Real-Time Inference

While Mamba-based models have shown strong potential for long sequence modeling, adapting them to vision is challenging due to the requirement of local neighborhood correlations and multi-directional spatial contexts for visual understanding. In this paper, we present LiAuto-MindViT, a novel hybrid vision backbone that synergizes the strengths of CNNs, Mamba, and Transformers. The core of our design is the Adaptive Bidirectional Mamba (ABM), which eliminates the directional bias of unidirectional SSMs through bidirectional selective scanning with learnable alpha blending, enabling content-adaptive directional fusion without the overhead of exhaustive multi-path routing. To further accelerate inference, we propose a deployment-friendly Reparameterized ConvSE (RepConvSE) module that leverages structural reparameterization to reduce latency and memory access overhead. Extensive experiments demonstrate that LiAuto-MindViT achieves state-of-the-art performance on image classification, object detection, and semantic segmentation while enabling efficient inference through reparameterization.

Diffusion-based Vision-Language-Action (VLA) models achieve strong performance in embodied tasks, but their iterative sampling imposes heavy computational and memory-access cost, blocking real-time deployment on edge platforms. Existing acceleration methods either require expensive training (e.g., distillation, flow matching) or degrade perception via statically scheduled pruning and caching, ignoring the dynamic workload variance of robotic interactions. This paper presents FASA (Feedback-Aware Sampling Adaptation), a training-free runtime framework that treats real-time multimodal feedback as a control signal for the denoising pipeline: an interaction-driven range adaptor modulates the global sampling-step budget based on visual and gripper-force feedback, and a proprioception-aware step adaptor pinpoints the optimized step within the adapted range. This co-designed framework allows the underlying hardware architecture to adaptively match the workload demands of different execution phases. Comparative evaluations across several benchmarks show that the inference speed can be increased by up to 1.45$\times$ while maintaining competitive success rates, providing a novel dynamic runtime architecture paradigm for deploying heavy generative embodied AI workloads onto resource-constrained computing platforms.

Factory work is a promising early scenario for embodied AI: assigning repetitive manual jobs to robots has clear economic payoff, and a structured station keeps the jobs tractable for current policies. Vision-Language-Action (VLA) models now dominate as the policy paradigm for these robots. The inference latency of VLA models directly affects robot responsiveness and motion smoothness. However, existing VLA inference frameworks do not fully exploit the characteristics of embodied workloads or account for the distinct bottlenecks across different stages of VLA inference. In this paper, we first characterize embodied workloads and identify substantial task similarity across repeated robot executions. We further find that such similarity extends beyond observations and action trajectories to internal model states. Drawing on these observations, we present rMuscle, a real-time VLA inference framework inspired by human muscle memory. It exploits cross-execution similarity through a dual-phase muscle-memory cache. The Context Cache reuses visual-token outputs to reduce computation, while the Action Cache reuses neuron activation patterns to reduce weight accesses. We keep both the cache memory footprint and access overhead low through online cache recomputation, sliding-window cache retrieval, and mask sharing across consecutive denoising steps. rMuscle achieves 1.29-1.42X speedup on RTX 4090 and Jetson Thor across LIBERO, RoboTwin, and physical manipulation tasks, while maintaining the original success rates on real-world robots.

Billion-parameter vision--language--action (VLA) policies demand substantial onboard power, while communication delays in remote inference hinder timely responses. We propose VLA-ULAP, which interleaves remote VLA calls with an Ultra-Lightweight Local Action Predictor (ULAP). With approximately 7.4M parameters including the frozen vision encoder, ULAP combines current views, proprioception, and executed action history to predict chunks in one pass. Trained independently, it requires no VLA hidden states, online verification, or server round trips. On Jetson Orin Nano, ULAP takes 19.9 ms and 0.183 J per inference, compared with 284.3 ms and 50.55 J for GR00T on RTX A6000. Across three simulated base-policy/benchmark pairs, selected operating points remove 48.8--76.7% of VLA calls while retaining 95.0--97.5% of the baseline success rate. Against local VLA-acceleration alternatives on VLA-JEPA, ULAP uses an estimated 49.2% less inference time and 51.0% less GPU energy per successful episode than ACT at comparable success rates, and 77.1% less time and 79.9% less energy than SP-VLA at equal success rates. Physical SO-101 experiments retain 95.2--100% of the baseline success rate across seen and held-out placements while reducing inference time by an estimated 47.9--58.0% and inference-device energy by 52.1--62.5%, based on successful-episode call counts and measured device costs. Faster responses also improve dynamic-task success rates: in latency-aware LIBERO-Safety simulation, VLA-ULAP exceeds $π_{0.5}$ by 11.0 and 15.5 percentage points on two tasks while approximately halving VLA calls.

FastWAM-style world action models enable efficient action-only inference, but generalize poorly under visual distribution shifts. Their reconstruction-oriented representations emphasize appearance-specific details, limiting generalization to unseen scenes and objects. Without observation history, the model also lacks temporal evidence for robustly identifying task-relevant state changes and motion in unfamiliar visual conditions. To address these limitations, we present the Causal Semantic World Action Model (CSWAM), which augments FastWAM with a causal semantic expert built on V-JEPA 2.1. V-JEPA provides temporally grounded representations of semantic state changes and motion with less dependence on appearance-specific details. The expert learns their future evolution from a sparse history of current and past observations and shares the history-derived context with both the video and action streams through causal attention. At inference, CSWAM conditions action denoising on the current video state and observed semantic history, retaining efficient action-only inference. We conduct simulation and real-robot experiments to evaluate generalization under distribution shifts. With embodied pretraining, CSWAM raises Randomized success on RoboTwin 2.0 Clean-to-Randomized transfer from 10.16% to 45.18%, a gain of 35.02 percentage points over FastWAM. Across two real-robot tasks and three OOD difficulty levels, CSWAM improves average success over FastWAM by 42.5 percentage points, from 27.5% to 70.0%.

We introduce TabPFN-3.5, our new flagship Tabular Foundation Model. It significantly outperforms its predecessor, TabPFN-3, and all existing baselines across a broad range of tabular problems. TabPFN-3.5 sets a new state of the art on standard tabular prediction in TabArena, and extends it to the data practitioners encounter in practice: non-i.i.d. data with temporal or grouped splits, tables with strings, text and images, high-cardinality categorical features, and wide tables with many features. These gains carry over to our task-specific harnesses: state of the art on relational data and stronger time-series forecasting. For faster inference, our variant TabPFN-3.5-Fast runs up to 3x faster than TabPFN-3 while keeping most of the accuracy gains. In addition, we upgrade TabPFN-3.5-Plus, expanding our multimodal capabilities with advanced text and date handling alongside proprietary inference optimizations. Finally, we release a new version of our Thinking mode, TabPFN-3.5-Thinking, which scales inference-time computation to push the state of the art further. It benefits from our stronger base model and from inference-time improvements that make it up to 12x faster than TabPFN-3-Thinking.

As unmanned aerial vehicles (UAVs) become increasingly prevalent in consumer and defense settings, classifying them reliably from limited, modality-specific data is an urgent challenge. The dominant approach, large pretrained networks fully fine-tuned on task data, carries a substantial computational and memory weight that is hard to bear in resource-constrained UAV deployments, where edge inference and rapid retraining for emerging platforms are both required. This paper systematically scales across both model architectures and fine-tuning methods for UAV audio classification, asking when that weight is justified and when lighter alternatives prevail. Using a custom dataset of 3,100 audio clips spanning 31 drone classes, we evaluate transformer (ViT, AST) and convolutional (custom CNN, ResNet-18/152, MobileNet-V3-S/L, EfficientNet-B0/B7) backbones under full fine-tuning, classifier-only fine-tuning, and four parameter-efficient fine-tuning (PEFT) methods: SSF, IA3, OFT, and selective batch-norm tuning. All configurations are evaluated with 5-fold cross-validation across accuracy, training time, trainable-parameter share, and inference-time memory footprint. Selective batch-norm fine-tuning of EfficientNet-B7 with three-fold augmentations achieves the highest validation accuracy (97.65% +- 0.30) while updating under 0.5% of model parameters. Across the sweep, lightweight CNNs consistently outperform transformers on both accuracy and efficiency. For UAV audio classification under data scarcity, scaling the method outperforms scaling the model.

Diffusion language models (dLLMs) promise fast inference by generating multiple tokens in parallel, but suffer severe performance degradation when parallel decoding is pushed too aggressively. We introduce Temporal Self-Distillation (TSD), a simple on-policy method that trains dLLMs for fast inference by distilling predictions across time. Specifically, TSD distills the model's denoising distribution at earlier timesteps toward its distribution at the final timestep at which a token is committed. This encourages earlier predictions to better anticipate the model's eventual output, enabling much more aggressive parallel decoding. Because its teacher signal comes from the model itself, TSD requires no offline teacher generation and applies seamlessly to both base and post-trained policies. Across seven benchmarks in mathematics, planning, and code, TSD substantially shifts the speed--quality frontier toward the low-compute regime. TSD thus provides a simple, single-stage approach to accelerating dLLMs, achieving speedups competitive with offline distillation while avoiding a complex two-stage pipeline.

EdgeAI systems are increasingly employing computer vision applications to enable intelligent, on-device decision-making in real-time. However, these deployments face highly dynamic operational conditions, with fluctuating constraints on latency, power availability, and memory resources. Deep Neural Networks (DNN), which follow fixed computational execution flows, lack the flexibility to adapt to such variability, resulting in inefficient and suboptimal performance in edge scenarios. This underscores the need for architectures that are not only efficient but also dynamically scalable at runtime. In this paper, we propose Elastoformer: A framework that transforms conventional neural networks (NN) into Elastic NN capable of real-time elastic inference. Unlike the conventional bag-of-models approach, which requires maintaining multiple independent models for different operating conditions, Elastoformer offers a single, modular solution that dynamically switches between multiple modes of operation at runtime, adapting efficiently to the changing computational budgets of edge devices without the overhead of managing separate models. Experiments reveal that our framework achieves up to 85% reduction in computation FLOPs, 50% reduction in latency and 76% reduction in memory overhead, while showcasing the architecture agnostic nature of the framework across both Vision Transformers and CNNs. Our code is available at https://github.com/sudaksh14/Elastoformer.

Mixture-of-Experts (MoE) models enable efficient scaling of large language model (LLM) inference but suffer from substantial data-movement overhead when deployed on neural processing unit (NPU)-based systems. Near-Data Processing (NDP) provides a promising way to mitigate this bottleneck via cooperative NPU-NDP execution. However, existing NPU-NDP MoE systems do not fully account for hardware heterogeneity, dynamic expert-level concurrency, and temporal expert reuse during batched inference. This paper presents DynaNDE, a dynamic near-data expert scheduling framework that exploits NPU-NDP collaboration to accelerate batched MoE inference. DynaNDE introduces an analytical performance model that captures hardware heterogeneity, data-movement costs, and communication-computation overlap in cooperative NPU-NDP execution. Guided by this model, DynaNDE determines per-layer expert scheduling across the NPU and NDP while accounting for expert-level concurrency. DynaNDE also incorporates a reuse-aware runtime that avoids redundant parameter movement when experts reside in NPU memory. Experimental results show that DynaNDE achieves substantial throughput improvements over the state-of-the-art NPU-NDP MoE serving framework, with average speedups of 2.6$\times$ and 2.2$\times$ for the prefill and decoding stages, respectively.

Autoregressive large models have recently advanced Text-to-SVG generation from simple icons to complex, long-context graphics, yet standard autoregressive decoding often fails to maintain global consistency across geometry, layout, occlusion, and composition. We introduce CANVAS (Consistency-Aware Navigation via Visual Adaptive Sampling), a training-free, render-aware inference framework that combines power-sharpened trajectory likelihood with visual feedback from rendered futures and derives a stroke-wise navigation rule. It effectively estimates each candidate stroke's future value under a limited generation and rendering budget and adaptively allocates samples according to candidate uncertainty, decision influence, and rollout cost. Experiments across multiple autoregressive SVG backbones and complementary benchmarks demonstrate improvements in global consistency, which includes sound geometric relationships, spatial layouts, occlusion ordering, and overall composition, without additional training, demonstrating the effectiveness and generalization ability of our framework.

Text-rich image understanding requires multimodal large language models (MLLMs) to organize OCR (Optical Character Recognition)-grounded evidence across words, layout, fields, charts, and visual correspondences. Existing evaluations often conflate extraction with reasoning and rarely test whether models follow the required reasoning direction: applying visible rules, abstracting hidden regularities, or recovering missing premises. We introduce OCR-MetaReasoning, a controlled single-image benchmark that treats deduction, induction, and abduction as distinct directions and separates final-answer correctness from reasoning-process compliance. The benchmark contains 1,500 verified samples in a balanced $3\times5$ taxonomy crossing three reasoning types with five OCR-object categories, along with reference reasoning steps, automatic answer scoring, the Meta-Reasoning Macro Score (MRMS), and the Reasoning Process Compliance Score (RPCS). Experiments with representative closed-source and open-source MLLMs show that OCR-grounded meta-reasoning remains far from saturated: models struggle with visible-rule application and layout-sensitive inference, while process-compliant rationales can accompany incorrect final answers under exact-match evaluation. The code is available at https://github.com/gengxuli/OCR-MetaReasoning.

Deploying robot policies in the physical world requires satisfying two fundamental desiderata: reliability and smooth real-time execution. However, deploying state-of-the-art generalist models presents challenges on both fronts. Achieving the precision and robustness required for real-world deployment necessitates sample-efficient online reinforcement learning (RL) to adapt pretrained models. Meanwhile, the increasing scale of robot foundation models has led to higher inference latency. To satisfy real-time constraints under high latency, modern systems adopt asynchronous inference with action chunking, overlapping policy computation with chunk execution to hide latency and enable smooth control. Despite their complementary roles, integrating asynchronous execution with gradient-based online RL remains underexplored. We present SmoothRL, an online RL framework that fine-tunes a pretrained policy within an asynchronous inference loop. SmoothRL follows a value-gradient paradigm, directly updating policy parameters using gradients of the action-value function with respect to policy actions. To enable correct optimization under asynchronous execution, SmoothRL explicitly models the asynchronous inference process during training. Specifically, each generated action chunk is partitioned by frame index into three regions: a committed region, consisting of actions committed by the previous inference cycle; an execution region, containing newly generated actions executed by the robot; and a discarded region, containing actions superseded by the next inference cycle. Gradients are propagated only through the execution region, ensuring policy optimization aligns with the trajectory distribution induced by asynchronous execution. We evaluate SmoothRL on real-world robotic tasks requiring high precision, as well as highly dynamic tasks that necessitate asynchronous execution.

As edge-side vision services continue to expand toward low-latency, high-throughput scenarios, reducing the inference cost of vision models without sacrificing reliability has become a central concern. Existing semantic caching methods largely rely on empirical similarity thresholds; while such thresholds improve hit rates, they tend to introduce silent misclassifications near decision boundaries. To address this, we propose \texttt{LipCache}, a certified semantic caching framework for image classification. Without modifying the existing deployed main model, \texttt{MainNet}, the framework introduces a lightweight network, \texttt{GuardNet}, that maps inputs into a low-dimensional feature space subject to a Lipschitz constraint. It then computes a per-sample certified reuse radius from the local classification margin and the spectral norm of the classification head. At runtime, a cached result is reused only when the query feature falls inside the certified reuse ball; otherwise, the query falls back to \texttt{MainNet}. Thus, cache hits are transformed from empirical threshold tests into geometric certification decisions with explicit theoretical boundaries. Across standard image classification tasks like CIFAR, Tiny-ImageNet, and SVHN, \texttt{LipCache} achieves a measured speedup of up to $1.65\times$ with limited end-to-end accuracy degradation, while all accepted cache hits satisfy the \texttt{GuardNet}-side certified-consistency condition. Furthermore, an enhanced \texttt{GuardNet} training recipe substantially improves cache hit rates in the Tiny-ImageNet multi-class extension while maintaining a certified-consistency rate of $100\%$. These results demonstrate that per-sample certified reuse can reduce main-model fallback while preserving theoretical consistency, providing a feasible approach to reliable cache-assisted inference at the edge.

Vision-Language-Action (VLA) models promise to bring end-to-end reasoning to autonomous driving, but their computational cost remains far too high for real-time control. The core challenge is structural: VLA inference is not a single bottleneck but a cascade of four. Visual encoding wastes compute on overlapping video frames; language-model prefill recomputes context that could be carried over from the previous timestep; reasoning tokens are generated serially despite low entropy; and flow-matching denoising applies uniform compute to a non-uniform velocity field. Addressing any one stage in isolation leaves the others untouched. We propose FlashDrive, an algorithm-system co-design framework that targets all four stages simultaneously. Our key insight is that each bottleneck admits a distinct, lightweight algorithmic shortcut: temporal overlap enables streaming KV-cache reuse across frames; the low per-token entropy and strong intra-block correlations of driving-domain reasoning make a non-autoregressive diffusion drafter highly effective for speculative decoding; and the velocity field's structure---sharp at the endpoints, flat in the middle---permits adaptive step caching that concentrates compute where it matters. Layered on system-level CUDA Graph compilation and kernel fusion, these techniques compound. Applied to Alpamayo 1.5-10B with W4A8 quantization, FlashDrive reduces end-to-end latency from 717ms to 151ms (4.7x) while leaving accuracy essentially unchanged: minADE6@6.4s shifts by only 0.08m, minADE1 improves, and closed-loop collision and off-road rates improve in simulation. By raising a 10B-parameter reasoning VLA from 1.4Hz to 6.6Hz on a single GPU, FlashDrive moves end-to-end autonomous driving substantially closer to real-time deployment.

Mixture-of-Experts (MoE) models are attractive for edge deployment because they provide high model capacity while activating only a small subset of parameters per token, improving compute efficiency. However, MoE inference at the edge is fundamentally limited by memory. Expert parameters are large and often reside in off-chip memory due to capacity, cost, and power constraints, putting expert loading to the critical path. We present APEX: Adaptive Expert Prefetching, a predictive resource management framework that overlaps expert loading with useful computation. APEX introduces a lightweight prefetch router that predicts candidate experts before the attention block to dynamically fetch additional experts using a learned confidence model. This adaptive strategy achieves over 99% overlap accuracy, significantly outperforming fixed top-k prefetching techniques. APEX supports two execution modes: a correctness-preserving mode that guarantees exact routing semantics, and a stall-free mode that eliminates residual stalls by operating on available experts with negligible impact on application accuracy. Across multiple MoE models, the correctness-preserving mode reduces per-token latency by up to 26% and improves energy-delay product (EDP) by up to 41% over state-of-the-art baselines, while the stall-free mode provides additional efficiency gains with negligible impact on application accuracy. These results establish adaptive, confidence-driven expert prefetching as an effective approach for efficient MoE inference on edge systems.

This work investigates whether Electroencephalograph (EEG) foundation models (EFMs) can be made faster and locally deployable without sacrificing accuracy. EEG foundation models are a major trend, offering strong general-purpose representations. However, their computational burden grows quadratically with input length, hindering deployment on resource-constrained scenario, particularly for real-time clinical monitoring. EEG's low SNR further suggests many of these tokens are redundant and compressible with little accuracy cost. We propose ZIPBrain, a novel redundancy-aware EEG token pooling module that leverages this low-SNR characteristic to reduce token count. Given a token sequence, ZIPBrain partitions tokens into redundant and unique groups, then merges each redundant token with its most similar counterpart in the unique group. Furthermore, ZIPBrain serves as a training-free, plug-and-play module that seamlessly integrates into standard Transformer encoders with negligible computational overhead. Extensive experiments across multiple EEG foundation models show ZIPBrain's strong versatility, achieving 1.3%-10.5% average improvement over baselines, while reducing wall-clock inference time by 32.7% (up to 41.8% with CUDA Graph) compared to the original EEG foundation models.

Mobile image denoising requires both good restoration quality and low computational cost. In addition, it's annoying to collect large-scale LQ-GT clean pairs. As a result, we propose LiteKD-Net, a lightweight knowledge-distilled network for mobile image denoising. First, a physics-guided noise simulation pipeline generates paired training data by adding pixel crosstalk compared with pipelines applied to cameras. Next, we adapt the Real-ESRGAN to identity-resolution denoising and construct a lightweight Student using Lite-RRDB blocks based on depthwise separable convolutions. Third, feature-level knowledge distillation is applied to transfer the Teacher's restoration capability to the Student without introducing additional inference cost. Experiments on real-world datasets show that our model reaches great reduction in runtime and increase in the inference rate with good restoration quality. Our model also reaches the best in all metrics compared with SwinIR. These results indicate that LiteKD-Net provides a great trade-off between restoration quality and computational efficiency.

Can computer vision help make classrooms safer? In this pilot study, we investigate privacy-aware and computationally efficient classroom incident recognition from CCTV-style observations. This setting remains underexplored, with limited benchmarks and few methods designed for the privacy, efficiency, and generalization demands of real-world deployment. We introduce a novel hybrid benchmark combining generative CCTV-style videos with real-world classroom pose data, and propose a lightweight, but robust motion-reasoning framework motivated by the observation that many incidents differ more in motion direction, speed, acceleration, and intensity than in pose alone. To that end, our method first constructs hierarchical kinematic representations of human actions. Our method then distills hierarchical, multi-order kinematic reasoning from a large teacher into a much smaller single-order student, enabling efficient per-person inference while preserving expressive motion understanding. Experiments show that our model outperforms substantially larger baselines at less than one-tenth of their computational cost, while also demonstrating stronger out-of-domain motion reasoning and zero-shot synthetic-to-real generalization. We will publicly release the benchmark, codebase, and supporting tools to facilitate further research in privacy-aware classroom safety.

Vision-language-action (VLA) models have emerged as a key component in embodied AI. Among existing approaches, diffusion-based VLA models achieve superior motion quality and generalization. However, diffusion-based VLA models are compute-intensive and must run at high control frequency, e.g., 50-200 Hz. Thus, it imposes strict latency and energy constraints on edge devices. In this work, we present Deltoris, an algorithm-hardware co-design framework for efficient diffusion-based VLA inference. First, we exploit the temporal similarity of consecutive inputs and propose a \textit{temporal-aware bit-sparsity} algorithm that computes only the differences between consecutive inputs, eliminating redundant bit-level operations. To further address the extra off-chip traffic introduced by our algorithm, we propose a \textit{speculative inference} technique, which amortizes data loading across multiple control steps. Lastly, to support these techniques, we co-design a dedicated accelerator with customized 1D systolic bit-serial PE arrays that eliminate PE workload imbalance. Our evaluation shows that Deltoris achieves up to 34.2$\times$ speedup over mobile GPUs and 6.1$\times$ over prior accelerators, while maintaining comparable accuracy.

Unified multimodal object tracking has achieved remarkable robustness by leveraging complementary sensor data (e.g., RGB, Thermal, Depth), yet the heavy computational burden of state-of-the-art models hinders their deployment on resource-constrained edge devices. In this work, we identify the prediction head as a critical but often overlooked efficiency bottleneck. By strategically streamlining the decoder architecture, we unlock the potential for real-time inference but simultaneously introduce a capacity gap between the lightweight student and the heavy teacher. To resolve this, we conduct a systematic analysis of 17 distillation strategies and introduce a Dual-Alignment Distillation framework. Our key insight is that effective compression requires decoupling knowledge transfer into two complementary streams: (1) Spatial Representation Alignment, which employs feature distillation to sharpen the student's spatial focus on foreground targets ("Where to track"); and (2) Semantic Distribution Alignment, which utilizes logit-based distillation to align decision boundaries and transfer discriminative dark knowledge ("What to track"). Extensive experiments across five benchmarks demonstrate that our approach significantly outperforms complex state-of-the-art methods. Notably, our distilled model achieves 91.5% MPR on RGBT234 and operates at 54 FPS on a single RTX 4090, representing a 5x speedup over the teacher model while maintaining superior accuracy.

The emergence of transformer-based deep learning models has brought unprecedented performance across various domains, particularly in natural language processing and computer vision. However, deploying these models, especially on resource-constrained devices, poses significant challenges due to their high computational complexity and large memory size and bandwidth requirements. This complexity has led researchers to use low-bit model weights to reduce memory usage and improve efficiency. In addition to reducing processing and memory demands, quantization introduces another useful property: value locality, where the extremely large number of parameters are restricted to a limited range of values. To fully take advantage of this locality, this paper presents DeVIT, an acceleration method for vision transformers that leverages differential computation to enable multiplier-less matrix multiplication.

Mapping neural networks to FPGAs enables low-latency, energy-efficient inference, particularly for lookup table (LUT)-based models that eliminate multipliers and map directly to reconfigurable fabric. While prior work achieves high compute efficiency, it typically assumes full-sample availability, causing pipeline stalls in bandwidth-limited streaming scenarios. Here, the bottleneck shifts from computation to data movement, as large input transfers limit throughput and energy efficiency. We present CascadeLUT, an information-structured inference framework organized around bandwidth constraints. Instead of buffering the full input, features are partitioned into ordered subsets and predictions are progressively refined as subsets arrive. The cascade statically controls which layers consume incoming features, enabling deterministic streaming inference without runtime branching. By co-designing feature scheduling with hardware dataflow, CascadeLUT reduces data movement while maintaining accuracy. Across datasets, it achieves 4.0 to 12.5 times lower latency, 3.0 to 5.0 times higher throughput and up to 13.8 times lower energy/sample than prior LUT baselines, using 1.2 to 4.4 times the LUTs of the smallest DWN baseline per task. We also demonstrate on-device input quantization integrated with LUT-based inference and present end-to-end FPGA results on real-world workloads, with 5 times reductions in quantization overhead.

Spiking Neural Networks (SNNs) communicate through sparse binary spike events rather than dense activations, enabling energy-efficient inference on neuromorphic hardware and motivating their use in always-on, battery-powered edge systems. We show that this same efficiency advantage creates a distinct security risk: sponge attacks can increase inference-time spike activity and synaptic workload, inflating energy consumption while remaining difficult to detect through correctness-based monitoring alone. Prior input-space efficiency attacks on SNNs have focused on per-sample optimization, primarily in rate-coded settings. We extend this threat to native event-based binary inputs and study two attack models. First, we develop a per-sample sponge attack that crafts a custom adversarial spike train for each input via gradient-based optimization. This attack increases per-inference SynOps by 1.5-2.6x on three SNN models for the NMNIST, SHD, and IBM DVS Gesture datasets, while preserving the predicted class on at least 98% of evaluated samples. Second, to the best of our knowledge, we introduce the first universal sponge attack for native event-based SNN inputs: a fixed binary perturbation computed offline and applied via XOR to all subsequent inputs. Although weaker, it still inflates SynOps by 1.09-1.24x across all three datasets and represents a more realistic deployment threat because it requires no per-input optimization. Mapping SynOp inflation to estimated Loihi-1 energy yields per-inference overheads from 14 $μ$J to 13.24 mJ. These results show that native event-based SNNs are vulnerable to practical input-space efficiency attacks, and that reusable universal perturbations can accumulate into meaningful battery drain in continuously deployed edge systems.

Audio-video generative models achieve impressive quality but suffer from high latency, making them unsuitable for real-time applications. Although several streaming audio-video generation methods have been proposed, they remain costly and fail to support long-form generation. To address this, we propose \textbf{Ripple}, a real-time joint audio-video generation system with a cross-modal recurrent memory mechanism. To enable efficient streaming inference while preserving long-term context, Ripple combines a fixed-length sliding-window attention with modality-specific memory states that continuously summarize audio and video context. Cross-modal memory interaction is further introduced to enhance audio-visual synchronization. To learn this memory-augmented model effectively, we devise a three-stage training recipe: (1) adapting a bidirectional audio-video teacher to block-wise causal attention with simulated memory, (2) optimizing the memory construction and interaction pipeline through end-to-end distillation, and (3) applying online reinforcement post-training tailored for streaming audio-video generation. As a result, Ripple achieves ~28 FPS at 480P resolution, over faster than the teacher, while capable of coherent long-form generation. Extensive experiments on both short-video and long-video benchmarks demonstrate our superior performance over existing offline and online joint audio-video generation methods.

Speculative decoding accelerates large language model inference without changing the target distribution, but no single drafting structure performs best across real-world workloads. Autoregressive multi-token prediction (MTP) is a lightweight, stable proposal mechanism, whereas block-parallel diffusion amortizes drafting latency over much longer candidate sequences; the better choice depends strongly on the output distribution. We present AngelSpec, a unified training framework for MTP and block-parallel speculative decoding that addresses this heterogeneity at three levels. At the training level, rather than fitting one universal drafter to a uniform data mixture, we co-specialize structure and data: the MTP drafter is trained on diverse conversational data for high-entropy open-ended chat, and the block-diffusion drafter on code and mathematics data for longer predictable continuations. At the architecture level, we propose DFly, a block-diffusion framework combining a hybrid target-conditioning backbone with a predecessor-conditioned autoregressive head, improving target-feature utilization and intra-block dependency modeling while keeping generation parallel. At the inference level, both acceptance length and verification cost vary with domain, request, online load, and hardware, so DFly treats verification as a shared batch-level resource: it reallocates compute toward high-confidence prefixes across requests and combines expected utility with a profiled cost model to adapt verification depth online. Across the Hy3 series, DFly raises the average accepted length on Hy3-A21B by roughly 30% and attains the highest average throughput at every tested concurrency from 4 to 64, a 1.98-2.40x speedup over autoregressive decoding and 10.5-11.8% higher throughput than DFlash. We release AngelSpec to support training and extending these methods.

This is an implementation and measurement study of what it costs to run a streaming speech enhancer on a CPU. We port FastEnhancer-Medium at 48 kHz to faster-enhancer.c, a C runtime with six int8 GEMM tiers selected at initialization, leaving architecture and weights untouched. One Apple M2 core reaches 0.069 real-time factor, against 0.230 for the fp32 ONNX Runtime graph on the same machine, a 3.3x speedup. A Galaxy S23+ (Snapdragon 8 Gen 2) reaches 0.096. The speedup comes from specializing every layer of the runtime around one fixed model. Activation ranges are recomputed per frame, so no calibration set is needed; the k=3 convolutions use Winograd F(2,3); cross-stage state is fp16; the GRU and the dequantization epilogues are fused; and nothing is allocated after startup. Over 824 VoiceBank-DEMAND utterances the engine tracks fp32 to within -0.006 PESQ and -0.08 dB SNR. Speed alone does not settle deployment cost. The enhancer holds a fraction of a core for as long as the microphone is open, so its real-time factor is a duty cycle. A benchmark races through a file; an audio callback does not. Pacing to the 6.67 ms deadline costs 4.2x per frame, saves 49% of the energy, and leaves the cheapest core placement missing 96% of its deadlines. All SIMD tiers within an architecture family emit byte-identical output. The runtime is released as a dependency-free library.

Stereo depth estimation for driving, robotics and augmented reality must run at high resolution under tight latency budgets, yet in transformer-based matchers the global self-attention that aggregates scene context grows quadratically with the number of pixels and comes to dominate runtime. We show that the joint self-attention stage of a stereo transformer, whose role is to spread context across both views, can be replaced by a data-independent Walsh-Hadamard token mixer that mixes tokens globally in the transform domain at log-linear cost, while the data-dependent cross-attention that performs left-right correspondence is retained. On synthetic driving data the mixer matches the attention baseline in end-point error while reducing model compute by a factor of 2.46 and single-image inference latency by a factor of 2.65. A complexity analysis shows the benefit is governed by the ratio of sequence length to channel width, which explains why high-resolution stereo matching is a particularly favorable setting and why classification transformers are not; we confirm this token-to-channel scaling on non-stereo long-sequence benchmarks. Furthermore, we introduce a hybrid log-disparity loss function designed to up-weight small-disparity pixels corresponding to long-range objects. This approach reduces the error on distant objects without incurring any additional computational overhead.

Vision-Language-Action (VLA) models have demonstrated strong potential for embodied AI, yet their high inference latency on GPUs limits real-time deployment. Existing accelerators, such as Dadu-Corki, improve efficiency but treat VLA models as full-precision workloads, leaving substantial redundancy in both memory and computation underexploited. In this paper, we propose VQVLA, an algorithm-hardware co-design framework that accelerates VLA inference by exploiting weight similarity and execution dynamics. We first introduce MotionVQ, a motion-aware vector quantization scheme that dynamically adjusts quantization precision based on the robot's execution state, reducing memory access while preserving task success rate. We then propose a merged-centroid vectorized GEMM paradigm that operates on the codebook-index representation, eliminating redundant multiplications through spatial aggregation and temporal reuse of centroids. To realize these optimizations, we design an accelerator that efficiently supports dynamic precision selection and centroid-reuse computation. Experimental results show that VQVLA achieves 6.5x, 2.8x, 1.9x, 3.3x, and 4.3x speedup over the A100 GPU, Dadu-Corki, LUT-DLA, CodeGEMM, and ShiftAddLLM, respectively, with negligible accuracy degradation.

Multi-task inference models share a single backbone across diverse tasks, yet execute identical computation regardless of which task is active - wasting energy and cycles on task-irrelevant operations. We observe that the task command, typically available before inference begins, provides a free signal that can be exploited to skip unnecessary computation at the hardware level. We present a HW/SW co-designed approach in which a lightweight gating network, trained jointly with the backbone, predicts per-tile binary execution masks conditioned on the task input. Each tile corresponds to a fixed group of output channels (the native scheduling granularity of the accelerator), enabling masked tiles to be skipped with zero overhead. This yields a task-dependent reduction in compute, where each command activates only the subset of the network it requires, without changes to the model architecture or inference pipeline. We co-design the full system stack: a command-conditioned training procedure that learns hardware-aligned tile masks under a sparsity objective; an instruction set architecture whose instructions carry per-tile bitmask fields, allowing the hardware to skip masked tiles without software intervention; and a tiled inference accelerator with configurable parallelism, double-buffered memory, and INT8 datapath that natively supports sparse tile execution. We prototype on an AMD/Xilinx Alveo U50 FPGA and evaluate on a closed-loop visuomotor driving task in CARLA autonomous driving simulator. Task-conditional sparsity reduces FLOPs by 66-76% while maintaining driving quality. On-device latency decreases by 51-59%, from 9.12 ms to 3.74-4.44 ms (2.1-2.4x speedup), with energy per inference dropping from 263 to 108-128mJ.

We present VibeVoice-ASR-BitNet, a compressed variant of VibeVoice-ASR optimized for real-time inference on edge CPUs. We apply heterogeneous quantization tailored to the computational characteristics of each stage: the VAE acoustic tokenizer uses full-pipeline INT8 quantization (I8_S) with kernel fusion and SIMD optimization, while the autoregressive language model adopts BitNet-style ternary weights (I2_S). To preserve accuracy under aggressive compression, we employ a progressive quantization-aware training strategy. For inference, we implement custom SIMD kernels and fused operators within the ggml framework targeting both ARM and x86 platforms, achieving real-time recognition (RTF < 1) on low-thread-count CPUs. VibeVoice-ASR-BitNet is 1.6--2.3x faster than Whisper.cpp at comparable model sizes (~1.6 GB), with only modest accuracy degradation compared to the FP16 baseline.

Autoregressive text-to-speech models achieve strong naturalness but suffer from slow inference due to sequential token generation, limiting their deployment in production applications that require low latency. IndexTTS-2 is a state-of-the-art autoregressive TTS model consisting of a GPT, a flow-matching Diffusion Transformer, and a vocoder. Despite its high synthesis quality, its inference speed barely reaches real-time without streaming or batching support. We present Faster IndexTTS-2, which accelerates all neural network components of IndexTTS-2 for production deployment on GPUs using NVIDIA TensorRT and TensorRT-LLM. Faster IndexTTS-2 also enables streaming synthesis for latency-sensitive interactive applications, and batched inference across all components to maximize GPU utilization. Experiments on the Seed-TTS benchmark for both English and Chinese demonstrate up to 5.0$\times$ speedup on the autoregressive GPT and 3.6$\times$ end-to-end, with minimal degradation in word error rate, speaker similarity, and naturalness. Our methodology provides a practical reference for efficiently accelerating similar autoregressive speech models on GPUs.

Recent geometric foundation models (e.g., Metric3D, Depth Anything and UniDepth) have substantially improved monocular depth estimation (MDE) in both cross-scene generalization and metric-scale prediction, yet these gains have not translated to tiny models. We bridge this gap with DepthART (Depth Anything Rethought for Tiny Models), which is a compact MDE model for on-device deployment across diverse scenes. We first identify two capacity-driven bottlenecks in tiny models: (i) overfitting to dataset-specific distribution bias and (ii) unstable metric adaptation under camera shift, where full fine-tuning easily damages transferable geometry. Accordingly, DepthART combines two simple but effective strategies: a bias-resistant data sampling scheme to reduce distribution bias under the same training budget, and a camera-conditioned fine-tuning protocol that freezes the distilled encoder and adjusts metric scale conditioned on intrinsics while better preserving cross-dataset generalization. Across datasets, DepthART consistently surpasses previous tiny baselines in both zero-shot generalization and metric accuracy (e.g., zero-shot $δ_1$=0.964 for DepthART-S on NYUD v2), and in some cases approaches heavy models. We further provide a scalable model family, with DepthART-S reaching 347/245 FPS (strict FP32) on an RTX A6000 at $224^2/448^2$, 102 FPS (TF32) on a Orin NX 8GB, and over 15 FPS (FP32) on a Jetson Nano 4GB.

Audio-visual event recognition (AVER) has achieved significant performance improvements through transformer-based multimodal architectures. However, the high computational complexity, large memory footprint, and inference cost of these models hinder their deployment on edge and resource-constrained devices. This paper presents an efficient compression framework for hybrid cross-attention-based audiovisual event recognition by combining architectural model compression, knowledge distillation, and dynamic INT8 quantization. A high-capacity teacher model integrates VideoMAE for visual representation learning, the Audio Spectrogram Transformer (AST) for audio feature extraction, and a hybrid cross-attention fusion network for multimodal feature integration. A lightweight student model is constructed by reducing the hidden feature dimension, the number of attention heads, and the feedforward network size while preserving the overall network architecture. The student model is trained using knowledge distillation to effectively transfer discriminative knowledge from the teacher. Finally, dynamic INT8 post-training quantization is applied to further reduce the model size for efficient deployment. Experimental results on the Audio-Visual Event (AVE) dataset show that the proposed framework reduces the number of trainable parameters in the multimodal fusion module by 59.06%, with only a 2.14% decrease in classification accuracy compared with the teacher model. Furthermore, dynamic INT8 quantization reduces the model size from 10.71 MB to 2.04 MB while maintaining competitive recognition performance. These results demonstrate that the proposed framework provides an effective trade-off between recognition accuracy and computational efficiency, making it a promising solution for deployment on resource-constrained edge AI platforms.

Recent FPGAs have improved deep learning (DL) inference efficiency through dedicated tensor blocks and in-BRAM computation. ReRAM-based analog in-memory computing (IMC) pushes efficiency further, offering an order-of-magnitude improvement in compute density and energy efficiency over conventional digital logic by performing vector-matrix multiplication (VMM) directly within the ReRAM crossbar; prior work has integrated such IMC blocks into FPGAs for DL inference. However, conventional IMC designs support only static-weight VMM, leaving nonlinear operations and dynamic matrix-matrix multiplication (DIMM) to the FPGA fabric. As a result, the benefits of IMC are largely confined to static-weight models, whereas Transformer-based models, which rely on frequent nonlinear and DIMM operations, gain only limited improvement. Moreover, the ADCs within each IMC block consume more than 70% of its area and power, further limiting system efficiency and scalability. To address these limitations, we propose a novel FPGA architecture that integrates an ADC-free IMC block, replacing the conventional ADC with analog content-addressable memories (ACAMs) that natively perform nonlinear operations inside the block. To fully exploit this block, we conduct an FPGA-aware design-space exploration that determines optimal crossbar dimensions while balancing FPGA area, flexibility, and DL performance, and we develop an efficient mapping that leverages ACAMs to carry out DIMM operations, extending the applicability of IMC to attention computation. On CNN and Transformer-based benchmarks, the proposed architecture achieves up to 40x and 1.9x higher energy efficiency and 4.1x and 2.5x higher area efficiency, respectively. Overall, it significantly improves FPGA DL inference efficiency and sustains robust gains on Transformer-based workloads across long input sequences, advancing domain-specialized FPGA design.

Flow matching Vision-Language-Action (VLA) models promise precise continuous control, but their iterative denoising nature introduces fundamental incompatibilities with real-time robotics: global timestep injection invalidates KV-caching, forcing a choice between slow $O(N^2)$ re-computation or mathematically incorrect cache reuse. We present \textbf{Reflex}, a framework that enables \textit{real-time streaming inference} for flow matching policies by exploiting the \textit{Timestep-Invariance Property} -- that perception encoders are functionally independent of the denoising loop. Reflex partitions the attention context into static, sliding, and dynamic regions, enabling $O(1)$ incremental cache updates while preserving full-batch-equivalent attention outputs for fixed inputs. To ensure stability under continuous high-frequency inference, we introduce \textit{AdaRMSNorm}, an adaptive normalization layer that prevents BFloat16 numerical collapse by gating on flow phase. We further maximize throughput through an \textit{async pipeline} that decouples visual encoding from action generation, combined with \textit{operator fusion} that reduces kernel overhead. On LIBERO and Kinetix benchmarks, Reflex achieves a 2.58$\times$ inference speedup and 50Hz stable streaming, reducing reaction latency by up to 54% and enabling efficient deployment without performance degradation.

We introduce Boogu-Image-0.1, an open-source unified multimodal understanding and generation model family, comprising Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in high-quality text-to-image generation, fast inference, instruction-based editing, and bilingual (Chinese-English) text rendering. Closed-source multimodal systems like Nano-Banana-Pro and GPT-Image-2 achieve strong performance through system-level integration rather than a single model, yet their internal practices remain largely undisclosed. In this work, we demonstrate that targeted improvements in model understanding, data quality, and training pipelines, coupled with agentic inference-time scaling, can substantially enhance generation and editing performance even under highly constrained compute budgets. Comprehensive evaluations show that Boogu-Image-0.1 consistently matches or surpasses other open-source models across standard benchmarks, and achieves results approaching leading closed-source systems. Notably, this is accomplished with only 208.62 million unique images. The base model's theoretical training cost is only approximately $400K. We share practical discussions that we believe are valuable to the broader research community, and release weights, code, and recipes under Apache 2.0 to advance the open ecosystem for unified multimodal understanding and generation. Our code is available here: https://github.com/Boogu-Project/Boogu-Image.

State space models (SSMs), notably Mamba, have recently emerged as efficient alternatives to self-attention with linear computational complexity. We investigate the integration of Mamba into YOLO26, the latest non-maximum suppression (NMS)-free object detection framework, by proposing MambaPSA, a lightweight Mamba-based replacement for the C2PSA block at the end of the backbone. To complement this study, we additionally insert a bidirectional Vision Mamba (BiViM) module at the P3, P4, and P5 levels of the neck. Experiments on PASCAL VOC 2007+2012 show that MambaPSA reduces parameters by 2.9%, FLOPs by 12.1%, and improves CPU inference throughput by 17.6% (from 17 to 20 FPS) with negligible accuracy change (-0.1 mAP50:95), while the P4 BiViM placement yields the best accuracy gain (+0.9 mAP50:95). These results suggest that SSMs offer a favorable efficiency-accuracy trade-off when replacing attention-based blocks in NMS-free lightweight detectors.

Vision-Language-Action (VLA) models have achieved impressive performance on diverse embodied tasks. However, deploying VLA models on low-power onboard devices, such as the Jetson Orin, remains challenging due to their high computational complexity, which leads to substantial inference latency and low control frequency. Asynchronous inference can partially mask this latency by parallelizing action execution and subsequent inference, but it introduces two critical issues: perception-execution misalignment and long reaction time. In this paper, we propose Jetson-PI, a method for efficient VLA deployment on onboard devices via Foresight-Aligned Asynchronous Correction. To address misalignment, we train a lightweight future correction module that predicts future environment representation conditioned on committed actions, enabling the action expert to directly predict actions from the future time step. To reduce reaction time, we introduce confidence-based scheduling optimization that adaptively balances VLM and action expert invocations, complemented by system-level accelerations including CUDA graph reuse, GPU-resident intermediate buffering, and flow unrolling. Extensive experiments demonstrate that Jetson-PI achieves 8.66x and 5.41x improvements in control frequency compared with naive PyTorch and vla.cpp on NVIDIA Jetson Orin, while outperforming VLASH by 14.8% in average success rate on the LIBERO benchmark. The code of our asynchronous algorithm is available on https://github.com/PKU-SEC-Lab/Jetson-PI, and our efficient llama.cpp-based inference engine is available on https://github.com/PKU-SEC-Lab/Jetson-PI-Edge.

We present WanToFight, a generative game engine that simulates real-time, two-player The King of Fighters '97 (KOF~'97) gameplay from keyboard input. Prior generative game engines target either single-player first-person settings or non-real-time cooperative scenarios; multi-player control, real-time inference, complex physical interaction, and adversarial gameplay have not been jointly addressed. WanToFight closes this gap with three components built on the Wan-1.3B video diffusion transformer: a streaming autoregressive generator with block-causal attention and a rolling KV cache; a visually grounded Player Association module that binds each player's keyboard signal to a character identity; and a gated, locally causal keyboard injection module trained with a single-player-to-full-gameplay curriculum. A four-step DMD-distilled student paired with a pruned VAE decoder sustains 30FPS at 512x384 on a single NVIDIA RTX 5090 over the duration of a complete match. To our knowledge, WanToFight is the first generative game engine to combine multi-player control, real-time inference, complex physical interaction, and adversarial gameplay in one system.

We present a full-pipeline inference optimization for the MiMo-V2.5 model family, which combines Hybrid Sliding Window Attention (Hybrid SWA), sparse Mixture-of-Experts (MoE), and multimodal encoders. While Hybrid SWA can ideally reduce both attention compute and KVCache storage significantly compared to Full Attention, realizing these gains in production requires substantial engineering effort. We systematically optimize the KVCache system with layerwise prefetch, SWA-aware prefix cache trees, and specialized placement strategies, achieving strict $O(W)$ SWA storage and high cache hit rates. We further build GCache, a high-performance distributed cache infrastructure with RDMA-optimized networking, and develop a KVCache-affinity router to reduce computation while preserving load balancing. We also optimize for multimodal inputs, including GPU image preprocessing, parallel video decoding, and multimodal cache sharing. Together, these optimizations constitute the first large-scale LLM serving system in production that efficiently covers the Hybrid SWA + MoE + multimodal composite architecture.

The recent large video foundation model, SAM2, enables segment anything in both images and videos, serving as a powerful base model for various applications. However, many of such use cases require to operate on resource-constrained devices like mobile phones and laptops. In this work, we aim to make SAM2 more mobile-friendly by distilling the heavyweight SAM2 into a lightweight model, facilitating segment anything in both images and videos on mobile devices. To this end, we propose Hypergraphical Knowledge Distill (HyperKD), which introduces the idea of hypergraph into knowledge distillation, aiming to effectively model and transfer SAM2's generalizable and comprehensive knowledge. HyperKD consists of Temporal HyperKD and Granularity HyperKD that construct hypergraphs to explicitly model and extract the generalizable temporal knowledge and the comprehensive multi-granularity knowledge from SAM2 respectively, which are then distilled into the lightweight student model by aligning it with the constructed hypergraphs. Besides, we present MobileSAM2, a new family of lightweight SAM2 that balances efficiency and effectiveness via searching the best model architectures with HyperKD during model size reduction. Extensive experiments validate MobileSAM2 across multiple benchmarks and show promising generalization performance on embodied AI tasks.

On-device LLM inference faces a trilemma of response latency, limited hardware resources and user privacy. Full cloud inference delivers strong computing power but exposes user prompts and dialogue data, while standalone on-device inference is unfeasible for most consumer and embedded edge devices. This paper presents a privacy-centric edge-cloud collaborative LLM inference framework built on endpoint-authenticated KV cache. Local endpoints handle input preprocessing, embedding computation, adaptive feature optimization, KV cache authentication, speculative decoding and low-dimensional model head calculation, while the cloud conducts authenticated decoder inference, KV cache management, token verification and high-dimensional vocabulary projection. Endpoints fuse partial outputs, apply language-adaptive masking and sample target tokens. All transmitted data and truncated logits are quantized and AES-GCM encrypted for privacy, with core lightweight modules, draft parameters and cache access policies kept local to avoid leakage. The framework supports heterogeneous devices including CPU-only, GPU-equipped and embedded devices via optimized streaming, batching and quantized ONNX deployment. Evaluations demonstrate that the framework reduces per-token latency by up to 46.1% and downlink payloads by up to 67.4% over baseline split inference, retaining comparable performance to full cloud inference.

Sensor-rich data-driven applications increasingly use Bayesian approaches to infer latent states of dynamic systems from noisy sensor measurements and physical models. Yet the computation of the likelihood remains an essential bottleneck for accurate posteriors and performant inference. This paper presents a Bayesian filtering technique that uses processor-native uncertainty tracking for both uncertainty propagation and inference. The technique implements deterministic hierarchical importance restructuring through a native operation, giving deterministic latency and bounded memory use for arbitrary models written as program code. Benchmarks across three nonlinear state-space systems compare the approach against particle filters and Monte-Carlo-based likelihood estimators. The technique enables deterministic approximate filtering with as high as 805$\times$ average speedup against direct Monte Carlo work at matched result quality for model evaluation, and Pareto-dominant accuracy-latency trade-offs for posterior inference while remaining competitive in RMSE with baseline particle filters.

Edge deployment is often the preferred solution for industrial machine vision systems when low latency, data security, or limited connectivity are critical requirements. Several frameworks are available to optimise inference on edge devices; however, relatively few studies have systematically compared their inference-time performance under industrial deployment conditions. In this work, we present a comparative study of four widely used approaches for machine vision inference in industrial settings: plain PyTorch, ONNX Runtime, OpenVINO, and TensorRT. The evaluation focuses on inference time, covers several CPU- and GPU-based hardware platforms, and includes both conventional convolutional neural networks and a transformer-based vision model. For the hardware platforms and models evaluated, the results show that OpenVINO achieves the lowest inference time on CPUs, while TensorRT achieves the lowest inference time on GPUs. However, TensorRT does not outperform plain PyTorch for the transformer-based model considered in this study.

Large language models increasingly operate as tool-using agents, where small format, argument, or function-call errors can invalidate otherwise plausible responses. We study inference-time feed-forward network (FFN) intervention for improving structured outputs without retraining model weights. Our project began with Orthogonal Residual Projection (ORP), a direction-changing repair attempt that revealed sensitive SwiGLU FFN intervention sites but often caused more harm than fixes. We therefore propose Amplitude Gating (AG), a non-destructive alternative that preserves pretrained FFN weight directions and modulates only activation magnitudes during generation. We define a fine-grained intervention system spanning P1/P2/P3 and branch-specific P1s/P2a/P2b sites, and introduce an evaluation protocol that separates combination-oracle headroom from fixed configurations and learned gates, enforces sample-level accounting, and uses task-aware metrics for binary and partial-credit datasets. Across Qwen3.5-9B, Qwen3-8B, and Qwen2.5-7B, AG is weakly positive in aggregate but strongest on tool-structured tasks. On Qwen3.5-9B, a category-level learned gate improves tool/structured/agentic performance from 38.66% to 42.92% (+4.27 percentage points), with Hermes function-call tasks reaching about +7.6 points. On Qwen3-8B, Hermes JSON mode improves by +11.36 points. Qwen2.5-7B retains oracle headroom but current learned gates fail to capture it, showing that deployment requires model- and category-specific routing. Comparisons of entropy AG with Newton-Schulz-windowed AG show that neither family is uniformly dominant. These results identify tool-structured inference as the most credible first target for safe FFN-level inference optimization, while prospective online validation and broader cross-model evaluation remain necessary.

Physical AI systems, such as autonomous vehicles and intelligent machines, require transformer-based perception models that satisfy stringent edge latency and energy constraints. However, heterogeneous edge-GPU deployment remains limited by underutilized hardware engines and accelerator-incompatible operators, causing fragmented execution and lower throughput per watt. This paper presents Heterogeneous Frame Dispatch Scheduling (H-FraDS), a hardware-aware frame scheduling methodology for transformer inference on a recent NVIDIA edge GPU. H-FraDS routes frames across the GPU and dual deep learning accelerator (DLA) cores using fixed dispatch ratios to improve utilization under latency and power constraints. To enable scheduling, incompatible transformer components are adapted for DLA execution by reshaping tensors, approximating error function (ERF) with tanh, and replacing layer normalization with bounded tanh. The adapted model maintains a 92% F1 score, with only a 2% reduction from the original. Optical flow accelerator (OFA) is further used for inference-side optical-flow estimation. To the best of the authors' knowledge, prior work has not addressed these combined issues. Using Swin Transformer for autonomous-driving perception, H-FraDS Balanced Dispatch (1:2) achieves 125.93 FPS, a 2.36x speedup over standalone adapted-DLA execution, 4.0 FPS/W, and approximately 24 ms DLA latency, satisfying 30 FPS real-time operation; the GPU-DLA-OFA case achieves a 2.02x DLA throughput speedup.

Vision-Language Models (VLMs) have achieved success using homogeneous Transformers to process multimedia data. Recent studies show that heterogeneous structures interleaving efficient mechanisms, like linear attention, improve both performance and inference latency over homogeneous designs. However, these efforts rely on handcrafted static mixing patterns, which are sub-optimal and difficult to adapt to specific hardware. To bridge this gap, we propose Multi-Objective Search for Adaptive Inter-layer Composition (MOSAIC), a hardware-aware search method that automatically transforms homogeneous models into optimized heterogeneous architectures. MOSAIC integrates diverse efficiency mechanisms--including linear, sparse, and low-rank operators--into a unified search space. By formulating the selection as a multi-objective Mixed Integer Programming (MIP) problem, our method identifies optimal configurations that maximize downstream performance under strict hardware latency constraints. To mitigate performance degradation from structural transitions, we introduce a two-stage parameter recovery process: global off-policy distillation to stabilize internal representations, followed by a dual-teacher on-policy distillation leveraging a 235B oracle for knowledge expansion and the original 4B teacher for distributional stability. We validate MOSAIC through MOSAIC-4B, derived from Qwen3-VL-4B-Instruct. Results demonstrate that MOSAIC-4B matches the baseline's performance across multiple benchmarks while requiring less than 2% of the original training cost. Furthermore, it substantially improves inference efficiency, achieving 1.76x prefilling and 2.54x decoding speedups.

Monocular depth estimation has seen remarkable progress through foundation models achieving robust zero-shot generalization, yet their computational demands place them far beyond the reach of embedded and mobile platforms. Lightweight alternatives exist, but have been developed almost exclusively within single-domain, self-supervised paradigms, failing silently under domain shift. We present ZipDepth, a compact monocular depth network that bridges this gap by combining an efficient reparameterizable encoder-decoder with large-scale knowledge distillation from a foundation model over a large multi-domain training set. Comprising just 6.1M parameters, ZipDepth runs at real-time rates from server GPUs to power-constrained devices, achieving the best trade-off between zero-shot accuracy and deployment efficiency among lightweight models across five benchmarks, taking a significant step towards the accuracy of foundation models with 50x more parameters.

Vision-Language-Action (VLA) models have emerged as a promising approach for generalizable robotic manipulations. In particular, flow matching-based VLA models have shown remarkable success due to their capability to generate precise and smooth action sequences and capture multimodal distributions. However, the iterative denoising process in the action head acts as a major computational bottleneck, posing a critical challenge for real-time deployment. To address this challenge, we propose ActionCache, a plug-and-play external cache that opportunistically reuses past intermediate actions to warm-start generations from the vicinity of target actions, thereby drastically reducing the inference latency. Specifically, ActionCache stores the intermediate actions with compact multimodal keys, which enables retrieval from similar past contexts across different episodes or even different tasks. Experimental results in simulation and real-world environments demonstrate that ActionCache maintains high task success rates in a low-latency regime, achieving inference acceleration of up to $11.75\times$ and $34.43\times$ for representative flow-based VLA models, $π_{0.5}$ and GR00T-N1.6, respectively.

The future of World Models depends not only on scaling model capability, but also on scaling practicality and inference efficiency. High-frame-rate inference enables responsive perception, planning, and control in real-world autonomous systems. To this end, we present MoWorld, a cost-effective yet high-performance Flash World Model with an end-to-end framework spanning data generation, pre-training, distillation, and efficient inference, enabling up to 50 FPS real-time interaction with cinematic visual quality without the need of high-end GPUs. To enable large-scale real-world deployment, MoWorld jointly optimizes model capability and cost throughout the entire development pipeline. Specifically, unlike existing approaches that primarily rely on large-scale video corpora, MoWorld is built upon a scalable 3D-native data engine accumulated from our large-scale 3D vision and generative modeling pipeline, enabling the efficient construction of geometrically consistent training data across diverse real-world and synthetic environments. Based on this foundation, a curriculum cross-frame pre-training strategy for stable and scalable World Model learning, an efficient denoising-step distillation algorithm to reduce diffusion training cost, and a mixed-precision parallel inference framework for low-cost real-time deployment. MoWorld is the first real-time interactive World Model built on the Neural Processing Unit (NPU) and can achieves up to 50 FPS in such the devices, enabling practical and efficient deployment at scale. Comprehensive evaluations demonstrate that MoWorld achieves leading performance; notably, its average inference cost is only 30%-50% of that of existing World Models, providing a practical foundation for large-scale real-world applications of World Models. We also demonstrate diverse applications of MoWorld.

HAR is increasingly expected to run continuously on edge devices, yet recent LLM-based methods remain hard to deploy: raw sensor prompts are long, cloud inference adds latency and privacy risk, and fine-tuned LLM pipelines turn general-purpose models into task-specific classifiers. We present STELLA, an efficient sensor-to-LLM translation framework for on-device HAR that shifts the burden from LLM adaptation to sensor tokenization. A lightweight hierarchical tokenizer compresses an entire multi-channel inertial window into a fixed set of compact latent sensor tokens, which are projected into the embedding space of a frozen pretrained LLM and combined with a natural-language prompt for label scoring. This preserves activity-relevant temporal and cross-channel structure while keeping LLM-side computation predictable across sensor configurations. STELLA also supports on-device personalization, adapting only the lightweight tokenizer on small amounts of user-specific labelled data and augmenting inference with a local retrieval context, keeping the LLM, user data, and retrieval on device. Across seven public HAR datasets and eight benchmark settings, STELLA achieves new state-of-the-art performance, improving over prior methods by up to 11.83% F1; on-device personalization yields up to a further 21.91% F1 as user data accumulates after deployment. STELLA also outperforms representative time-series tokenizers under the same LLM pipeline and achieves real-time inference under practical mobile and edge budgets, showing that efficient sensor tokenization is a practical path toward accurate, private, and personalized LLM-based HAR on edge devices.

We present EVA-Client, an open-source framework for deployment, data collection, and evaluation of trained manipulation policies on real robots. Sitting between a policy server and the physical hardware, EVA-Client unifies the real-robot stages of the policy iteration loop within a single codebase. It makes three contributions. First, a component-decoupled architecture in which robot backends, inference strategies, and transport middlewares form an orthogonal grid: adding a robot or a strategy touches only its own layer. Second, inspectable execution through Debug, Collect, and Eval workflows, with modes ranging from open-loop simulation to continuous real-time control. Third, every evaluation run doubles as a data collection, recording full rollouts in training-ready format alongside exhaustive logs and a side-by-side comparison viewer, so each evaluation feeds the next round of training rather than ending as an unrecorded impression. EVA-Client further consolidates major real-time inference strategies, synchronous and asynchronous execution, ACT-style temporal ensembling, Real-Time Chunking, and a naive-async ablation baseline, behind a single configuration surface.

Embodied AI models now span vision-language-action (VLA) models and world-action models (WAMs), but practical deployment remains fragmented across model-specific Python stacks, backend assumptions, and robot-side glue code, especially on heterogeneous edge devices. Existing inference runtimes are designed mainly for request-response serving and therefore do not satisfy the runtime contract of embodied deployment: multi-rate execution inside closed-loop control, latency-first batch-1 inference on heterogeneous hardware, and extensible embodied interfaces beyond fixed token I/O. We present Embodied$.$cpp, a portable C++ inference runtime for embodied models. Based on an architectural analysis of representative VLA models and WAMs, Embodied$.$cpp captures a shared execution path and organizes it into five layers: input adapters, sequence builders, backbone execution, head plugins, and deployment adapters. The runtime provides modular multi-rate execution, latency-first fused inference, and extensible operator and I/O support, enabling deployment across heterogeneous devices, robots, and simulators through one backend abstraction. We evaluate Embodied$.$cpp on two VLA models, HY-VLA and pi0.5, and on a preliminary WAM benchmark using a LingBot-VA Transformer block. The VLA deployments achieve successful closed-loop execution with 100.0% and 91.0% task success rates, respectively. The WAM benchmark reduces block memory from 312.2 MiB to 88.1 MiB. These results show that Embodied$.$cpp improves deployment efficiency while preserving high accuracy across diverse embodied model architectures.

Classic 3D scene graph generation approaches fail to work in real-time due to the heavy computational cost of environment mapping and the need to generate intermediate point-cloud representations. To alleviate this issue, a recent work eschews point clouds in favor of a lightweight Gaussian distribution for each object. This approximation drastically speeds up inference and enables real-time 3D scene graph generation. However, the representation has two key weaknesses. \textbf{1)} Each object is approximated by a single 3D Gaussian, which causes a severe loss of 3D geometric detail. \textbf{2)} The discrepancy between this approximation and the true object geometry exacerbates the inaccurate merging of object candidates during online inference. To address these issues, we propose \textbf{NoPA}, which represents each object as a separate non-parametric distribution. This formulation retains 3D geometric information while preserving real-time inference of the parametric Gaussian formulation. To build upon our novel object representation, we propose a tailored merging strategy to recover coherent object instances. Specifically, we leverage maximum mean discrepancy on kernel density estimates to enable robust merging of object candidates during online exploration while minimizing added computational complexity. The key is to maintain a fixed particle set per object. Furthermore, to rectify the relation loss caused by misclassified objects, NoPA propagates relationships between objects with high affinity. Experiments show that NoPA substantially outperforms current methods without sacrificing real-time inference speed.

Modern vision-language models (VLMs) for driving assistants typically treat vehicle dynamics as a black box, resulting in decisions that lack awareness of the vehicle's real-time electro-mechanical state. To bridge this gap, we introduce the Electro-Visual-Language Assistant (EVLA) -- a novel framework that combines multi-modal scene understanding with real-time perception of the electrified powertrain state (e.g., motor torque, battery SOC). Our approach features two key innovations: first, a Unified Co-State Encoder (UCSE) that fuses visual, textual, and vehicle-state inputs into a shared latent representation, augmented with an Energy-Efficiency Field to model spatial energy costs; and second, an Electro-aware Structured Reasoning Chain (ESRC), which replaces external chain-of-thought prompting with an internal, deterministic reasoning process grounded in physical constraints and optimization objectives. Trained end-to-end with a physics-guided joint loss, EVLA learns to generate context-aware and energy-optimal driving decisions. Extensive evaluations on a driving QA benchmark demonstrate that EVLA substantially outperforms strong fine-tuned VLM baselines, improving the final score by +0.0871 and accuracy by +5.6%. Ablation studies validate the necessity of each component, and efficiency analyses show that EVLA achieves 36% faster inference than multi-stage pipelines. This work underscores that integrating vehicle-state awareness and structured physical reasoning is crucial for developing next-generation, physically-grounded driving assistants.

Long-context LLM inference faces a fundamental conflict: head-adaptive compression algorithms (e.g., Top-$p$ nucleus sampling) offer superior accuracy by dynamically fluctuating memory budgets, yet modern inference engines (e.g., vLLM) demand rigid, static memory patterns to leverage CUDA Graphs and PagedAttention. We resolve this ``Static-Dynamic'' mismatch with HARD-KV, a unified framework that that bridges dynamic selection with rigid system constraints. HARD-KV introduces a Cascade Cache hierarchy, managing the token lifecycle across dense, sparse, and condensed tiers. Crucially, we propose a Logits Calibration mechanism that normalizes diverse importance metrics into a unified probability space, enabling consistent Top-$p$ budgeting across heterogeneous heads. To bridge the efficiency gap, we offer a system-level solution, which rewrites fragmented, dynamic indices into contiguous physical layouts compatible with high-performance inference engine. Extensive experiments on math-reasoning benchmarks (AIME, U-Math) verify that HARD-KV achieves up to 2$\times$ throughput improvement over static baselines while maintaining high-fidelity generation in 10k+ token scenarios. Code is available at https://github.com/SuDIS-ZJU/HARDInfer.

Vision-Language-Action (VLA) models have become a dominant paradigm for embodied intelligence. However, most existing approaches are built on large-scale transformers, resulting in substantial inference latency and energy consumption that limit their practical deployment in low-power, real-time scenarios. We propose SpikeVLA, a spiking VLA architecture for embodied navigation with energy-efficient inference, consisting of three key components. (i) A spiking vision encoder, Spike-V, that replaces dense continuous layers with event-driven spiking layers to reduce the energy consumption of visual representation learning. (ii) A multi-modal spiking large language model, Spike-L, that reformulates cross-modal reasoning with spiking dynamics and token-level event-driven sparsity to further lower computational cost. (iii) A spiking action policy network, Spike-A employs Laplacian-kernel population coding with a multi-layer fully connected SNN, and decodes spiking activities into stable and robust continuous control with energy-efficient inference under low-power constraints. Experiments on navigation and robotic control tasks show that SpikeVLA significantly reduces energy consumption and computational cost while maintaining competitive performance, highlighting its potential for low-power, real-time embodied intelligence.

While test-time adaptation (TTA) empowers vision-language models to adapt without costly retraining, it remains highly vulnerable to out-of-distribution (OOD) outliers prevalent in real-world applications. This discrepancy motivates Noisy TTA (NTTA), an online task to filter noisy OOD samples on the fly while maximizing in-distribution (ID) classification accuracy. Existing zero-shot NTTA approaches typically rely on test-time discriminative training, leading to overconfident misclassifications and significantly degraded inference efficiency. To address these limitations, we propose a novel framework named Dual Distribution Estimation (DDE), shifting the zero-shot NTTA paradigm from instance-level learning to training-free Gaussian distribution modeling. DDE incorporates two novel modules: Positive Feature Distribution Estimation (PFDE) and Negative Label Distribution Estimation (NLDE). PFDE explicitly models class-wise inclusion and exclusion Gaussian distributions to formulate a calibrated contrastive score, robustly enhancing ID accuracy. In parallel, NLDE improves OOD identification by explicitly modeling the negative label distribution to mine highly discriminative labels, effectively mitigating spurious correlations. Extensive experiments show that on the large-scale ImageNet benchmark, DDE achieves an improvement of 3.70% in harmonic mean accuracy and reduces the FPR95 for OOD detection by 6.20%, while ensuring highly scalable and efficient online inference. Furthermore, DDE is zero-shot and training-free, demonstrating remarkable robustness in data-scarce scenarios. Codes are available at https://github.com/ZhuWenjie98/DDE.

Recent advances in stereo matching have achieved remarkable accuracy, but often rely on large models, heavy computation, or additional foundation-model priors, making them difficult to deploy on resource-constrained platforms. In contrast, efficient stereo models offer faster inference but are commonly considered less capable of strong zero-shot generalization. In this paper, we challenge this assumption by introducing Lite Any Stereo V2 (LAS2), an ultra-fast model series designed for efficient zero-shot stereo matching. LAS2 is developed from both architecture and training perspectives. Architecturally, we revisit efficient stereo design under practical deployment settings and propose a 2D-only cost aggregation framework, optimized for real inference latency rather than theoretical MACs alone. For training, we develop a three-stage strategy that combines synthetic supervision, self-distillation, and real-world knowledge distillation. To improve the reliability of real-world pseudo supervision, we further introduce pseudo-label filtering and an error-clamping operation, enabling smoother synthetic-to-real transfer. We instantiate LAS2 as a family of models, including feed-forward variants for different efficiency budgets and an iterative variant for higher accuracy. Extensive experiments show that LAS2 achieves state-of-the-art accuracy among efficient stereo methods while maintaining significantly lower latency. Specifically, LAS2-H achieves stronger overall zero-shot performance than the iterative method Fast-FoundationStereo, with 1.8x and 2.7x faster inference on H200 and Orin, respectively. The project page, demos, and code are available at https://tomtomtommi.github.io/LiteAnyStereoV2/.