Low-Latency Systems

Edge devices are increasingly utilized for deploying deep learning applications on embedded systems. The real-time nature of many applications and the limited resources of edge devices necessitate latency-targeted neural network compression. However, measuring latency on real devices is challenging and expensive. Therefore, this letter presents a novel and efficient framework, named EvoLP, to accurately predict the inference latency of models on edge devices. This predictor can evolve to achieve higher latency prediction precision during the network compression process. Experimental results demonstrate that EvoLP outperforms previous state-of-the-art approaches by being evaluated on three edge devices and four model variants. Moreover, when incorporated into a model compression framework, it effectively guides the compression process for higher model accuracy while satisfying strict latency constraints. We open source EvoLP at https://github.com/ntuliuteam/EvoLP.

Coordinated beamforming in distributed 5G networks relies on the timely exchange of inter-cell scheduling information, but backhaul latency makes this information stale. Even a single transmission time interval (TTI) of delay can reduce CBF-SLNR performance below the uncoordinated baseline, because the precoder suppresses interference toward users that are no longer active. Coordination on stale information is therefore worse than no coordination at all. To address this, we propose a two-stage predictive framework in which a Spectral Temporal Graph Neural Network (StemGNN) predicts future user equipment (UE) scheduling states from delayed historical observations, and the predictions replace stale inputs to the CBF-SLNR precoder. Evaluated on a three-cell massive MIMO downlink with 60 UEs and 64 antennas per base station under Quadriga Urban Micro (UMi) channels and a proportional fair scheduler, StemGNN achieves a mean scheduling prediction accuracy of 87.57%, outperforming LSTM, GRU, Simple RNN, and Markov chain baselines at all evaluated horizons, with gains of up to 7.71% over LSTM at longer horizons where inter-UE structural dependencies dominate over temporal autocorrelation. When integrated into coordinated beamforming, the predictions recover 57-73% of the sum rate loss caused by one TTI of backhaul delay, improving sum rate by 9.58-14.35% over the no-prediction baseline and recovering up to 83% of the Lag-1 fairness loss for cell-edge users, with fairness gains persisting at higher lag values where throughput gains diminish. These results show that treating backhaul latency as a spatio-temporal forecasting problem is an effective approach for robust inter-cell coordination in delay-constrained networks.

Federated Learning (FL) empowers multiple clients to collaboratively learn a model, enlarging the training data of each client for high accuracy while protecting data privacy. However, when deploying FL in real-time edge systems, the heterogeneity of devices among systems has a severe impact on the performance of the inferred model. Existing optimizations on FL focus on improving the training efficiency but fail to speed up inference, especially when there is a latency constraint. In this work, we propose Collate, a novel training framework that collaboratively learns heterogeneous models to meet the latency constraints of multiple edge systems simultaneously. We design a dynamic zeroizing-recovering method to adjust each local model architecture for high accuracy under its latency constraint. A proto-corrected federated aggregation scheme is also introduced to aggregate all heterogeneous local models, satisfying the latency constraint of different systems with only one training process and maintaining high accuracy. Extensive experiments indicate that, compared to state-of-the-art methods and under a latency constraint, our extended models can improve the accuracy by 1.96% on average, and our shrunk models can also obtain a 3.09% accuracy improvement on average, with almost no extra training overhead. The related codes and data will be available at https://github.com/ntuliuteam/Collate

Production LLM agents often waste latency and reliability by regenerating code for the same procedural steps on every request. We replace this inference-time coding loop with an agentic tool-making pipeline that compiles repeated SOP steps into validated, versioned tools before deployment. The tool-maker grounds synthesis in the live environment as it collects execution traces, observes backend schemas and values, generates candidate tools, and repairs them against labeled cases. At runtime, the production agent calls these tools directly and falls back to code generation only when needed. We deploy the approach in a Fulfillment Center alarm-triage system, where an agent diagnoses alarms against a 44-node SOP over heterogeneous metric backends. In production, tool calls reduce p50 latency by 42%. On 1,500 historical alarms, they reduce end-to-end error rate by up to 53% by suppressing run-to-run variance in repeated steps. Because tools return compact structured verdicts, they also enable a simpler direct-call architecture, reducing p50 latency by a further 62% in a controlled ablation. Versioned tools also improve auditability and expose specification gaps and upstream data drift. Our results show that self-evolving agents can make industrial LLM systems faster, more reliable, and easier to operate.

Online truckload bid acceptance is a closed-loop stochastic decision problem in which a carrier or broker must, in real time, accept or reject a tendered load subject to operational feasibility, fleet repositioning costs, and opportunity cost against future demand. Public, reproducible benchmarks for this problem are scarce: existing routing benchmarks are static, while dynamic-fleet studies typically rely on private operator data. We introduce FreightBidBench, a public-calibrated, dependency-free, closed-loop benchmark in which feasibility (pickup reach, appointment windows, simplified hours-of-service, stochastic yard delays) and economics (service-failure penalty, terminal fleet value, daily price-premium window) are explicit, versioned, and reproducible from public Freight Analysis Framework and U.S. Department of Agriculture truck rate data. We develop two full-horizon hindsight ceilings: a simple LP style relaxation and a tighter Lagrangian-per-truck information relaxation that retains per-truck hours-of-service and sequencing structure and is 20.7% tighter than the LP relaxation on a tight-capacity scenario and 39.3% tighter on a scarce-capacity scenario. We introduce a parametric surrogate-rollout cascade with boundary-band and scarcity-pressure escalation triggers. On ten-seed tight and scarce scenarios, the best simple policy retains 91.0% and 86.5% of rollout profit and the standard-library surrogate 94.2% and 89.3%; a cascade at a single escalation band recovers roughly 98% on both at 40-56% of rollout's mean decision latency, and on the tight scenario is statistically indistinguishable from the rollout teacher (paired-bootstrap 95% CI on the profit delta spans zero).

Deep learning applications have been widely adopted on edge devices, to mitigate the privacy and latency issues of accessing cloud servers. Deciding the number of neurons during the design of a deep neural network to maximize performance is not intuitive. Particularly, many application scenarios are real-time and have a strict latency constraint, while conventional neural network optimization methods do not directly change the temporal cost of model inference for latency-critical edge systems. In this work, we propose a latency-oriented neural network learning method to optimize models for high accuracy while fulfilling the latency constraint. For efficiency, we also introduce a universal hardware-customized latency predictor to optimize this procedure to learn a model that satisfies the latency constraint by only a one-shot training process. The experiment results reveal that, compared to state-of-the-art methods, our approach can well-fit the 'hard' latency constraint and achieve high accuracy. Under the same training settings as the original model and satisfying a 34 ms latency constraint on the ImageNet-100 dataset, we reduce GoogLeNet's latency from 40.32 ms to 34 ms with a 0.14% accuracy reduction on the NVIDIA Jetson Nano. When coupled with quantization, our method can be further improved to only 0.04% drop for GoogLeNet. On the NVIDIA Jetson TX2, we compress VGG-19 from 119.98 ms to 34 ms and even improve its accuracy by 0.5%, and we scale GoogLeNet up from 20.27 ms to 34 ms and achieve higher accuracy by 0.78%. We also open source this framework at https://github.com/ntuliuteam/ZeroBN

Large language models deployed in open-world applications require safety guardrails that are both robust to complex risks and efficient enough for low-latency runtime moderation. Existing guardrails face a practical trade-off between lightweight classification-based models, which are efficient but often struggle with concealed intent, ambiguous semantics, and borderline safety decisions, and reasoning-based guards, which improve judgment quality but introduce additional token generation and inference latency. We present DT-Guard, a content safety guardrail model based on a Reasoning-Active Training, Reasoning-Free Inference paradigm. The key idea is to use reasoning supervision during training while emitting only structured safety labels at inference time. DT-Guard formulates safety judgment as a progressive decision process, Intent - Category - Safety, and constructs an intent-driven dataset with intent labels, risk categories, safety labels, and structured reasoning trajectories. To further improve hard-case robustness, we propose Rollout-Guided Progressive Hard-Case Optimization (RG-PHO), which uses multi-rollout consistency to identify stably mastered, persistently failed, and preference-unstable samples, and applies targeted supervised and preference optimization accordingly. At inference time, DT-Guard directly generates structured labels without explicit reasoning traces, preserving deployment efficiency. Experiments on prompt-side and response-side safety benchmarks show that DT-Guard achieves average F1 scores of 0.886 and 0.870, respectively. With only a 4B backbone, it reaches a dual-side average F1 of 0.878, outperforming strong 8B guardrail baselines. These results demonstrate that reasoning supervision can be effectively internalized into low-latency safety discrimination.

LLM serving optimization typically benchmarks many configurations and reaches for heavy profilers when latency targets are missed. We argue for the reverse discipline: estimation is the analytical layer of profiling -- without it, optimization degenerates to grid search. Floor First is a residual-driven triage workflow. Each decode step is modeled as a five-dimensional resource vector (HBM bytes, FLOPs, network bytes, network messages, KV capacity); summing within a resource and maximizing across resources gives an optimistic floor, the plain sum a pessimistic one. Where a measurement lands inside this [max, sum] interval reads out overlap quality before any profiler is opened, and profilers escalate only on residuals above a stated threshold. Deployment alternatives are compared by wall ordering -- which resource wall binds first as load grows -- rather than by point benchmarks. The account is compositional: new attention or state-space variants enter by declaring one module, and the workflow ships as a zero-dependency calculator plus an agent skill that enforces the discipline in agentic optimization loops. As a case study we analyze a DeepSeek-V3.2-style 671B MoE/MLA model on 16 NVIDIA H20 GPUs, whose ridge point of ~74 FLOP/byte (vs ~590 for H100) makes it an extreme decode-oriented part. The floors show TP16 decoding is KV-capacity-limited to ~70 concurrent 8K requests; sparse attention removes the KV-bandwidth term but not the capacity wall; an EP16+DP-attention layout accepts slightly worse same-batch weight traffic for an order-of-magnitude higher capacity wall (~644) -- while single-stream latency favors TP by 2.4x. The layout judgment is thus a computable function of the operating point, explaining why production deployments on identical hardware have shipped opposite attention layouts.

We present Wan-Streamer v0.2, a latency-preserving upgrade of the native-streaming, end-to-end audio-visual interaction model. v0.2 keeps the v0.1 modeling formulation, but raises the interactive output stream from 192x336 to 640x368 while preserving approximately 200 ms model-side signal-to-signal latency at 25 FPS. The higher-resolution stream supports scene-grounded mid-shot agents whose posture, gaze, hands, nearby objects, and local scene layout remain legible during real-time conversation. To support the larger visual stream without adding user-visible delay, v0.2 keeps the thinker as a single-GPU low-latency path for streaming perception, the short language/state Transformer pass that builds the generation cache, and final decoding. The performer becomes a multi-GPU Ulysses-style context-parallel group for the expensive next-unit latent generation. Each performer rank writes incoming K/V into a pre-sharded local cache. The long high-resolution latent video sequence is split across ranks for denoising and gathered through Ulysses communication, while the much shorter audio latent sequence is generated without sequence sharding. In this split, the thinker's language/state computation reaches the performer only as K/V conditioning, so no separate language sequence has to be communicated inside the performer group. This concentrates additional hardware on visual generation while preserving the compact thinker-performer boundary, keeping total remote interaction latency at approximately 550 ms when a 350 ms bidirectional network budget is included.

Extended Reality (XR) wearables require always-on perception within tight power envelopes of a few watts and motion-to-photon latency budgets below 20 ms, leaving only a few milliseconds for neural-network inference. Bit-serial computing is attractive for such energy-efficient neural network acceleration, but many existing architectures still process all bits even when ReLU sets the final output to zero. This paper presents BitFair, a software-hardware co-designed bit-serial CNN accelerator with learnable bit-level early termination and adaptive bit ordering, working under the ultra-low-power and strict latency requirements of XR applications. BitFair exploits dynamic bit-level sparsity by learning per-layer thresholds that trigger early termination when partial sums reliably predict that the final ReLU output will be zero. Furthermore, it searches for layer-wise bit orders that prioritize informative bits, maximizing early termination without sacrificing accuracy. A GlobalFoundries 12nm FinFET implementation with a core area of 0.34 mm^2, 104 KB on-chip memory, and voltage scaling from 0.55 to 0.70 V achieves sub-millisecond latency, up to 117.0 BTOPS/W, and 0.07 pJ/SOP. On IBM DVS128 Gesture and N-MNIST, BitFair achieves 96.5% and 97.7% accuracy, respectively, while improving effective energy efficiency by 4.0-22.1x and accuracy by up to 9.2% over prior fabricated XR vision accelerators.

Large Language Model (LLM) inference workloads are a rapidly growing contributor to data center energy consumption. Optimizing these deployments requires matching specific LLMs to the most efficient GPUs, but operators currently lack the tools to do so without exhaustively profiling each combination. While some predictive models exist, they still require profiling data and struggle to generalize to hardware unseen during training. To address this, we introduce \textit{WattGPU}, featuring two predictive models for mean GPU power draw and Inter-Token Latency (ITL). Our approach leverages only publicly available LLM metadata and GPU specifications, eliminating the need for hardware access or profiling while enabling generalization to unseen NVIDIA server-grade GPUs and LLMs. We evaluate our models using rigorous leave-one-GPU-out and leave-one-LLM-out cross-validation on a dataset of 42 open-source LLMs (0.1B--27B parameters) and 8 GPUs under both offline and server scenarios. The mean power draw model achieves a median absolute percentage error of $\leq3.4\%$ for offline and $\leq13.5\%$ for server scenarios on unseen GPUs, while the latency model achieves $\leq8.5\%$ in server mode, both maintaining strong GPU ranking correlations for server scenarios (Kendall $τ\geq0.76$). Compared to standard physically grounded baselines -- Load-Scaled Thermal Design Power (TDP) for power draw and roofline for latency -- our models reduce median absolute percentage error by approximately 4$\times$ on unseen LLM-GPU combinations for server scenarios or approximately 2$\times$ for completely unseen GPUs. WattGPU's data and code are publicly available at https://github.com/maufadel/wattgpu.

Real-time interaction models -- Moshi, MiniCPM-o, Qwen-Omni -- turn serving into a periodic real-time task: on every frame a session ingests streaming audio and must respond by a recurring wall-clock deadline, while its KV cache grows monotonically and stays pinned for the whole conversation. This regime hides a dangerous failure mode. On a real full-duplex stack, sustained load does not degrade serving gracefully: it falls off a cliff, jumping in one step from milliseconds per frame to a stalled engine when accumulated session state exhausts the KV pool. The collapse is metastable -- identical five-minute runs collapse or survive on run-to-run variance -- and silent: latency and deadline-miss metrics read healthy throughout. We show one move restores both stability and observability: bound each session's resident state, and latency starts telling the truth. Metronome's in-engine KV window eliminates the collapse (0/20 vs. 14/20 runs across two batches) and turns per-frame latency into a monotone load signal, on which an online admission controller discovers the schedulable concurrency; without the window, the identical controller over-admits into the wall. A first-order model predicts the collapse time within a few percent on the headline model, and a quality probe validates the bound's design by ablation: the window alone is quality-free in turn-based decoding, and its few pinned attention-sink tokens are what keep free-running generation healthy. Everything is measured end-to-end on real audio, across four interaction models on one GPU.

Large language models (LLMs) are increasingly deployed in domains requiring guardrails to detect unsafe, off-topic, or adversarial prompts. Existing guardrails predominantly rely on fine-tuning to build classifiers, which often suffer from low generalization and high inference latency. We present kNNGuard, a training-free guardrail that utilizes the activation space of an off-the-shelf LLM. Given a small bank of 50 safe and unsafe prompts, kNNGuard extracts hidden activations and performs multi-layer kNN fusing activation-space and embedding-space scores for classification. Across six domains spanning topical and security prompts, kNNGuard achieves competitive or superior F1 compared to fine-tuned state-of-the-art guardrails while running 2.7x faster than the best comparable guardrail, and 10x faster than a fine-tuned safety classifier without gradient updates or fine-tuning. Domain adaptation requires only updating the labeled bank, which can be constructed in under 10 seconds and several orders of magnitude faster than established guardrails. We also analyze the impact of system prompts, layer selection, and integration into production LLM pipelines as a configurable, low-latency guardrail.

Modern automotive software architectures comprise large sets of mixed-criticality functions executing on shared multi-core platforms with strict real-time and end-to-end timing requirements. Sensor-to-actuator data propagation in such systems is typically expressed via cause-effect chains with worst-case data-age budgets. Job-level dependencies (JLDs) have been introduced to provide a schedule-agnostic mechanism for bounding the data age independently of the underlying scheduler. The state-of-the-art methods for synthesizing JLDs, however, do not check whether the produced JLDs are enforceable under a concrete scheduling policy or jointly schedulable at the system level. In this paper we propose the first machine-learning-based JLD synthesis method, built around a two-level Graph Neural Network with temperature-controlled sampling that learns the structural patterns connecting cause-effect chain configurations to their JLD solutions. Since learned outputs may not be correct by construction, we embed the GNN in a novel Generate-and-Verify architecture in which a safe DP data-age checker, together with a per-chain EDF feasibility checker and a system-level demand-bound test, accept or reject each candidate. We show that the ML-based generator substantially outperforms the original greedy heuristic while achieving orders-of-magnitude lower synthesis time, demonstrating that learned structural priors can effectively replace exponential propagation-tree enumeration on this class of real-time scheduling problems.

Communication systems designed for reliable data reconstruction, rather than task-oriented communication, typically rely on separate source and channel coding and incur high latency under limited spectrum availability and fading channels. To address this, we propose a transmission framework with opportunistic spectrum access, in which the transmitter sends discrete latent representations learned via a vector-quantized variational autoencoder (VQ-VAE) over idle licensed channels using standard digital modulation. The AI-powered receiver is still able to reconstruct task-related information from the heavily compressed data. We develop a cross-layer latency model that accounts for compression, block errors, retransmissions, and stochastic channel access. Results on latency-accuracy trade-offs show that the proposed scheme achieves at least 79- and 3.3-fold latency reductions with only 5.7% and 2.4% drops in classification accuracy compared to benchmarks using conventional source and channel coding. The framework enables low-latency communication and reliable task execution even under limited spectrum availability and challenging channel conditions.

Real-time binaural speech enhancement is constrained by latency, computational cost, and inter-device communication, yet existing efficient solutions predominantly address single-channel settings. In this paper, we introduce RT-Tango, a real-time distributed binaural speech enhancement framework designed for streaming on resource-constrained platforms and specifically for hearing aids. RT-Tango relies on a two-stage distributed architecture combining perceptually motivated ERB feature compression, lightweight grouped recurrent mask estimation, and temporal sparsification to reduce computational cost. Stringent latency constraints are addressed by decoupling spectral resolution from algorithmic delay using an asymmetric STFT, together with causal recurrent inference and online estimation of spatial statistics. Experimental results show that RT-Tango achieves competitive speech enhancement while significantly reducing MACs operations and functioning at ultra-low latencies as low as 8 ms.

We present NPUsper, a live transcription system that makes Whisper efficient on mobile NPUs by eliminating redundant computation. To avoid the heavy padding used by prior streaming systems, NPUsper detects hallucinated tokens online from temporal patterns in decoder cross-attention, allowing each inference round to process short audio inputs with minimal carryover. For efficient mobile-NPU execution, we propose controlled unrolling, which executes autoregressive decoding as K-step chunk graphs, removing unnecessary KV-cache computation and reducing graph-dispatch overhead. NPUsper achieves up to 4.84x lower per-word latency, up to 33.2x lower time-to-first-token (TTFT), and up to 88.64% lower average power consumption compared with baselines, while maintaining comparable transcription accuracy. The code is available at https://github.com/npusper/NPUsper.

Image-guided therapies, including radiotherapy, biopsy and deep brain stimulation, rely on real-time targeting of anatomical structures. However, in the presence of motion, imaging latencies create a temporal misalignment between observed and true anatomy, compromising treatment accuracy. Artificial intelligence-based frameworks have increasingly been presented to close this latency gap, but leading personalised models can fail due to a lack of stable anatomical grounding. Foundation models can provide grounded behaviour, but they do not adapt to real-time, individual patient dynamics. Here we introduce a closed-loop coupling framework that synergises patient-specific temporal prediction with continuous segmentation-based anatomical interpretation from a foundation model. A personalised model predicts future anatomy to compensate for system latency, while a streaming foundation model provides anatomical supervision used to continuously update the temporal predictor in real time during treatment. We validate the framework using a digital phantom and intrafraction magnetic resonance imaging (MRI) from patients undergoing MRI-guided radiotherapy. For a prediction horizon of 400 ms, the proposed method improves anatomical prediction and reduces dosimetric error compared with existing approaches, within clinically relevant latency constraints. These results establish closed-loop coupling as a general strategy for real-time image-guided intervention.

The emerging techniques of semantic communications and edge computing in 6G networks necessitate a paradigm shift toward co-designed semantic-aware and adaptive resource allocation for short-packet transmissions. However, there is a fundamental gap between the semantic layer and the physical layer under low-latency finite blocklength (FBL) effects. To bridge this gap, we introduce the Quantized Semantic Age of Information (QSAoI), a novel metric that rigorously captures the trade-offs among freshness and semantic efficiency of high-level features in real-time communication in the FBL regime. Guided by this metric, we propose a novel foundation model-based efficient co-designed framework to minimize the expected QSAoI over wireless fading channels in latency-constrained semantic communication. Specifically, we formulate a non-linear joint optimization problem to dynamically optimize the block-wise mixed-precision quantization (MPQ) strategy and the physical blocklength. To efficiently resolve this complex problem, we develop a high-efficiency low-complexity algorithm based on fixpoint inspection and bisection search. Extensive simulations validate that our proposed algorithm dynamically adapts the semantic quantization precision to varying channel conditions, effectively minimizing the expected QSAoI compared to baselines.

Edge-cloud inference collaborations are often designed with a routing estimator that decides whether to offload each frame from weak models at the edge to stronger models in the cloud. Existing systems place the routing estimator after the weak detector, so the weak forward pass still runs even on frames that are later offloaded. In this paper, we argue that this weak-conditioned design can be suboptimal when the offload budget varies. First, we present a competitive weak-skipping estimator (0.153 GFLOPs, about 29x lighter than the weak detector at 4.49 GFLOPs) that extracts routing signal from raw pixels, outperforming the common after-weak placement weak-conditioned baselines. Second, we show that neither weak-skipping nor weak-conditioned placement dominates across the full operating curve, and we propose budget-adaptive routing, which selects between them by offload budget via two offline-tuned thresholds. On PASCAL VOC, our budget-adaptive router traces the upper accuracy envelope of both fixed placements across the operating range. Our method reduces per-frame latency by up to 19.1 ms (about 30% lower at rho = 0.9). Besides outperforming SOTA methods, it is surprisingly stronger than the strong model (+1.7 pp over the strong model's peak mAP) at some operating points with far less compute. Artifacts are available at https://github.com/ViGeng/bgt-ada

Coding agents are rapidly becoming a major application of agentic LLMs, but serving them efficiently remains challenging. Progress on this challenge requires understanding real workload patterns, yet the data needed for such analysis is largely absent. Existing public traces and benchmarks do not capture real, day-to-day coding-agent usage across multiple agents and model families for serving-system analysis. To help fill this gap, we collect and release a trace of roughly 4,300 coding-agent sessions, containing about 350,000 LLM steps and 430,000 tool calls from our own day-to-day use of Claude Code and Codex. Our analysis shows that coding-agent workloads feature long autonomous loops, long contexts with short outputs, diverse and heavily-tailed tool calls, and high but imperfect prefix cache hit rates. These findings point to concrete opportunities for optimizing serving, including lower-overhead tool calling, append-length-aware prefill, semantic-aware tool-latency prediction, and improved KV-cache management around human-paced gaps. We release the dataset, trace collection pipeline, and analysis code at https://github.com/uw-syfi/TraceLab.git the project website is https://tracelab.cs.washington.edu.

Inference-time safety methods for large language models have proliferated, yet no systematic comparison exists. We evaluate five defense paradigms (no defense, static steering, CAST, AlphaSteer, probe-gated) across seven instruction-tuned models (7-31B) and five attack types (GCG, AutoDAN, DeepInception, prefilling, intent laundering). Our central finding: prompt-time activation defenses are structurally blind to prefilling attacks. AlphaSteer achieves 0% attack success on GCG, AutoDAN, and intent laundering but 50% on prefilling. We prove a corollary: any defense that gates intervention on a single layer's activation alignment with a benign reference (cone, subspace, or null-space) is blind to attacks that craft activations to lie inside that reference, whether checked at prompt time or per token. As its constructive contrapositive we introduce response-time probing: a linear probe on the model's hidden state at the first generated tokens, with AUROC 0.97-1.00 across all seven models. Combined with a halt, it cuts prefilling attack success to 0/40 on every model with 0% benign false positives, outperforming Llama Guard 3. Cross-template generalisation depends on probe depth, so we scope the claim to the canonical prefilling-template family. Composing the response-halt with AlphaSteer's null-space steering gives an orthogonal split (the halt catches prefilling, AlphaSteer catches semantic attacks), reaching defense success 0.983 on Mistral and 0.994 on Llama and dominating both components. We further show MMLU fails to capture steering's true utility cost, which appears as behavioral hedging rather than factual loss, and that diverse negative training sets cut probe false positives from 80-100% to near zero. Code, attacks, per-sample results, and the judge prompt are released.

Vision-language models and vision-language action models endow the robot with unprecedented capabilities. However, the input of video and high-resolution images yields a massive number of visual tokens, leading to extremely high inference latency and severely hindering the robot's real-time control. To break through this computational bottleneck, we propose ST-Merge, a plug-and-play, training-free framework that efficiently fuses redundant tokens directly during the visual encoding phase. By explicitly constructing 3D spatiotemporal coordinates, it employs a multi-queue parallel matching and weighted aggregation mechanism to achieve efficient and geometrically consistent fusion of redundant tokens across frames. In addition, we introduce a post-merge positional correction mechanism that effectively eliminates spatial deviation caused by merging by dynamically re-evaluating the rotational position code of the weighted centroid of the vision token, thereby ensuring the high-precision spatial awareness required for dexterous operation. In the Video Question Answering task on the mainstream VLM, Qwen2.5-VL, ST-Merge achieves a 2$\times$ inference speedup with only a tiny 1% loss in precision. When deployed on the $π_{0.5}$ VLA policy, ST-Merge achieves an 8.3$\times$ speedup at 1024 $\times$ 1024 resolution and matches the baseline success rate at this high-resolution setting. At lower resolutions, it introduces a small drop in accuracy.

Low-Earth orbit (LEO) satellite Internet has become an indispensable infrastructure that provide growing coverage for global users. Despite extensive measurement efforts, the principles underlying region-level performance characteristics remain insufficiently understood, limiting the ability to identify region-specific latency signatures under dynamic network conditions. In this paper, we formulate the problem of region-level latency characterization using Starlink round-trip time (RTT) measurements from the public LENS dataset. We then propose a hierarchical analytical framework that transforms raw RTT sequences into multi-scale statistical features for cross-region comparison. Using data from five geographically representative regions, we demonstrate that latency differences are strongly associated with deployment factors, particularly infrastructure availability and Starlink dish-to-Point-of-Presence distance. Mutual information analysis identifies minimum RTT as the most discriminative feature, which is further supported by XGBoost-based feature importance. The proposed model well achieves 83% accuracy on short-term data. However, its performance degrades over longer periods, indicating limited temporal generalization and motivating the need for adaptive models and feature representations for long-term performance in the future.

Diffusion language models (DLMs) have recently emerged as a promising alternative to conventional autoregressive language models. By generating multiple tokens in parallel during each denoising step, they offer higher inference throughput while maintaining competitive quality. However, realizing these throughput gains while meeting latency SLOs in a serving system requires addressing challenges introduced by DLMs' unique characteristics. These include navigating the speed-quality tradeoff created by confidence-based denoising, choosing appropriate parallelization levels across model instances under fluctuating load, and coordinating approximate KV caching mechanisms that introduce non-uniform per-step costs. To address these challenges, we present DiLaServe, a cluster-level serving system for DLMs. DiLaServe enables deadline-aware scheduling and adaptive load control through confidence-threshold adjustment, and dynamically reconfigures the cluster by solving a quality-aware optimization problem, while explicitly modeling the step-level heterogeneity introduced by approximate KV caching. Across multiple benchmarks and real-world traces, DiLaServe improves SLO attainment by up to 56.6 percentage points and reduces end-to-end request latency by up to 46% while incurring less than 1% accuracy drop.

Embodied foundation models have recently been widely used to improve robot generalization and task success rates. Previous works apply lossy efficient-inference techniques such as quantization, pruning, and asynchronous inference, accepting small action quality degradation in exchange for lower per-step computation cost and inter-action latency. However, unlike traditional static ML tasks, embodied tasks involve repeated interaction with the environment, and task-level performance is determined not only by per-step cost, but also by closed-loop effects unique to embodied execution, which remain insufficiently characterized in current efficient-inference studies. In this work, we propose TISED (\underline{T}ask-level \underline{I}nference \underline{S}peedup \underline{E}ffect \underline{D}ecomposition), an analytical framework that unifies diverse lossy inference optimization techniques and decomposes their effects on static and dynamic tasks, and uncovers some paradoxical effects on task-level performance: (1) on \textit{static tasks}, optimization sometimes can lengthen end-to-end per-task completion time even as per-step latency drops; (2) on \textit{dynamic tasks}, moderate lossy optimization can raise task success rate even above the baseline; and (3) the monotonicity and sweet-spot location of both effects can shift with hardware configuration. Together, our findings provide a new perspective on adapting inference optimization techniques to embodied tasks.

Matrix-orthogonalization-based optimizers, exemplified by Muon, have demonstrated strong convergence behavior across a wide range of modern deep learning workloads. The matrix-aware updates offer a compelling alternative to conventional element-wise optimization, particularly as model architectures continue to grow in scale and heterogeneity. Yet contemporary distributed training infrastructure built around the assumption of element-wise optimizers is poorly matched to matrix-level optimizers such as Muon, whose updates couple entire weight matrices and require costly Newton-Schulz iterations. Vanilla Muon implementations incur more than 2x the cost of forward and backward passes. To close this gap, we present DMuon, an open-source distributed Muon implementation that integrates into existing training pipelines as a drop-in module, with no framework-level modifications. Across both embodied foundation model and large language model (LLM) training workloads, DMuon achieves a 1.48x-3.01x speedup in end-to-end step time and a 6.85x-163.00x speedup in optimizer-step time, bringing per-step latency to near-AdamW levels and enabling efficient scaling in our model training.

Streaming video understanding (SVU) must answer queries that arrive asynchronously while visual tokens stream continuously under strict GPU-memory and query-time latency budgets. A key challenge is delayed query: decisive cues may appear briefly, yet many subsequent updates occur before the query arrives, increasing the risk that those cues are evicted or diluted under bounded memory. We propose ProtoKV, a constant-footprint SVU memory that represents far history as a fixed-capacity summary state rather than retaining token instances. ProtoKV keeps an exact near-window KV cache and aggregates older content into a semantic-spatial prototype bank with residual statistics. At query time, each prototype is exposed through a bounded pseudo-token interface that is drop-in compatible with standard attention. Under matched budgets and comparable query-time cost, ProtoKV improves accuracy by up to 12.5 points over token-retention baselines on SVU benchmarks in the long-delay regime, with gains that grow as query delay increases.

A mesh of sovereign agents has no center: no shared clock, no shared model, and no coordinator to gather data or retrain. Its competence rests on each agent folding the projections its peers emit into a single internal state, online, from observations that arrive at irregular, unscheduled times, on a substrate whose weights it cannot retrain. Any one of these constraints is tractable on its own; folding optimally under all three at once is not. We ask what such a substrate must be, and prove two necessary conditions from one model of a self-evolving latent observed at irregular, exogenous times. Because the latent changes, its optimal estimator is time-varying: an adaptive timescale is necessary, and every fixed-gain filter is strictly suboptimal. And because arrivals are clock-free, the optimal estimate depends on the elapsed gap between them, which no gap-blind network recovers at any width or depth. This second condition is capacity-independent: scale cannot substitute for the missing dependence. The two conditions intersect in the continuous-time liquid class. An LSTM satisfies the first, a fixed continuous-time filter the second, and a multi-timescale liquid network both. Synthetic experiments confirm each: the network attains the timescale, and the separation is computed exactly. The characterization is necessary, not sufficient, and binds fixed-weight substrates: a network free to retrain reaches the class by other means. Proved per agent, the necessity binds every agent of a mesh, a structural condition on mesh intelligence.

Existing object-aware SLAM systems force a trade-off between real-time performance, multi-class support, and the generation of high-fidelity, semantically coherent object models. To address this trade-off, we present DSP-SLAM++, which extends the DSP-SLAM framework with an asynchronous mapping pipeline for real-time performance and dedicated sensor fusion adaptations for a monocular fisheye-LiDAR suite. Experiments demonstrate that our system generates fine-grained, geometrically-complete shapes for multiple object classes while eliminating severe mapping thread bottlenecks by reducing maximum object processing latency by up to 70% compared to the state-of-the-art baseline, enabling robust, real-time performance on a challenging 25 Hz multi-class datasets. This work makes high-fidelity, multi-class object SLAM more practical for real-world applications like autonomous driving and robotic manipulation by enabling its use on platforms with common fisheye-LiDAR sensor setups. The open-source code is available at: [github.com/AUBVRL/DSP-SLAMpp].

Different real-time speech applications impose distinct latency budgets, often requiring separately trained enhancement models for each scenario. In this paper, we propose a one-for-all, real-time universal speech enhancement model that provides explicit control over both algorithmic and computational latency. Algorithmic latency is flexibly adjusted via configurable look-ahead frames. To avoid learning inefficiency caused by varying padding configurations, we introduce parallel convolutional layers corresponding to different look-ahead settings. Computational latency is controlled through an early-exit mechanism, enabling inference at different network depths. To narrow the performance gap between specialized and flexible models, we propose a two-stage training strategy with a shared-to-multiple decoder transition. Overall, the proposed framework enables a single model to be deployed across diverse latency budgets without retraining separate models. Model weights are available for download at: https://huggingface.co/nvidia/Real-time_RE-USE

This paper presents an energy-efficient hardware acceleration of the convolutional layers in the U-Net architecture for image segmentation, implemented on FPGA. While digit-serial arithmetic, particularly most-significant-digit-first (MSDF) techniques, offers a compact hardware footprint, it suffers from initial latency before producing the first output digit. This delay accumulates in cascaded operations like multiplication followed by addition, where each unit introduces its own startup overhead. To overcome this, we propose a merged multiply-add (MMA) architecture that fuses these operations into a unified pipeline. Instead of incurring separate delays, the MMA introduces a single streamlined latency per iteration, shorter than the combined latency of conventional cascaded units, resulting in enhanced throughput and efficiency. The MMA units are designed to process spatial input depths in parallel, achieving significantly higher performance than both standalone MSDF-based and conventional designs. We evaluate the proposed design using U-Net as a target application. Despite operating at a lower frequency than a CPU, the FPGA-based accelerator achieves up to an order of magnitude higher energy efficiency, delivering up to $15.14$ GOPS/W compared to $1.93$ GOPS/W for CPU-based inference. The design also shows approximately $9\times$ reduction in energy consumption compared to MSDF-based FPGA implementations. These results highlight the efficacy of the merged arithmetic approach for resource-constrained, latency-sensitive edge applications in medical imaging and computer vision.

Tactile sensing is essential for humanoid robots to achieve safe physical interaction, dexterous manipulation, and truly human-like responsiveness. However, the design of such systems remains challenging. Conventional approaches often suffer from complex multilayer structures, intricate wiring, high cost, and poor scalability, making it difficult to realize full-body tactile sensing with real-time, low-latency detection while maintaining minimal computational load on the robot's main processor. In this work, we present a simple, scalable and hardware friendly tactile sensing system for a companion humanoid robot based on the self-capacitance principle. The proposed sensor system employs a single conductive fabric layer with a conductive fabric wire architecture and does not require intricate electrode patterning. Scalability was demonstrated by fabricating a 100-point sensor array on a flexible printed circuit (FPC). Evaluation across sampling frequencies showed that 10 Hz is insufficient and misses transient events, whereas 100 Hz and 1000 Hz reliably capture and clearly distinguish all interaction types: gentle touch, slow tapping, fast tapping, and hitting. A decision-tree classifier was implemented directly on the FPGA, offloading real-time inference from the Raspberry Pi 4 with minimal latency and negligible power overhead. This design fully meets the tactile sensing requirements of the HIRO-chan robot and is well-suited for full-body tactile sensing in HIRO-chan and other companion robots.

Real-time texture transfer expands the creative horizon for interactive applications, enabling seamless detail projection in scenarios that range from digital character cosmetics to procedural automotive texturing. Yet, its practical application is governed by inherent trade-offs between processing speed and suppression of artifacts. Low-latency transfer methods frequently fail to suppress artifacts, and robust alternatives rely on large-scale models that are costly in training and memory. Our proposed method bridges the gap between efficiency and robustness by using a cage-based geometric filtering method to identify Non-Cosmetic Zones (NCZs) for artifact suppression. While other models are resource-intensive and require multiple days of training on manually annotated datasets, we are able to successfully suppress artifacts and achieve immediate deployment on consumer-grade hardware. Our framework achieved highly efficient runtimes of ~70ms on mobile devices for a ~4.8k triangle mesh.

Federated split learning (FSL) enables collaborative training across bandwidth-constrained IoT devices, but repeated activation and gradient exchange creates a communication bot-tleneck. Prior work optimises either activation compression or synchronisation frequency in isolation. This paper presents an FSL framework for IoT rainfall prediction that jointly regulates activation compression and the synchronisation interval \r{ho} via a latency driven scheduler on a server with per client EMA smoothing. The system is evaluated on hourly ERA5 data from 11 weather stations through a 17 scenario simulation matrix and a four scenario Raspberry Pi deployment over a real wide-area link. The simulation matrix validates scheduler switching across low, high, and mixed latency profiles, while the Pi deployment validates the high latency endpoint selected by the same policy. AUPRC varies only slightly across configurations (0.6381-0.6484 in simulation; within 0.011 on Pi), indicating that aggressive quantisation and sparser aggregation do not materially degrade predictive quality in this setting. On Pi, the selected endpoint (int8 with rho=3) achieves an 87% reduction in activation upload payload and a 54% reduction in synchronisation traffic relative to the float32 baseline, while reducing runtime jitter from +/-688 s to +/-10 s.

Adapting a streaming speech recognition model to a new language requires choosing between two plausible warm starts: a multilingual (ML) encoder or an English-only (EN) encoder. The common intuition is that the multilingual encoder should help most at low data, but it is unclear how long that advantage persists, whether tight streaming latency amplifies it, and whether it survives deployment quantization. We answer these questions with a controlled sweep of a 0.6 B-parameter cache-aware FastConformer transducer across eight European languages, up to five target-language data scales (100 h to 2500 h), three streaming tiers plus offline decoding, and up to four public test sets. The main result is that multilingual initialization is a data-limited advantage, not a latency-limited one. On FLEURS at 160 ms, the mean EN-ML word error rate (WER) gap falls from +4.21 percentage points (pp) at 100 h to +0.20 pp at 2500 h; a power-law fit summarizes this decay, with each doubling of target-language data roughly halving the remaining advantage. Across the three streaming tiers, the across-language mean EN-ML gap is approximately stable at each scale from 100 to 1000 h, and is near zero by 2500 h. Finally, 4-bit weight-only encoder quantization at the matched 560 ms streaming tier reduces the encoder footprint by about 3x, with an average FLEURS WER increase of about 0.5 pp. The resulting guideline is simple: use multilingual initialization in low-data regimes, treat the choice as effectively irrelevant at large data, and make latency and quantization decisions independently.

Remote photoplethysmography (rPPG) estimates the blood volume pulse (BVP) signal from facial videos, enabling contact-free health monitoring. Conventional clip-wise approaches, which use video clips as input, require capturing over one hundred frames before inference, thus introducing several seconds of delay and hindering real-time use. Meanwhile, frame-wise approaches struggle to capture long-range temporal and periodic features of physiological rhythms, and therefore lead to reduced estimation accuracy. To overcome these issues, we propose StreamPPG, a unified architecture that enables low-latency frame-wise physiological signal estimation while achieving competitive accuracy compared with clip-wise approaches. StreamPPG is trained under a consistent privileged learning (CPL) strategy, which leverages ground-truth rPPG signals as privileged information to enhance the model's representation capability. Extensive experiments demonstrate that StreamPPG achieves state-of-the-art accuracy across multiple datasets while maintaining real-time throughput on edge devices.

Outer-loop link adaptation (OLLA) is widely deployed in 5G NR to track channel variations, yet its reliance on first-order, single-bit feedback degrades performance significantly under high-mobility and fast-varying channels. This paper presents LOLLA (Learned Outer-Loop Link Adaptation), a deep reinforcement learning framework that replaces the conventional OLLA staircase with a learned, continuous SINR offset conditioned on rich PHY/MAC telemetry inaccessible to OLLA. The offset modulates the SINR-to-MCS lookup table, preserving 3GPP-compliant MCS selection and provably subsuming the conventional OLLA update rule. A Proximal Policy Optimization (PPO) policy trained under a Lagrangian block error rate (BLER) constraint automatically enforces tunable reliability targets from 1% to 15% without manual penalty calibration. The framework is realized as the first closed-loop AI-native control dApp on a GPU-accelerated 5G NR stack, achieving end-to-end control latencies under 500 microseconds. Evaluations under 3GPP TDL channel models demonstrate 15% to 92% throughput gains over OLLA across Doppler frequencies up to 400 Hz, while attaining a Pareto frontier that strictly dominates OLLA across all evaluated reliability targets. The learned policy generalizes to unseen channel models and scales to eight concurrent UEs under shared-resource scheduling. In the uplink formulation, the gNB directly observes decoding outcomes, enabling simulation-to-deployment parity.

Online learning is crucial for handling complex data streams in big data applications. Recent research has begun to focus on dynamic scenarios, i.e., non-stationary environments. However, a crucial yet often overlooked aspect is label latency, where new data may not receive labels in time due to the slow and expensive labeling process, thus hindering rapid adaptation to dynamic environments. To resolve this impasse, we propose Dual-Track Geometry Online Learning (DT-GOL), a novel framework that shifts from temporal compensation to spatial reasoning to bridge the supervised latency gap. By modeling the delay challenge as a semi-supervised task, we leverage real-time topological evolution of features as a reliable geometric surrogate for unobservable conceptual changes to achieve proactive supervised adaptation within the delay window. Unlike rigid self-training, we introduce a dynamic evidence calibration mechanism that distills geometric information into soft labels that perceive uncertainty, effectively mitigating the confirmation bias inherent in hard pseudo-labels. Furthermore, to resolve the stability-plasticity dilemma, we design a decoupled dual-track architecture in which a master learner serves as a stable anchor, updated strictly from delayed ground truth, while a transient branch leverages soft geometric knowledge for low-risk forward adaptation. Extensive experiments on real and synthetic datasets demonstrate that DT-GOL significantly outperforms existing state-of-the-art baseline methods, especially in scenarios with concept drift.

Agentic AI applications compose multiple model calls and tool executions, creating new scheduling challenges for GPU-CPU clusters. Their inference time and model-call structure often depend on prompt semantics, making conventional scheduling approaches ineffective for low-latency serving. This paper presents SwarmX, a system that implements agentic scheduling for low-latency agentic applications. SwarmX uses scheduling-specific neural predictors to capture prompt, device, runtime, and target-model features; exposes distributional predictions to routers and scalers for tail-aware decisions; and provides mechanisms for predictor training and online adaptation. These predictors and mechanisms are integrated into a scheduler-agent framework that provides a common substrate for integration with existing scheduling and model-serving infrastructure. We evaluate SwarmX using production deployment (nearly one thousand GPUs and one million CPU cores) and controlled experiments on a 128-GPU testbed. Across multi-agent code generation, deep research, and multimodal agentic workflows, SwarmX reduces tail latency by up to 61.5% compared to state-of-the-art schedulers and sustains up to 2x the throughput of production schedulers under the same SLO.

Mainstream LLM serving systems reuse prefix work mainly through paged or radix key-value (KV) caches. This is highly effective for high-throughput, high-concurrency serving, but it manages only one positional fragment of execution state: the KV cache. We study the opposite regime: low-latency, small-batch, on-device physical-AI serving, where interactive LLM agents, speech systems, and robot policies repeatedly branch, reset, interrupt, and re-enter under tight responsiveness budgets. We introduce execution-state capsules, a graph-bound checkpoint and restore mechanism for the complete restorable state at a committed boundary. FlashRT is a white-box, backend-facing kernel runtime whose evaluated NVIDIA CUDA backend runs captured graph plans over contiguous static buffers with no block-table indirection. Because the live state is a closed set of named buffers, a capsule can snapshot, restore, fork, or roll back the whole execution boundary, including KV, recurrent state, convolution state, MTP state, and metadata. This moves reuse from token-addressed KV fragments to graph-bound execution-state boundaries. On an RTX 5090, capsule restore is byte-exact at the stored-state level and token-identical under greedy decode. A KV-only ablation diverges, showing that recurrent state is load-bearing. GPU-resident snapshot and restore are sub-millisecond, and TTFT speedup over cold prefill grows from 3.9x at 2k tokens to 27x at 16k tokens. On Jetson AGX Thor and DGX Spark, the same correctness and structural properties hold. Capsules are not a replacement for high-throughput KV-cache serving; they define a complementary latency-first serving point for explicit execution-state reuse.

Learning-based planners for sidewalk navigation can generate diverse candidate trajectories in real time, yet their scoring functions often fail to select the best trajectory in challenging situations, outputting trajectories that make the mobile robot drive onto grass, toward pedestrians, or in the wrong direction, even when better candidates exist in the same set. We call this the trajectory scoring gap: in real-world sidewalk navigation, the gap between an anchor-based planner's top choice and the best possible candidate is substantial, likely due to limited high-level scene understanding capability of the planner. Rather than replacing the planner with an end-to-end Vision-Language-Action model, we propose a VLM-Planner interface that uses a VLM to select a candidate index from the planner's proposal set and then fuse it with the planner's initial output. However, VLMs take 1--3s per query and so cannot directly drive a 5--20Hz control loop. We contribute a training-free, latency-resilient trajectory-level fusion layer that turns a stale VLM selection into real-time planner scoring via geometric similarity with exponential decay. On $\sim$2,000 challenging real-world scenarios (e.g., junctions, pedestrian encounters), VLM selection achieves 30% ADE reduction versus the planner's best selection, while the planner remains competitive in routine situations. In simulation, Score Fusion maintains >80% success rate with delays up to 5s. We demonstrate the full system on a mobile robot navigating challenging campus sidewalks with varied network latency.

Streaming Retrieval-Augmented Generation (Streaming RAG) hides tool latency by issuing retrieval queries in parallel with the user's still-arriving input, before the utterance is complete. Speculation can only help, though, when the correct query becomes determinable before the user stops speaking or typing -- a property of the query, not the system. We name and measure this property, tool-intent stabilization: the point in the input stream at which a speculative query's retrieval converges on the answer-bearing result. On the CRAG benchmark (1371 validation questions) we (i) characterize how stabilization is distributed across queries; (ii) derive a model-agnostic bound H on the share of tool latency hideable behind the remaining input, given tool latency L and input cadence delta; (iii) validate it against a working streaming pipeline; and (iv) ask which query properties predict early versus late stabilization. Stabilization is typically early: at a realistic operating point a 73.9% streamable fraction of the benchmark admits latency hiding, and H acts as a conservative aggregate floor that realized savings meet or exceed -- though it does not predict savings query by query. Question type yields a statistically significant but small early/late split. The study needs no model training and runs on commodity CPU hardware; a dense-retriever replication confirms the early-stabilization effect is not a BM25 lexical artifact.

Streaming speech enhancement requires balancing algorithmic latency against quality, yet existing approaches largely treat this as a binary causal versus non-causal choice. LaCo-SENet addresses this issue with two mechanisms parameterized by a single training-time hyperparameter. First, asymmetric temporal padding redistributes past and future context in convolutions, enabling systematic latency configuration. Second, dual-buffer streaming combines state buffers for past context with lookahead buffers that supply future context at both the input and feature levels. Selective state updates also prevent future-frame leakage into the streaming state, ensuring training-inference consistency. On VoiceBank+DEMAND, a fixed-budget (1.37M parameters) backbone yields a family of models spanning 12.5-75.0 ms, with PESQ rising from 3.35 to 3.43. At just 12.5 ms (fully causal), a PESQ of 3.35 matches or exceeds the prior causal state-of-the-art (3.27 at 46.5 ms).

Cloud-hosted large multimodal models (LMMs) can provide strong open-vocabulary perception for Vehicle-to-Everything systems, but naively transmitting full-resolution frames from edge to cloud causes severe communication overhead and high cloud-side prefill latency. We present CABLE, a cloud-assisted bandwidth-efficient LMM-based encoding framework for edge-cloud perception. CABLE propagates the previous cloud segmentation mask on the edge using ego-motion compensation, refines it with residual-motion cues, and consolidates disconnected regions via a corridor envelope to form a robust region of interest (ROI). Only ROI-masked images are uploaded, while the cloud segmentation output is fed back as the prior for the next frame, forming a mask-to-ROI-to-LMM feedback loop. Experiments on five datasets (nuScenes, WOD-ZB, Waymo, KITTI, and CADC) show consistent communication savings while largely preserving perception, achieving $73$--$87\%$ ROI pixel-coverage reduction with $5$--$8\times$ estimated LMM prefill speedup at a modest detection-quality trade-off relative to full-frame inference.

LLM serving exhibits extreme length variability, making size-based scheduling difficult in practice. Recent LLM schedulers approximate SJF/SRPT using predicted decode lengths or ranks and primarily report mean-centric metrics such as TTFT and TBT. We show that these prediction-driven policies can be fragile under distribution shifts, bursty arrivals, and GPU memory pressure, while offering limited control over the tail latency (P90-P99) that dominates user experience, even with perfect decode-length knowledge. We introduce a distribution-aware, prediction-free scheduling framework that replaces explicit length prediction with soft priority boosting driven by lightweight statistical signals. Our design co-optimizes scheduling and cache-aware preemption to account for memory-coupled decode dynamics across workload mixes. Evaluated on production and open-source traces, our method reduces P99 TTLT by up to 35-50% relative to SRPT with perfect length knowledge and reduces TTFT by 34-47% across workloads, including reasoning-heavy and chat-heavy tasks. These results demonstrate a robust alternative for optimizing tail latency in online LLM serving.

Banks face two threat families with fundamentally different detection requirements: signature-based fraud (card-not-present attacks, account takeover, ATM cloning) and behavioural financial crime (structuring, layering, mule networks, business email compromise). Static rule engines catch high-velocity events but remain blind to BEC payment redirection, session hijacking, and laundering layering, which are engineered to resemble legitimate activity at the individual level. This paper presents an AI security agent for retail and corporate banking using a three-component fusion architecture across two parallel event streams: transactions (card fraud, ACH/wire fraud, AML) and sessions (account takeover, hijacking, SIM-swap, insider abuse). Each stream combines an LSTM sequence model of per-account behaviour, a statistical velocity/threshold monitor, and a graph module capturing account-counterparty patterns (fan-in, fan-out, pass-through ratio) for laundering detection. Experiments on a synthetic log of 237,669 transactions and 113,508 sessions across 13 threat categories and 3,470 accounts show overall F1 of 0.787 (transaction) and 0.867 (session), versus 0.562/0.733 for a rule-based baseline and 0.655/0.713 for an LSTM-only baseline. The agent also includes a customer-facing verification chatbot (96.6% identity accuracy, 86.8% mass-reset detection) and an analyst case-summary assistant (99.3% action recommendation F1), with Critical-tier response latency under 0.43 ms at the 95th percentile.

Large Language Models (LLMs) are increasingly deployed in edge-cloud inference systems to handle diverse user tasks with heterogeneous accuracy, latency, and cost profiles. Selecting the appropriate LLM for each incoming task is critical for ensuring service quality and efficient resource utilization. However, model heterogeneity, stochastic and unknown performance characteristics, and time-varying task demands make static selection strategies inadequate. Real-world deployments often impose hard resource budgets such as monetary expenditure limits, along with soft service-level requirements such as latency guarantees. These constraints introduce additional challenges for online decision-making. We formulate this problem as a constrained stochastic bandit learning task, where the learner sequentially selects models under both packing-type (hard) and covering-type (soft) constraints, while adapting to time-varying task demand. The learner operates without access to the underlying reward, cost, or latency distributions and must rely on partial feedback. We develop a novel online learning algorithm that leverages confidence-bound estimates and demand predictions to balance reward maximization with long-term constraint satisfaction. We provide theoretical guarantees showing sublinear regret and sublinear covering constraint violations compared to an offline benchmark with full information. Experimental results on synthetic workloads demonstrate the effectiveness and robustness of our approach in dynamic, resource-constrained environments.

Streaming data systems increasingly underpin Machine Learning workflows that maintain large numbers of continuously updated aggregations. In production settings, each incoming event typically triggers read-modify-write operations to persistent storage, making high-frequency state updates a dominant source of latency, contention, and operational cost. In this work, we decouple inference from state persistence in streaming Machine Learning pipelines via probabilistic thinning: every event is scored, but durable state updates are selectively triggered by informative events. Unlike approaches that shed input or state, we show that persistence-path control is achievable without a high-frequency in-memory control plane or cross-worker coordination, relying exclusively on approximate statistics retrieved from disk-backed key-value stores. We model the resulting stochastic processes, derive bounds on filtering rates, and prove that common time-based aggregations remain unbiased under variance-aware formulations, preventing systemic error accumulation. We evaluate the approach in a controlled setting that isolates per-event costs, demonstrating substantial reductions in storage Input/Output and serialization overhead. Across experiments, up to 90% of events are excluded from the persistence path while preserving and in some cases improving downstream utility.

Expert parallelism has become the prevailing paradigm to serve Mixture-of-Experts (MoE) models. Its efficiency depends on the communication and computation latencies of the GPUs, which are linked to the placement of experts in the GPUs. Existing works for optimizing expert placement focus on leveraging past requests' expert activation patterns. However, they demonstrate deficiencies facing diverse and rapidly changing request patterns, calling for an online, proactive approach. Implementing such an approach requires addressing several challenges: the uncertainty associated with incoming requests' expert activation, the cost of expert migration, and the NP-hard complexity in optimization. Therefore, we present Director, a new distributed MoE serving system that minimizes end-to-end latency via prediction-driven, online expert placement. Director uses either a lightweight cascaded predictor or a low-bit quantized replica for expert activation patterns of incoming requests. An online migration module then enacts the changes with near-zero downtime by executing migrations in compute-bound phases, keeping disruption bounded. At its core, a relaxation-based expert placement optimizer operates under capacity constraints, runs in polynomial time, and achieves a $(1+ε)$ approximation ratio. Finally, we implement a prototype and demonstrate, through extensive experiments, a reduction in end-to-end latency of $11\sim55\%$ for popular MoE models (e.g., Mistral, DeepSeek and Qwen) compared to existing work.

Disaggregated inference architectures physically separate prefill and decode phases onto distinct GPU pools, creating competing "agents" that share a fixed hardware budget. We provide, to our knowledge, the first formal game-theoretic analysis of this architecture, using NVIDIA Dynamo as a concrete case study. We model disaggregated serving as three coupled games: a two-player resource game between prefill and decode pools, a selfish caching game over the hierarchical KV cache, and a congestion game with positive externalities for request routing. We empirically validate the latter two; the P/D resource game is treated analytically (Section 9.2). We characterize how GPU saturation induces regime transitions that shift the game's payoff structure: below saturation, selfish behavior has bounded Price of Anarchy (PoA); at saturation, superlinear latency and cache externalities drive our empirical estimator PoA-hat (defined in Section 6.4) upward. Based on this analysis, we design an adaptive controller that detects saturation transitions in real time and adjusts routing parameters accordingly, shifting from cache-affinity exploitation to load-balanced congestion avoidance. We instantiate our framework on a 3-node NVIDIA B200 cluster running Dynamo with two models, Nemotron-4-340B (TP=8, full-node workers with cross-InfiniBand KV transfers) and Llama-3.1-70B (TP=4), and find the same three-regime PoA-hat structure with the same first post-knee grid point (C=128) on both models. Adaptive routing shifts each model to a better operating point. Our strongest result is on the 70B 1P/5D topology, where PoA-hat drops 3.1x (66.4 to 21.5) in the saturated phase at a 13% throughput cost. On the 70B 1P/2D, PoA-hat drops 2.2x and TTFT P99 drops 7.6x (see Section 8.5).

While low-latency interaction is critical for spoken dialogue, cascaded architectures are often bottlenecked by reactive turn-completion detection. We propose Endpoint Anticipation, shifting from reactive detection to proactive forecasting of end-of-turn signals. Our speech-based model anticipates endpoints upto 2.56 seconds in advance, enabling speculative execution of LLM and TTS pipelines on partial context. We introduce metrics to quantify the trade-off between realized latency reduction and computational redundancy. Evaluation across conversational and task-oriented datasets shows our model consistently outperforms competitive VAP-based baselines. Integration with the Unmute framework demonstrates a 505 ms average latency reduction with a 28.4% increase in speculative computation, effectively masking sequential bottlenecks to enable complex reasoning in real-time speech-to-speech interaction.

Diffusion large language models (dLLMs) accelerate generation by denoising multiple tokens in parallel, making them attractive for latency-sensitive mobile inference. However, repeated denoising introduces substantial computation on smartphones. Mobile neural processing units (NPUs) offer high-throughput dense matrix computation, but efficiently exploiting them remains challenging: token commitment shrinks per-block effective workloads, token revision complicates KV cache reuse, and limited NPU-visible address space incurs costly remapping and data transfer overheads. In this paper, we propose llada.cpp, the first NPU-aware inference framework for accelerating dLLMs on smartphones. llada.cpp aligns block-wise dLLM inference with the execution characteristics of mobile NPUs through three techniques. (1) Multi-Block Speculative Decoding fills the shrinking workload in late-stage current-block decoding with speculative future-block tokens. (2) Dual-Path Progressive Revision keeps committed tokens revisable until stable and refreshes unstable tokens through a CPU-side path without stalling dense NPU execution. (3) Swap-Optimized Memory Runtime compacts NPU-visible address layouts and overlaps data staging with NPU computation to reduce remapping and transfer overheads. We implement llada.cpp as an end-to-end framework and evaluate it across diverse hardware platforms and dLLM workloads. llada.cpp reduces LLaDA-8B generation latency by 17x-42x over the CPU baseline with prefix KV cache reuse, while preserving generation quality.

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.

Simultaneous speech-to-speech translation aims to enable near-real-time communication by minimizing latency, offering a compelling, real-time alternative to the high latency of consecutive translation. However, the excessive pursuit of low latency often results in fragmented chunk-wise speech. Consequently, listeners are subjected to an unnatural acoustic flow punctuated by frequent pauses, which could increase their cognitive load. To bridge this gap, we introduce a fluency-aware optimization framework designed to discover the sweet spot between the low-latency benefits of simultaneous translation and the natural flow of consecutive translation. Our framework minimizes inter-chunk silences by leveraging model-internal signals, including linguistic diversity and induced temporal variability in speech durations. Experiments on short- and long-form benchmarks show that our framework produces natural speech flow while maintaining competitive latency and translation quality.

Distributed machine learning (ML) is a key paradigm for today's large-scale artificial intelligence applications. As model inference arises as an important use case, faithful modeling of latency-sensitive collective communication has never been more important. Capturing the device architecture and modeling control and data paths at high fidelity is therefore a necessity today. Having a common, detailed representation for distributed ML infrastructure is also crucial. We revisit the promising open-source, community-driven simulator: ASTRA-sim. In this work, we identify limitations of the current ASTRA-sim simulator and augment it with new features. To this end, we enable fine-grained, high-fidelity simulation with a standardized infrastructure representation, opening new design space exploration opportunities. We propose the simulation at cache-line-sized load-store granularity, with a detailed graphics processing unit (GPU) execution model, to balance simulation scalability and fidelity. We also introduce InfraGraph, a standardized representation to capture distributed ML network infrastructure in detail. Using the updated ASTRA-sim 3.0 simulator, we showcase interesting design space explorations for designing optimized collective algorithms, network requirements, and GPU architectures.

Ask a pretrained biomedical language model whether "cortisol 28 ug/dL" and "stock-market volatility" are related, and it returns a cosine similarity of 0.83 on a scale where 1.0 means identical. The two share no mechanism. This is not a corner case: every off-the-shelf biomedical encoder we tested (BioBERT, PubMedBERT, BioM-ELECTRA) scores unrelated cross-domain pairs between 0.76 and 0.92 when the answer should be near zero. Accuracy on cross-domain discrimination is 0%. Retrieval systems survive this, because a language model downstream filters the noise. A Large Behavioural Model (LBM), a foundation model whose subject is a person rather than a sentence, does not: it reasons over a graph of a user's life and treats embedding proximity as evidence that two events are causally linked. False proximity writes a false causal edge, and everything downstream inherits the error. Here, embedding geometry is not a tuning knob; it is correctness. We report the fix. A contrastive pass over 72,034 pairs raises PubMedBERT BIOSSES correlation from 0.633 to 0.828 and within-vs-across-domain separation from 1.05x to 1.63x. A second pass, BODHI, mines hard negatives from edges absent in a biomedical knowledge graph and lifts separation to 2.30x and the discrimination gap to +0.392, at a 4.5% BIOSSES cost. On an Intel Xeon 6737P with AMX, OpenVINO cuts single-query latency from 1367 ms to 10 ms (133x) and reaches 555 sentences/sec. One finding contradicts standard advice: FP16 beats INT8 on this silicon at every serving batch size, and we explain why. The same model on a no-AMX Ice Lake instance runs 13-27x slower. We release the benchmark suite, training corpora, the BODHI generator, and the OpenVINO scripts.

Tensor program optimization is essential for modern machine learning systems, but its search space is enormous. Existing auto-schedulers reduce measurement cost with learned cost models, yet they usually evaluate each candidate as a static code snapshot, ignoring the schedule trajectory that produced it. This makes them insensitive to action dependencies and vulnerable to superficial code variations. We propose a \emph{world-model-inspired} evaluator that models schedule evaluation as action-conditioned latent dynamics over program states. Starting from the initial program, it rolls out scheduling actions in a continuous latent space with a lightweight transition model, avoiding expensive AST mutation and repeated code encoding. The final dynamic representation is combined with action and hardware features to rank candidates. Implemented in TVM AutoScheduler, our method improves representative-subgraph latency over Ansor by 1.37$\times$ on GPU and 1.54$\times$ on CPU under the same 64-trial budget. It also matches Ansor-10K within 2.2% geometric mean using 10$\times$ fewer measurements, and accelerates full-model inference over PyTorch/PyTorch-opt(cuDNN) by 4.61$\times$/3.67$\times$ geometric mean.

The Robot Operating System 2 (ROS 2) is a widely used middleware for robotic systems, characterized by a publish-subscribe (pub-sub) communication mechanism in which computation is structured as callbacks dispatched by ROS 2 executors. Despite its popularity, the pub-sub pattern in ROS 2 is inherently nondeterministic: the order in which these callbacks run is nondeterministic even within a single executor, and distributed deployments add further nondeterminism from the interleaving of messages across nodes and from network latency. Such nondeterminism often leads to concurrency issues and makes it virtually impossible to analyze for safeness and provide guarantees. We present a framework that is able to convert an unmodified ROS 2 application and run it under Lingua Franca (LF), a coordination language for deterministic execution using logical time, so that the same input always produces the same deterministic execution order. We first describe which ROS 2 features can be executed deterministically under logical time. Such features enable the possibility to establish an automatic conversion framework to extract information from a ROS 2 application and directly convert it into an LF program. The rich features of LF, such as logical-time delays, federated execution across processes, and fault handling, can then be applied to make the ROS 2 application be executed in a deterministic and timing-predictable manner without changing the ROS 2 code. We evaluate the framework on a synthetic example and on the Autoware reference system. We show that the order in which callbacks are executed differs in default ROS 2, while also having end-to-end latencies that vary across executions. In contrast, our LF-controlled ROS 2 system produces a deterministic execution order and consistent end-to-end latencies.

This paper reads six engine-level measurements together -- 1.1 host attack surface, 1.2 information leakage, 1.3 defense-in-depth stackability, 1.4 public CVE history, 1.5 patch cadence, and 1.6 upstream fuzzing posture -- to describe how five AI-sandbox products isolate guest code from the host kernel. No single axis is a sufficient basis for a comparative judgement; the cross-axis reading is the load-bearing analysis. Three high-level findings: (1) engine classes (microVM, userspace kernel, OCI container) separate cleanly on every architectural axis, but products within a class do not; (2) product pin policy is the dominant operator-facing variable -- engine-side patch latency aggregates to ~0 days for coordinated disclosures, while downstream lag spans 0 days to 471+ days to "opaque" to infinity; (3) fuzzing investment splits into three tiers, and the strongest combination -- microVM x continuous public fuzzer -- is unoccupied in this set, leaving the "0 published CVEs x no upstream fuzzer x no academic study" intersection structurally unmeasured. We report per-axis orderings, per-product portraits, and a threat-model qualification matrix; no overall ranking is proposed. Companion repository (code, Apache-2.0): https://github.com/orbitalab/RnD-ai-sandboxes-sec-study-part-1. License: CC BY 4.0.