors and confidence degradation. The framework is evaluated on ten Split Computing configurations based on ResNet-50 trained on ILSVRC-2012. Experimental results show substantial differences across compression strategies, with MRAD ranging from 44.3% to 61.2% under fault injection, while adversarial attacks achieve up to 98.8% ASR. Furthermore, the proposed joint metrics reveal vulnerability trends that remain hidden when reliability and security are analyzed independently, providing a more comprehensive methodology for designing dependable Split Computing systems.

Abstract: Unstructured pruning at extreme sparsity often suffers from representational collapse, causing sharp drops in accuracy. To address this, we study Backbone Contrastive Pruning (BaCP), which regularizes the sparse network's embedding space by aligning it with pretrained, fine-tuned, and historical snapshot models. Building on the contrastive decomposition of the CAP framework (Xu et al., 2022), we provide a rigorous matched-budget characterization of this approach across multiple pruning criteria. Evaluated across 90 settings, BaCP improves accuracy substantially in extreme sparsity regimes where standard pruning fails, and is close to baseline where representations remain intact.

Abstract: Video diffusion models are rapidly scaling and exhibiting enhanced generation capabilities. Among these recent advancements, MiniMax-H3 stands out as a highly capable, production-level open-source model. However, its 33-billion parameters and multi-step iterative denoising process introduce substantial computational overhead. Consequently, their practical production is hindered by generation latency in the cloud deployment like NVIDIA-GB200, alongside strict memory limits that pose further challenges at the edge device like DGX-Spark. To address these diverse hardware bottlenecks from cloud to edge device, we present a full-stack inference pipeline that integrates efficient algorithmic design with optimized operator implementations. Algorithmically, we introduce a cross-resolution two-stage generation scheduler that exploits the step-wise nature of diffusion: early low-resolution steps rapidly establish the global layout, while later high-resolution steps focus refinements of local and perceptual details. These stages are connected by a learned latent-to-latent mapping module, completely eliminating the computationally expensive VAE decode-reencode cycle for resolution transferring cross different resolutions. For operator implementation, we deploy a Recursive Self-Improvement (RSI) loop that searches kernel fusions and memory layouts, evaluating latency together with numerical agreement. Together, these optimizations deliver up to 30x end-to-end speedup and 20% lower memory: a 5-second 1344x768 video with audio is generated 3.5x faster than real time on an 8xGB200 node, and in under a minute fully memory-resident on a single DGX Spark.

Abstract: Parameter-efficient adaptation enables continual learners to acquire task-specific knowledge through compact model updates while maintaining strong within-task performance. However, class-incremental inference requires each input to be classified among all classes seen so far without access to its task identity. For learners equipped with task-specific parameter-efficient modules, this introduces a critical task-routing challenge beyond catastrophic forgetting. We study post-hoc task routing without retraining the learner or introducing a separately trained router. Such training-free inference-time calibration remains comparatively underexplored in parameter-efficient class-incremental learning. We identify three sources of routing error (feature-level, task-level, and class-level misalignment) and propose Feature Distribution Calibration (FDC). Its three components address these misalignments: Task Subspace Filtering (TSF) suppresses feature components outside each task's principal subspace, Residual Likelihood Calibration (RLC) evaluates the typicality of its subspace residual, and Prototype Affinity Calibration (PAC) measures compatibility with the task's class prototypes. Experiments demonstrate plug-and-play applicability to eight parameter-efficient class-incremental methods using a shared encoder. With one component configuration selected per method across all five benchmarks, FDC improves final accuracy in all 40 method-dataset pairs by 4.39 percentage points on average. Enabling all components improves 35 of the 40 pairs, with an average gain of 4.45 points. When applied to a simple baseline, FDC achieves strong overall performance.

Abstract: Confidence-gated edge--cloud inference accepts confident local predictions and offloads uncertain inputs to a stronger cloud model. We show that this routing decision creates an availability attack surface. We call this attack \emph{GateDrain}: bounded input perturbations lower calibrated confidence and redirect requests that would otherwise be answered locally into a shared cloud queue, without increasing the application request rate. Because escalated requests share a cloud service, an increase in per-request cloud demand can move a near-capacity deployment across a queueing knee, causing disproportionate tail-latency degradation for benign users. We evaluate white-box, transfer, decision-only, universal, and multi-gate attacks on the public EdgeBoost artifact. A fixed-application-volume comparison isolates the effect of confidence manipulation from added client traffic, while perturbation-budget and arrival-process sweeps show that the queueing transition persists across several workload models but its amplification depends on the operating point. Adaptive attacks also defeat the evaluated training-free preprocessing defenses. To contain the resulting cloud demand, we evaluate Bounded Escalation, which combines per-source admission budgets, protected capacity, and non-preemptive trusted-class priority; an optional global bucket adds an identity-independent bound on untrusted admissions. The evaluation makes the resulting policy trade-off explicit: authenticated clients receive latency isolation, whereas tighter aggregate containment can reject legitimate unauthenticated offloads and reduce overall expected accuracy through edge fallback.

Abstract: Three-dimensional (3D) Gaussian splatting (3D-GS) has emerged as a promising technique for large-scale scene reconstruction due to its high rendering efficiency and fidelity. However, the training of large-scale 3D-GS models at wireless edge faces various technical challenges including the limited communication, computation, and graphics processing unit (GPU) memory resources at edge devices, the structural inconsistency issue across local models hindering their effective aggregation, as well as privacy leakage risks associated with raw visual content and camera parameters. To address these challenges, this paper proposes a novel resource-efficient federated learning framework for efficiently training 3D-GS models of large scenes under severe resource constraints. First, we propose an on-device model lightweighting mechanism that adaptively selects and prunes Gaussian points to balance the rendering quality and training efficiency. In this mechanism, we quantitatively evaluate the importance of different Gaussian points at each device to facilitate the pruning, and use a novel importance-to-latency ratio criterion to determine the number of pruned Gaussian points under GPU memory and computation/communication latency constraints. Furthermore, we develop a 3D-GS model recovery mechanism that restores structural consistency across local 3D-GS models without accessing private camera parameters, enabling their effective aggregation towards a global model. Finally, extensive experiments show that our approach significantly accelerates convergence, maintains high rendering quality, and reduces training latency compared to state-of-the-art federated 3D-GS baselines.

Abstract: Large neural networks can be compressed by rounding or "quantizing" their weights to numbers that admit representations with fewer bits. One algorithm for quantization, OPTQ, progressively quantizes the weights of a neural network so that the squared quantization error on a specified calibration dataset is as small as possible. We study the performance of OPTQ and a variant algorithm, stochastic OPTQ, in a generalization setting and derive bounds for the expected squared error accrued by the algorithm when a test point is drawn from a fixed distribution. We prove two results. One result relates the generalization error to the error on a calibration dataset comprising independent samples from the same distribution as the test distribution. The other result bounds the generalization error of stochastic OPTQ for all sufficiently nice distributions, regardless of the calibration dataset. In both of these results, the regularization term $λ$ plays an important role. We use insights from these results to make a new recommendation for the choice of $λ$ and see that this choice of $λ$ preforms favorably in experiments when compared to prior recommendations in the literature.

Abstract: Neural networks are typically adapted by computing gradients and updating model parameters. We investigate whether task-specific adaptation can instead emerge from a meta-learned self-organising process that requires no gradients at adaptation time. We instantiate this idea with a Neural Cellular Automaton in which locally interacting recurrent cells maintain both a recurrent state and a fast associative memory. During meta-training, backpropagation is used to learn the recurrent dynamics together with how the memory is read and written. Once training is complete, the slow model parameters remain fixed, and online adaptation occurs only through cellwise memory updates driven by local prediction errors and a delta rule. We evaluate whether the learned mechanism can adapt to semantically distinct held-out classification tasks. A single pass over the support data produces substantial improvements in held-out performance without gradient computation or parameter updates during adaptation, and the mechanism remains effective across large changes in the number of examples processed jointly. These results show that task-specific adaptation can be achieved through explicit fast-memory updates while keeping the slow model parameters fixed.

Abstract: Deploying deep learning models on edge CPUs is bottlenecked by computational and memory constraints. Mixed-precision quantization promises to reduce inference latency while preserving accuracy. However, quantization affects different layer types in inconsistent ways, so identifying where accuracy loss is minimized and latency reduction is maximized is critical, as the effect accumulates over a full deployment into substantial savings or unacceptable task degradation. Such identification relies on sensitivity metrics, proxies that estimate layer-wise degradation without evaluating the task accuracy of every candidate policy. Nevertheless, widely used metrics fail systematically on modern architectures. We present a systematic empirical study of 13 sensitivity metrics for layer-wise INT8 quantization across four distinctly different neural networks, and validate the resulting policies on two ARM64 platforms. Gradient-based sensitivity methods fail on 4 out of 8 model-hardware configurations and weight-based statistics on 2. In contrast, the Jensen-Shannon Divergence achieves zero catastrophic failures, reliably isolating the layers that cannot be safely quantized. A sensitivity metric alone does not define a policy, and the fixed thresholds typically used for that step are fragile over the highly skewed distributions of modern architectures. We address this with K-Means clustering, achieving near-lossless accuracy and a mean speed-up of $1.81\times$ over the full-precision model. Finally, we reveal that excluding from quantization the layers whose speed-up is negligible, regardless of their sensitivity, can be counterproductive, as it induces computational graph fragmentation and disables operator fusion. Our results yield concrete allocation policies for practitioners and researchers deploying quantized vision models on heterogeneous edge CPUs, without GPU access or gradient computation.

Abstract: GPU robotics lacks the reusable numerical infrastructure of mature CPU stacks, instead relying on compiler frameworks that introduce overhead or repeatedly reimplementing numerical libraries. To address this, we introduce GLASS (GPU Linear Algebra Simple Subroutines), a header-only CUDA C++ library that provides thread-, warp-, block-, and NVIDIA-backed implementations of robotics-scale linear algebra and geometric computations under one composable device API. GLASS treats implementation choice, execution scope, and launch packing as architecture-specific placement decisions determined by offline measurement and resolved statically at compile time. This is critical as the best and worst placements differ by a median of 4.9x (max 81x), with 145 of 396 recommended placements changing between a Jetson AGX Orin and an RTX 5090, and 162 of 396 versus an AGX Xavier. These stakes are highest at the edge as GLASS's advantage over the best of PyTorch and JAX is as much as 73x on the Orin versus 12x on the RTX 5090. GLASS is released open source with independent numerical oracles and source-bound local-GPU test attestation. Finally, integrating GLASS with published robotics systems both exposed a pre-existing numerical bug and improved embedded runtimes by up to 1.5x.

Abstract: Edge deployment motivates forecasting models with compact parameter storage and low-bit representations. Deep equilibrium models (DEQs) obtain implicit depth by repeatedly applying a shared layer, reducing the parameter cost of explicit layer stacking. Their usual Anderson solver, however, searches for update coefficients in the continuous real domain. We propose Q-DEQ, which formulates local updates in DEQ forward solving as discrete optimization problems. Candidate directions are constructed from the current state and iteration history, and a local quadratic residual model is used to evaluate their combinations. Binary encoding of the direction coefficients yields a quadratic unconstrained binary optimization (QUBO) problem that can be solved by simulated annealing (SA) or a coherent Ising machine (CIM). After fixed-point solving, a re-forward pass applies W8A8 fake quantization to the shared layer's weights and activations. We evaluate Q-DEQ with an iTransformer backbone on five multivariate time series forecasting datasets. Relative MSE differences from the explicit multi-layer baseline range from $-1.16\%$ to $+2.90\%$, with lower MSE on two datasets. DEQ parameter sharing reduces parameter counts by factors of $1.80\times$--$3.82\times$; combined with W8A8, static weight storage is reduced by factors of $4.3\times$--$12.8\times$. Local QUBO problems solved using CPU-based SA and the Kaiwu CIM physical backend produce closely matching downstream forecasts. These results establish local discrete solving as a viable component of DEQ time series forecasting and provide a route for executing fixed-point updates through different combinatorial optimization backends.

Abstract: Large vision-language models (VLMs) enable recognition beyond a fixed class set, but their computational demands prevent them from running on many edge devices. Cloud offload makes this capability accessible, but sending every image consumes scarce bandwidth and communication energy. We ask how to bring the open-world recognition capability of VLMs to the edge while operating within tight compute, energy, and bandwidth budgets. Wildlife monitoring provides a natural setting for exploring this question because camera traps encounter species not known at deployment. We present Scout, an autonomous open-world recognition system that invokes a cloud VLM intermittently to teach new classes to a compact edge model. Given only the deployment location and empty site frames, Scout autonomously turns each species identified by the VLM into persistent, site-conditioned recognition capability in a resource-efficient edge model, without a predefined species list, human labeling, or manual tuning. Across 30 camera-trap deployments in three regions on an NVIDIA Jetson Orin Nano, the accuracy of Scout remains within 0.1-2.5% of a model given a predefined species list. On species outside its initial class set, Scout achieves 53.7-59.1% accuracy, compared with 56.5-65.1% for full cloud offload, while using 59-71% less deployment energy.

Abstract: Bird-strike mitigation at wind farms requires identifying distant birds that span only tens of pixels, so the classifier's input resolution N is a design variable, not a fixed specification. We study it with a factorial design over 14 side lengths N (16 to 224), six architectures, two training and evaluation regimes and 30 random seeds -- 2,520 checkpoints and 5,040 evaluations -- plus latency measured on an NVIDIA Jetson Orin Nano. Four results. (1) The selected N depends on the target: 0.90 is met on validation by ResNet50 at N=112 in an estimated 1.85 ms (0.8980 on test) and 0.95 by DINOv2-L at N=144 in 12.70 ms; changing the model buys more accuracy than raising N (+5.93 versus +2.33 points at N=112). (2) The benefit of lowering N depends on the assumed preprocessing path: N=224 -> 80 saves 13.5% when each individual is decoded from its own file but 46.2% when the detector decodes the 4K frame once; the Pareto set grows from 20 to 23 configurations. (3) Accuracy must be measured on the deployed engine: half precision costs ViT-S/16 alone 4 to 7 points at N>=96 while the CNNs stay within 0.1 points, and with selection held at validation the choice differs at 26 of 176 targets. A broken FP16 engine can run faster than a correct one, undetectable from latency; admitting 14 ViT-S/16 FP32 configurations moves the recommendation over the 0.931-0.938 band and under the 10 ms budget. (4) ViT-L-scale models fit this device, but activations exceed the FP16 range; splitting the graph at transformer-block boundaries confines FP32 to the affected segments, making the deployed DINOv2-L chain 1.85x faster than the single-engine build. We also quantify how the regime-difference sign stabilises with seed count; a sensitivity split removing some forms of group sharing preserves all 14 non-trivial signs at the selection boundary.

Abstract: The lottery ticket hypothesis posits the existence of winning tickets: sparse subnetworks that, when trained in isolation from their original initialization, match the accuracy of the full dense network. The predominant method for discovering such tickets, iterative magnitude pruning, alternates pruning with full retraining from scratch until convergence over many cycles. Similarly, deep active learning also retrains a model from scratch after each acquisition round as new labels become available. Despite this shared reliance on iterative retraining with a substantial computational overhead, the two paradigms have been studied separately. We observe that the iterative training loop inherent to pool-based active learning already provides the exact computational structure that iterative magnitude pruning exploits, and propose Improve & Prune (I&P), a method that integrates magnitude pruning into each active learning retraining cycle at practically no additional cost. This raises a key empirical question: can iterative magnitude pruning produce winning tickets under the non-stationary data regime of active learning? We investigate this question across multiple acquisition functions, architecture families, and image classification datasets, including an active fine-tuning scenario. Our results demonstrate that I&P yields sparse, deployable models at each active learning iteration. Those match the accuracy of their dense counterparts at sparsities up to 95%, effectively obtaining winning tickets as a byproduct of the active learning pipeline. These per-iteration sparse models can address two computational bottlenecks - per-round model retraining and acquisition scoring over the unlabeled pool - that currently prevent the practical adoption of DAL on large architectures and large unlabeled pools.