Meng Li

Zechun Liu, Barlas Oguz, Aasish Pappu et al. · meta-ai, pku

Modern pre-trained transformers have rapidly advanced the state-of-the-art in machine learning, but have also grown in parameters and computational complexity, making them increasingly difficult to deploy in resource-constrained environments. Binarization of the weights and activations of the network can significantly alleviate these issues, however, is technically challenging from an optimization perspective. In this work, we identify a series of improvements that enables binary transformers at a much higher accuracy than what was possible previously. These include a two-set binarization scheme, a novel elastic binary activation function with learned parameters, and a method to quantize a network to its limit by successively distilling higher precision models into lower precision students. These approaches allow for the first time, fully binarized transformer models that are at a practical level of accuracy, approaching a full-precision BERT baseline on the GLUE language understanding benchmark within as little as 5.9%. Code and models are available at: https://github.com/facebookresearch/bit.

Ganesh Jawahar, Haichuan Yang, Yunyang Xiong et al. · meta-ai, pku

Weight-sharing supernets are crucial for performance estimation in cutting-edge neural architecture search (NAS) frameworks. Despite their ability to generate diverse subnetworks without retraining, the quality of these subnetworks is not guaranteed due to weight sharing. In NLP tasks like machine translation and pre-trained language modeling, there is a significant performance gap between supernet and training from scratch for the same model architecture, necessitating retraining post optimal architecture identification. This study introduces a solution called mixture-of-supernets, a generalized supernet formulation leveraging mixture-of-experts (MoE) to enhance supernet model expressiveness with minimal training overhead. Unlike conventional supernets, this method employs an architecture-based routing mechanism, enabling indirect sharing of model weights among subnetworks. This customization of weights for specific architectures, learned through gradient descent, minimizes retraining time, significantly enhancing training efficiency in NLP. The proposed method attains state-of-the-art (SoTA) performance in NAS for fast machine translation models, exhibiting a superior latency-BLEU tradeoff compared to HAT, the SoTA NAS framework for machine translation. Furthermore, it excels in NAS for building memory-efficient task-agnostic BERT models, surpassing NAS-BERT and AutoDistil across various model sizes. The code can be found at: https://github.com/UBC-NLP/MoS.

Wenxuan Zeng, Meng Li, Wenjie Xiong et al.

Secure multi-party computation (MPC) enables computation directly on encrypted data and protects both data and model privacy in deep learning inference. However, existing neural network architectures, including Vision Transformers (ViTs), are not designed or optimized for MPC and incur significant latency overhead. We observe Softmax accounts for the major latency bottleneck due to a high communication complexity, but can be selectively replaced or linearized without compromising the model accuracy. Hence, in this paper, we propose an MPC-friendly ViT, dubbed MPCViT, to enable accurate yet efficient ViT inference in MPC. Based on a systematic latency and accuracy evaluation of the Softmax attention and other attention variants, we propose a heterogeneous attention optimization space. We also develop a simple yet effective MPC-aware neural architecture search algorithm for fast Pareto optimization. To further boost the inference efficiency, we propose MPCViT+, to jointly optimize the Softmax attention and other network components, including GeLU, matrix multiplication, etc. With extensive experiments, we demonstrate that MPCViT achieves 1.9%, 1.3% and 3.6% higher accuracy with 6.2x, 2.9x and 1.9x latency reduction compared with baseline ViT, MPCFormer and THE-X on the Tiny-ImageNet dataset, respectively. MPCViT+ further achieves a better Pareto front compared with MPCViT. The code and models for evaluation are available at https://github.com/PKU-SEC-Lab/mpcvit.

Yonggan Fu, Haichuan Yang, Jiayi Yuan et al.

Efficient deep neural network (DNN) models equipped with compact operators (e.g., depthwise convolutions) have shown great potential in reducing DNNs' theoretical complexity (e.g., the total number of weights/operations) while maintaining a decent model accuracy. However, existing efficient DNNs are still limited in fulfilling their promise in boosting real-hardware efficiency, due to their commonly adopted compact operators' low hardware utilization. In this work, we open up a new compression paradigm for developing real-hardware efficient DNNs, leading to boosted hardware efficiency while maintaining model accuracy. Interestingly, we observe that while some DNN layers' activation functions help DNNs' training optimization and achievable accuracy, they can be properly removed after training without compromising the model accuracy. Inspired by this observation, we propose a framework dubbed DepthShrinker, which develops hardware-friendly compact networks via shrinking the basic building blocks of existing efficient DNNs that feature irregular computation patterns into dense ones with much improved hardware utilization and thus real-hardware efficiency. Excitingly, our DepthShrinker framework delivers hardware-friendly compact networks that outperform both state-of-the-art efficient DNNs and compression techniques, e.g., a 3.06% higher accuracy and 1.53$\times$ throughput on Tesla V100 over SOTA channel-wise pruning method MetaPruning. Our codes are available at: https://github.com/facebookresearch/DepthShrinker.

Xin Dong, Barbara De Salvo, Meng Li et al. · pku

We design deep neural networks (DNNs) and corresponding networks' splittings to distribute DNNs' workload to camera sensors and a centralized aggregator on head mounted devices to meet system performance targets in inference accuracy and latency under the given hardware resource constraints. To achieve an optimal balance among computation, communication, and performance, a split-aware neural architecture search framework, SplitNets, is introduced to conduct model designing, splitting, and communication reduction simultaneously. We further extend the framework to multi-view systems for learning to fuse inputs from multiple camera sensors with optimal performance and systemic efficiency. We validate SplitNets for single-view system on ImageNet as well as multi-view system on 3D classification, and show that the SplitNets framework achieves state-of-the-art (SOTA) performance and system latency compared with existing approaches.

Can Peng, Kun Zhao, Tianren Wang et al.

The continual appearance of new objects in the visual world poses considerable challenges for current deep learning methods in real-world deployments. The challenge of new task learning is often exacerbated by the scarcity of data for the new categories due to rarity or cost. Here we explore the important task of Few-Shot Class-Incremental Learning (FSCIL) and its extreme data scarcity condition of one-shot. An ideal FSCIL model needs to perform well on all classes, regardless of their presentation order or paucity of data. It also needs to be robust to open-set real-world conditions and be easily adapted to the new tasks that always arise in the field. In this paper, we first reevaluate the current task setting and propose a more comprehensive and practical setting for the FSCIL task. Then, inspired by the similarity of the goals for FSCIL and modern face recognition systems, we propose our method -- Augmented Angular Loss Incremental Classification or ALICE. In ALICE, instead of the commonly used cross-entropy loss, we propose to use the angular penalty loss to obtain well-clustered features. As the obtained features not only need to be compactly clustered but also diverse enough to maintain generalization for future incremental classes, we further discuss how class augmentation, data augmentation, and data balancing affect classification performance. Experiments on benchmark datasets, including CIFAR100, miniImageNet, and CUB200, demonstrate the improved performance of ALICE over the state-of-the-art FSCIL methods.

Meng Li, Shangyin Gao, Yihui Feng et al. · pku

In recent years, with the development of deep neural networks, end-to-end optimized image compression has made significant progress and exceeded the classic methods in terms of rate-distortion performance. However, most learning-based image compression methods are unlabeled and do not consider image semantics or content when optimizing the model. In fact, human eyes have different sensitivities to different content, so the image content also needs to be considered. In this paper, we propose a content-oriented image compression method, which handles different kinds of image contents with different strategies. Extensive experiments show that the proposed method achieves competitive subjective results compared with state-of-the-art end-to-end learned image compression methods or classic methods.

Yufei Jia, Zhanxiang Cao, Mingrui Yu et al.

Simulation-based RL for contemporary robot control is increasingly organized around GPU-resident simulation: physics, rollout collection, and learning are placed on a single GPU-centric execution path. This paradigm has greatly improved training speed, but it has also encouraged a default assumption that efficient training requires physics to reside on the GPU. We revisit this assumption. Our view is that, in simulation-dominated robot control, the essential question is not which processor runs physics, but whether simulation throughput, policy learning, and runtime synchronization form an efficient end-to-end loop. We present UniLab, a heterogeneous CPU-simulation / GPU-learning architecture that decouples CPU-parallel simulation from GPU policy updates through a unified runtime for data movement, buffering, and synchronization. UniLab is implemented as a complete and extensible training system using MuJoCoUni and MotrixSim CPU-batched physics backends, supporting PPO, SAC, FlashSAC, TD3, and APPO. On representative simulation-based robot control tasks, UniLab improves end-to-end training efficiency by 3--10$\times$ under the same hardware configuration, while reducing dependence on the NVIDIA CUDA-based software stack and supporting cross-platform execution on the Apple macOS platform and the AMD ROCm and Intel XPU accelerator backends. These results show that GPU simulation is an effective path to efficient training, but not a necessary one, broadening the practical system choices available for robot RL training. Project page: https://github.com/unilabsim/UniLab.

Shichao Fan, Kun Wu, Zhengping Che et al.

Recent progress in large-scale robotic datasets and vision-language models (VLMs) has advanced research on vision-language-action (VLA) models. However, existing VLA models still face two fundamental challenges: (i) producing precise low-level actions from high-dimensional observations, (ii) bridging domain gaps across heterogeneous data sources, including diverse robot embodiments and human demonstrations. Existing methods often encode latent variables from either visual dynamics or robotic actions to guide policy learning, but they fail to fully exploit the complementary multi-modal knowledge present in large-scale, heterogeneous datasets. In this work, we present X Robotic Model 1 (XR-1), a novel framework for versatile and scalable VLA learning across diverse robots, tasks, and environments. XR-1 introduces the \emph{Unified Vision-Motion Codes (UVMC)}, a discrete latent representation learned via a dual-branch VQ-VAE that jointly encodes visual dynamics and robotic motion. UVMC addresses these challenges by (i) serving as an intermediate representation between the observations and actions, and (ii) aligning multimodal dynamic information from heterogeneous data sources to capture complementary knowledge. To effectively exploit UVMC, we propose a three-stage training paradigm: (i) self-supervised UVMC learning, (ii) UVMC-guided pretraining on large-scale cross-embodiment robotic datasets, and (iii) task-specific post-training. We validate XR-1 through extensive real-world experiments with more than 14,000 rollouts on six different robot embodiments, spanning over 120 diverse manipulation tasks. XR-1 consistently outperforms state-of-the-art baselines such as $π_{0.5}$, $π_0$, RDT, UniVLA, and GR00T-N1.5 while demonstrating strong generalization to novel objects, background variations, distractors, and illumination changes. Our project is at https://xr-1-vla.github.io/.

Lemeng Wu, Dilin Wang, Meng Li et al. · pku

Fusing 3D LiDAR features with 2D camera features is a promising technique for enhancing the accuracy of 3D detection, thanks to their complementary physical properties. While most of the existing methods focus on directly fusing camera features with raw LiDAR point clouds or shallow-level 3D features, it is observed that directly combining 2D and 3D features in deeper layers actually leads to a decrease in accuracy due to feature misalignment. The misalignment, which stems from the aggregation of features learned from large receptive fields, becomes increasingly more severe as we delve into deeper layers. In this paper, we propose PathFusion as a solution to enable the alignment of semantically coherent LiDAR-camera deep feature fusion. PathFusion introduces a path consistency loss at multiple stages within the network, encouraging the 2D backbone and its fusion path to transform 2D features in a way that aligns semantically with the transformation of the 3D backbone. This ensures semantic consistency between 2D and 3D features, even in deeper layers, and amplifies the usage of the network's learning capacity. We apply PathFusion to improve a prior-art fusion baseline, Focals Conv, and observe an improvement of over 1.6% in mAP on the nuScenes test split consistently with and without testing-time data augmentations, and moreover, PathFusion also improves KITTI $\text{AP}_{\text{3D}}$ (R11) by about 0.6% on the moderate level.

Tianshi Xu, Meng Li, Runsheng Wang et al.

Efficient networks, e.g., MobileNetV2, EfficientNet, etc, achieves state-of-the-art (SOTA) accuracy with lightweight computation. However, existing homomorphic encryption (HE)-based two-party computation (2PC) frameworks are not optimized for these networks and suffer from a high inference overhead. We observe the inefficiency mainly comes from the packing algorithm, which ignores the computation characteristics and the communication bottleneck of homomorphically encrypted depthwise convolutions. Therefore, in this paper, we propose Falcon, an effective dense packing algorithm for HE-based 2PC frameworks. Falcon features a zero-aware greedy packing algorithm and a communication-aware operator tiling strategy to improve the packing density for depthwise convolutions. Compared to SOTA HE-based 2PC frameworks, e.g., CrypTFlow2, Iron and Cheetah, Falcon achieves more than 15.6x, 5.1x and 1.8x latency reduction, respectively, at operator level. Meanwhile, at network level, Falcon allows for 1.4% and 4.2% accuracy improvement over Cheetah on CIFAR-100 and TinyImagenet datasets with iso-communication, respecitvely.

Meng Li, Senbo Wang, Weihao Yuan et al.

Monocular depth estimation is an ambiguous problem, thus global structural cues play an important role in current data-driven single-view depth estimation methods. Panorama images capture the complete spatial information of their surroundings utilizing the equirectangular projection which introduces large distortion. This requires the depth estimation method to be able to handle the distortion and extract global context information from the image. In this paper, we propose an end-to-end deep network for monocular panorama depth estimation on a unit spherical surface. Specifically, we project the feature maps extracted from equirectangular images onto unit spherical surface sampled by uniformly distributed grids, where the decoder network can aggregate the information from the distortion-reduced feature maps. Meanwhile, we propose a global cross-attention-based fusion module to fuse the feature maps from skip connection and enhance the ability to obtain global context. Experiments are conducted on five panorama depth estimation datasets, and the results demonstrate that the proposed method substantially outperforms previous state-of-the-art methods. All related codes will be open-sourced in the upcoming days.

Junyi He, Meimei Wu, Meng Li et al.

Multimodal emotion recognition has attracted much attention recently. Fusing multiple modalities effectively with limited labeled data is a challenging task. Considering the success of pre-trained model and fine-grained nature of emotion expression, it is reasonable to take these two aspects into consideration. Unlike previous methods that mainly focus on one aspect, we introduce a novel multi-granularity framework, which combines fine-grained representation with pre-trained utterance-level representation. Inspired by Transformer TTS, we propose a multilevel transformer model to perform fine-grained multimodal emotion recognition. Specifically, we explore different methods to incorporate phoneme-level embedding with word-level embedding. To perform multi-granularity learning, we simply combine multilevel transformer model with Albert. Extensive experimental results show that both our multilevel transformer model and multi-granularity model outperform previous state-of-the-art approaches on IEMOCAP dataset with text transcripts and speech signal.

Shuzhang Zhong, Ling Liang, Yuan Wang et al.

Mixture-of-Experts (MoE) models are designed to enhance the efficiency of large language models (LLMs) without proportionally increasing the computational demands. However, their deployment on edge devices still faces significant challenges due to high on-demand loading overheads from managing sparsely activated experts. This paper introduces AdapMoE, an algorithm-system co-design framework for efficient MoE inference. AdapMoE features adaptive expert gating and management to reduce the on-demand loading overheads. We observe the heterogeneity of experts loading across layers and tokens, based on which we propose a sensitivity-based strategy to adjust the number of activated experts dynamically. Meanwhile, we also integrate advanced prefetching and cache management techniques to further reduce the loading latency. Through comprehensive evaluations on various platforms, we demonstrate AdapMoE consistently outperforms existing techniques, reducing the average number of activated experts by 25% and achieving a 1.35x speedup without accuracy degradation. Code is available at: https://github.com/PKU-SEC-Lab/AdapMoE.

Renjie Wei, Zechun Liu, Yuchen Fan et al.

Deep neural networks for image super-resolution (SR) have demonstrated superior performance. However, the large memory and computation consumption hinders their deployment on resource-constrained devices. Binary neural networks (BNNs), which quantize the floating point weights and activations to 1-bit can significantly reduce the cost. Although BNNs for image classification have made great progress these days, existing BNNs for SR still suffer from a large performance gap between the FP SR networks. To this end, we observe the activation distribution in SR networks and find much larger pixel-to-pixel, channel-to-channel, layer-to-layer, and image-to-image variation in the activation distribution than image classification networks. However, existing BNNs for SR fail to capture these variations that contain rich information for image reconstruction, leading to inferior performance. To address this problem, we propose SCALES, a binarization method for SR networks that consists of the layer-wise scaling factor, the spatial re-scaling method, and the channel-wise re-scaling method, capturing the layer-wise, pixel-wise, and channel-wise variations efficiently in an input-dependent manner. We evaluate our method across different network architectures and datasets. For CNN-based SR networks, our binarization method SCALES outperforms the prior art method by 0.2dB with fewer parameters and operations. With SCALES, we achieve the first accurate binary Transformer-based SR network, improving PSNR by more than 1dB compared to the baseline method.

Shuzhang Zhong, Meng Li, Yun Liang et al.

Memory-aware network scheduling is becoming increasingly important for deep neural network (DNN) inference on resource-constrained devices. However, due to the complex cell-level and network-level topologies, memory-aware scheduling becomes very challenging. While previous algorithms all suffer from poor scalability, in this paper, we propose an efficient memory-aware scheduling framework based on iterative computation graph optimization. Our framework features an iterative graph fusion algorithm that simplifies the computation graph while preserving the scheduling optimality. We further propose an integer linear programming formulation together with topology-aware variable pruning to schedule the simplified graph efficiently. We evaluate our method against prior-art algorithms on different networks and demonstrate that our method outperforms existing techniques in all the benchmarks, reducing the peak memory footprint by 13.4%, and achieving better scalability for networks with complex network-level topologies.

Meng Li, Brian Lovell

Current medical image synthetic augmentation techniques rely on intensive use of generative adversarial networks (GANs). However, the nature of GAN architecture leads to heavy computational resources to produce synthetic images and the augmentation process requires multiple stages to complete. To address these challenges, we introduce a novel generative meta curriculum learning method that trains the task-specific model (student) end-to-end with only one additional teacher model. The teacher learns to generate curriculum to feed into the student model for data augmentation and guides the student to improve performance in a meta-learning style. In contrast to the generator and discriminator in GAN, which compete with each other, the teacher and student collaborate to improve the student's performance on the target tasks. Extensive experiments on the histopathology datasets show that leveraging our framework results in significant and consistent improvements in classification performance.

Meng Li, Chaoyi Li, Can Peng et al.

Deep learning techniques have become widely utilized in histopathology image classification due to their superior performance. However, this success heavily relies on the availability of substantial labeled data, which necessitates extensive and costly manual annotation by domain experts. To address this challenge, researchers have recently employed generative models to synthesize data for augmentation, thereby enhancing classification model performance. Traditionally, this involves generating synthetic data first and then training the classification model with both synthetic and real data, which creates a two-stage, time-consuming workflow. To overcome this limitation, we propose an innovative unified framework that integrates the data generation and model training stages into a unified process. Our approach utilizes a pure Vision Transformer (ViT)-based conditional Generative Adversarial Network (cGAN) model to simultaneously handle both image synthesis and classification. An additional classification head is incorporated into the cGAN model to enable simultaneous classification of histopathology images. To improve training stability and enhance the quality of generated data, we introduce a conditional class projection technique that helps maintain class separation during the generation process. We also employ a dynamic multi-loss weighting mechanism to effectively balance the losses of the classification tasks. Furthermore, our selective augmentation mechanism actively selects the most suitable generated images for data augmentation to further improve performance. Extensive experiments on histopathology datasets show that our unified synthetic augmentation framework consistently enhances the performance of histopathology image classification models.

Tianshi Xu, Huifeng Wen, Meng Li

LLM agents are shaped not only by their language models, but also by the runtime harness that mediates observation, tool use, action execution, feedback interpretation, and trajectory control. While existing agent adaptation methods mainly update model parameters, many failures in deterministic, rule-governed domains stem from mismatches at the model--environment interface. We propose Life-Harness, a lifecycle-aware runtime harness that improves frozen LLM agents without changing model weights or evaluation environments. Life-Harness evolves from training trajectories by converting recurring interaction failures into reusable interventions across environment contracts, procedural skills, action realization, and trajectory regulation, and remains fixed during held-out evaluation. On seven deterministic environments from $τ$-bench, $τ^2$-bench, and AgentBench, Life-Harness improves 116 out of 126 model--environment settings across 18 model backbones, with an average relative improvement of 88.5%. Harnesses evolved only from Qwen3-4B-Instruct trajectories transfer to 17 other models, showing that Life-Harness captures reusable environment-side structure rather than model-specific behavior. These results position runtime interface adaptation as a complementary alternative to model-centric agent training. Code is available at GitHub.

Zhibo Xing, Zijian Zhang, Ziang Zhang et al.

Over recent decades, machine learning has significantly advanced network communication, enabling improved decision-making, user behavior analysis, and fault detection. Decentralized approaches, where participants exchange computation results instead of raw private data, mitigate these risks but introduce challenges related to trust and verifiability. A critical issue arises: How can one ensure the integrity and validity of computation results shared by other participants? Existing survey articles predominantly address security and privacy concerns in decentralized machine learning, whereas this survey uniquely highlights the emerging issue of verifiability. Recognizing the critical role of zero-knowledge proofs in ensuring verifiability, we present a comprehensive review of Zero-Knowledge Proof-based Verifiable Machine Learning (ZKP-VML). To clarify the research problem, we present a definition of ZKP-VML consisting of four algorithms, along with several corresponding key security properties. Besides, we provide an overview of the current research landscape by systematically organizing the research timeline and categorizing existing schemes based on their security properties. Furthermore, through an in-depth analysis of each existing scheme, we summarize their technical contributions and optimization strategies, aiming to uncover common design principles underlying ZKP-VML schemes. Building on the reviews and analysis presented, we identify current research challenges and suggest future research directions. To the best of our knowledge, this is the most comprehensive survey to date on verifiable decentralized machine learning and ZKP-VML.

Roberto Goya-Maldonado, Tracy Erwin-Grabner, Ling-Li Zeng et al.

Major depressive disorder (MDD) is a complex psychiatric disorder that affects the lives of hundreds of millions of individuals around the globe. Even today, researchers debate if morphological alterations in the brain are linked to MDD, likely due to the heterogeneity of this disorder. The application of deep learning tools to neuroimaging data, capable of capturing complex non-linear patterns, has the potential to provide diagnostic and predictive biomarkers for MDD. However, previous attempts to demarcate MDD patients and healthy controls (HC) based on segmented cortical features via linear machine learning approaches have reported low accuracies. Here, we used globally representative data from the ENIGMA-MDD working group containing 7,012 participants from 30 sites (N=2,772 MDD and N=4,240 HC), which allows a comprehensive analysis with generalizable results. Based on the hypothesis that integration of vertex-wise cortical features can improve classification performance, we evaluated the classification of a DenseNet and a Support Vector Machine (SVM), with the expectation that the former would outperform the latter. We found that both classifiers exhibited close to chance performance (balanced accuracy DenseNet: 51%; SVM: 53%), when estimated on unseen sites. Slightly higher classification performance (balanced accuracy DenseNet: 58%; SVM: 55%) was found when the cross-validation folds contained subjects from all sites, indicating site effect. In conclusion, the integration of vertex-wise morphometric features and the use of the non-linear classifier did not lead to the differentiability between MDD and HC. Our results support the notion that MDD classification on this combination of such features and classifiers is unfeasible. Perhaps more sophisticated integration of multimodal information may lead to a higher performance in this diagnostic task.

Meng Li, Dongyang Shi, Junjun Wang

In this paper, a linearized Crank-Nicolson Galerkin finite element method (FEM) for generalized Ginzburg-Landau equation (GLE) is considered, in which, the difference method in time and the standard Galerkin FEM are employed. Based on the linearized Crank-Nicolson difference method in time and the standard Galerkin finite element method with bilinear element in space, the time-discrete and space-time discrete systems are both constructed. We focus on a rigorous analysis and consideration of unconditional superconvergence error estimates of the discrete schemes. Firstly, by virtue of the temporal error results, the regularity for the time-discrete system is presented. Secondly, the classical Ritz projection is used to obtain the spatial error with order $O(h^2)$ in the sense of $L^2-$norm. Thanks to the relationship between the Ritz projection and the interpolated projection, the superclose estimate with order $O(τ^2 + h^2)$ in the sense of $H^1-$norm is derived. Thirdly, it follows from the interpolated postprocessing technique that the global superconvergence result is deduced. Finally, some numerical results are provided to confirm the theoretical analysis.

Meng Li, Yahan Yu, Yi Yang et al.

Stroke extraction of Chinese characters plays an important role in the field of character recognition and generation. The most existing character stroke extraction methods focus on image morphological features. These methods usually lead to errors of cross strokes extraction and stroke matching due to rarely using stroke semantics and prior information. In this paper, we propose a deep learning-based character stroke extraction method that takes semantic features and prior information of strokes into consideration. This method consists of three parts: image registration-based stroke registration that establishes the rough registration of the reference strokes and the target as prior information; image semantic segmentation-based stroke segmentation that preliminarily separates target strokes into seven categories; and high-precision extraction of single strokes. In the stroke registration, we propose a structure deformable image registration network to achieve structure-deformable transformation while maintaining the stable morphology of single strokes for character images with complex structures. In order to verify the effectiveness of the method, we construct two datasets respectively for calligraphy characters and regular handwriting characters. The experimental results show that our method strongly outperforms the baselines. Code is available at https://github.com/MengLi-l1/StrokeExtraction.

DeepSeek-AI, Daya Guo, Dejian Yang et al. · stanford, tsinghua

We introduce our first-generation reasoning models, DeepSeek-R1-Zero and DeepSeek-R1. DeepSeek-R1-Zero, a model trained via large-scale reinforcement learning (RL) without supervised fine-tuning (SFT) as a preliminary step, demonstrates remarkable reasoning capabilities. Through RL, DeepSeek-R1-Zero naturally emerges with numerous powerful and intriguing reasoning behaviors. However, it encounters challenges such as poor readability, and language mixing. To address these issues and further enhance reasoning performance, we introduce DeepSeek-R1, which incorporates multi-stage training and cold-start data before RL. DeepSeek-R1 achieves performance comparable to OpenAI-o1-1217 on reasoning tasks. To support the research community, we open-source DeepSeek-R1-Zero, DeepSeek-R1, and six dense models (1.5B, 7B, 8B, 14B, 32B, 70B) distilled from DeepSeek-R1 based on Qwen and Llama.

DeepSeek-AI, Aixin Liu, Bei Feng et al. · pku

We present DeepSeek-V2, a strong Mixture-of-Experts (MoE) language model characterized by economical training and efficient inference. It comprises 236B total parameters, of which 21B are activated for each token, and supports a context length of 128K tokens. DeepSeek-V2 adopts innovative architectures including Multi-head Latent Attention (MLA) and DeepSeekMoE. MLA guarantees efficient inference through significantly compressing the Key-Value (KV) cache into a latent vector, while DeepSeekMoE enables training strong models at an economical cost through sparse computation. Compared with DeepSeek 67B, DeepSeek-V2 achieves significantly stronger performance, and meanwhile saves 42.5% of training costs, reduces the KV cache by 93.3%, and boosts the maximum generation throughput to 5.76 times. We pretrain DeepSeek-V2 on a high-quality and multi-source corpus consisting of 8.1T tokens, and further perform Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) to fully unlock its potential. Evaluation results show that, even with only 21B activated parameters, DeepSeek-V2 and its chat versions still achieve top-tier performance among open-source models.

Zebin Yang, Yijiahao Qi, Tong Xie et al.

Vision-Language-Action (VLA) models have shown remarkable success in robotic tasks like manipulation by fusing a language model's reasoning with a vision model's 3D understanding. However, their high computational cost remains a major obstacle for real-world applications that require real-time performance. We observe that the actions within a task have varying levels of importance: critical steps demand high precision, while less important ones can tolerate more variance. Leveraging this insight, we propose DySL-VLA, a novel framework that addresses computational cost by dynamically skipping VLA layers based on each action's importance. DySL-VLA categorizes its layers into two types: informative layers, which are consistently executed, and incremental layers, which can be selectively skipped. To intelligently skip layers without sacrificing accuracy, we invent a prior-post skipping guidance mechanism to determine when to initiate layer-skipping. We also propose a skip-aware two-stage knowledge distillation algorithm to efficiently train a standard VLA into a DySL-VLA. Our experiments indicate that DySL-VLA achieves 2.1% improvement in success length over Deer-VLA on the Calvin dataset, while simultaneously reducing trainable parameters by a factor of 85.7 and providing a 3.75x speedup relative to the RoboFlamingo baseline at iso-accuracy. Our code is available on https://github.com/PKU-SEC-Lab/DYSL_VLA.

Shuanghao Bai, Meng Li, Xinyuan Lv et al.

Humans achieve complex manipulation through coordinated whole-body control, whereas most Vision-Language-Action (VLA) models treat robot body parts largely independently, making high-DoF humanoid control challenging and often unstable. We present HEX, a state-centric framework for coordinated manipulation on full-sized bipedal humanoid robots. HEX introduces a humanoid-aligned universal state representation for scalable learning across heterogeneous embodiments, and incorporates a Mixture-of-Experts Unified Proprioceptive Predictor to model whole-body coordination and temporal motion dynamics from large-scale multi-embodiment trajectory data. To efficiently capture temporal visual context, HEX uses lightweight history tokens to summarize past observations, avoiding repeated encoding of historical images during inference. It further employs a residual-gated fusion mechanism with a flow-matching action head to adaptively integrate visual-language cues with proprioceptive dynamics for action generation. Experiments on real-world humanoid manipulation tasks show that HEX achieves state-of-the-art performance in task success rate and generalization, particularly in fast-reaction and long-horizon scenarios.

Fanqi Shen, Enhong Yang, Jiahe Li et al.

Brain Foundation Models (BFMs) are transforming neuroscience by enabling scalable and transferable learning from neural signals, advancing both clinical diagnostics and cutting-edge neuroscience exploration. Their emergence is powered by large-scale clinical recordings, particularly electroencephalography (EEG) and intracranial EEG, which provide rich temporal and spatial representations of brain dynamics. However, despite their rapid proliferation, the field lacks a unified understanding of existing methodologies and a standardized evaluation framework. To fill this gap, we map the benchmark design space along two axes: (i) from the model perspective, we organize BFMs under a self-supervised learning (SSL) taxonomy; and (ii) from the dataset perspective, we summarize common downstream tasks and curate representative public datasets across clinical and human-centric neurotechnology applications. Building on this consolidation, we introduce Brain4FMs, an open evaluation platform with plug-and-play interfaces that integrates 15 representative BFMs and 18 public datasets. It enables standardized comparisons and analysis of how pretraining data, SSL strategies, and architectures affect generalization and downstream performance, guiding more accurate and transferable BFMs. The code is available at https://anonymous.4open.science/r/Brain4FMs-85B8.

Ji-Xuan He, Jia-Cheng Zhao, Feng-Qi Cui et al.

Low-light image super-resolution (LLISR) is essential for restoring fine visual details and perceptual quality under insufficient illumination conditions with ubiquitous low-resolution devices. Although pioneer methods achieve high performance on single tasks, they solve both tasks in a serial manner, which inevitably leads to artifact amplification, texture suppression, and structural degradation. To address this, we propose Decoupling then Perceive (DTP), a novel frequency-aware framework that explicitly separates luminance and texture into semantically independent components, enabling specialized modeling and coherent reconstruction. Specifically, to adaptively separate the input into low-frequency luminance and high-frequency texture subspaces, we propose a Frequency-aware Structural Decoupling (FSD) mechanism, which lays a solid foundation for targeted representation learning and reconstruction. Based on the decoupled representation, a Semantics-specific Dual-path Representation (SDR) learning strategy that performs targeted enhancement and reconstruction for each frequency component is further designed, facilitating robust luminance adjustment and fine-grained texture recovery. To promote structural consistency and perceptual alignment in the reconstructed output, building upon this dual-path modeling, we further introduce a Cross-frequency Semantic Recomposition (CSR) module that selectively integrates the decoupled representations. Extensive experiments on the most widely used LLISR benchmarks demonstrate the superiority of our DTP framework, improving $+$1.6\% PSNR, $+$9.6\% SSIM, and $-$48\% LPIPS compared to the most state-of-the-art (SOTA) algorithm. Codes are released at https://github.com/JXVision/DTP.

Hui Cheng, Jinsheng Guo, Zhenhao Weng et al.

Recent research in time series forecasting frequently investigates the integration of textual and visual modalities with numerical models to better navigate non-stationary environments. Despite delivering solid numerical results, existing multi-modal approaches usually encounter a dilemma: prioritizing the minimization of average errors can result in excessively smooth forecasts that overlook essential fluctuations. To resolve this limitation, we introduce STaT, an innovative multimodal architecture for Symbolic-Temporal-Textual Alignment, which seamlessly unites three synergistic modalities. Specifically, the symbolic modality converts continuous time series into discrete tokens, facilitating the accurate identification of structural patterns and turning points; the temporal modality extracts inherent sequential dependencies; and the textual modality leverages domain semantics to steer the macroscopic forecasting trends. Comprehensive evaluations on eight real-world benchmarks indicate that STaT delivers exceptional performance, enhancing conventional magnitude indicators by up to 8.9% while simultaneously decreasing shape distortion by up to 8.5%.

Jinsheng Guo, Zhenhao Weng, Yibo Liu et al.

Graph fraud detection has long depended on Graph Neural Networks (GNNs) to propagate and aggregate information across relational data. A critical obstacle in practice, however, is that fraudsters frequently disguise themselves by forging numerous connections with benign users, causing fraud signals to be progressively diluted during neighborhood aggregation and undermining detection reliability. While recent efforts have used Large Language Models (LLMs) to provide rich semantic cues for fraud detection, the underlying intent behind suspicious connections remains insufficiently explored. Compounding this issue, the scarcity of annotated fraud samples makes it difficult to train detectors that remain robust under heavy camouflage. To address these gaps, we propose L2IR, an LLM-driven Latent Intent Revealing framework for graph fraud detection. By uncovering latent intent from both user behaviors and suspicious connections, L2IR extracts intent-aware representations from raw behavioral traces and reasons about the true purpose behind individual connections, effectively distinguishing supportive links from misleading ones. It further incorporates adaptive self-training to enhance robustness under limited supervision. Evaluations on two real-world datasets characterized by pervasive camouflage demonstrate that L2IR surpasses strong baselines and can function as a plug-in enhancement for a range of GNN-based detectors, improving AUPRC by up to 8.27%.

Yingkai Fu, Meng Li, Wenxi Liu et al.

Event cameras, or dynamic vision sensors, have recently achieved success from fundamental vision tasks to high-level vision researches. Due to its ability to asynchronously capture light intensity changes, event camera has an inherent advantage to capture moving objects in challenging scenarios including objects under low light, high dynamic range, or fast moving objects. Thus event camera are natural for visual object tracking. However, the current event-based trackers derived from RGB trackers simply modify the input images to event frames and still follow conventional tracking pipeline that mainly focus on object texture for target distinction. As a result, the trackers may not be robust dealing with challenging scenarios such as moving cameras and cluttered foreground. In this paper, we propose a distractor-aware event-based tracker that introduces transformer modules into Siamese network architecture (named DANet). Specifically, our model is mainly composed of a motion-aware network and a target-aware network, which simultaneously exploits both motion cues and object contours from event data, so as to discover motion objects and identify the target object by removing dynamic distractors. Our DANet can be trained in an end-to-end manner without any post-processing and can run at over 80 FPS on a single V100. We conduct comprehensive experiments on two large event tracking datasets to validate the proposed model. We demonstrate that our tracker has superior performance against the state-of-the-art trackers in terms of both accuracy and efficiency.

Yixue Zhang, Kun Wu, Zhi Gao et al.

The pursuit of general-purpose robotic manipulation is hindered by the scarcity of diverse, real-world interaction data. Unlike data collection from web in vision or language, robotic data collection is an active process incurring prohibitive physical costs. Consequently, automated task curation to maximize data value remains a critical yet under-explored challenge. Existing manual methods are unscalable and biased toward common tasks, while off-the-shelf foundation models often hallucinate physically infeasible instructions. To address this, we introduce RoboGene, an agentic framework designed to automate the generation of diverse, physically plausible manipulation tasks across single-arm, dual-arm, and mobile robots. RoboGene integrates three core components: diversity-driven sampling for broad task coverage, self-reflection mechanisms to enforce physical constraints, and human-in-the-loop refinement for continuous improvement. We conduct extensive quantitative analysis and large-scale real-world experiments, collecting datasets of 18k trajectories and introducing novel metrics to assess task quality, feasibility, and diversity. Results demonstrate that RoboGene significantly outperforms state-of-the-art foundation models (e.g., GPT-4o, Gemini 2.5 Pro). Furthermore, real-world experiments show that VLA models pre-trained with RoboGene achieve higher success rates and superior generalization, underscoring the importance of high-quality task generation. Our project is available at https://robogene-boost-vla.github.io.

Haojun Jiang, Teng Wang, Zhenguo Sun et al.

Echocardiography is an essential medical technique for diagnosing cardiovascular diseases, but its high operational complexity has led to a shortage of trained professionals. To address this issue, we introduce a novel probe movement guidance algorithm that has the potential to be applied in guiding robotic systems or novices with probe pose adjustment for high-quality standard plane image acquisition.Cardiac ultrasound faces two major challenges: (1) the inherently complex structure of the heart, and (2) significant individual variations. Previous works have only learned the population-averaged structure of the heart rather than personalized cardiac structures, leading to a performance bottleneck. Clinically, we observe that sonographers dynamically adjust their interpretation of a patient's cardiac anatomy based on prior scanning sequences, consequently refining their scanning strategies. Inspired by this, we propose a novel sequence-aware self-supervised pre-training method. Specifically, our approach learns personalized three-dimensional cardiac structural features by predicting the masked-out image features and probe movement actions in a scanning sequence. We hypothesize that if the model can predict the missing content it has acquired a good understanding of personalized cardiac structure. Extensive experiments on a large-scale expert scanning dataset with 1.67 million samples demonstrate that our proposed sequence-aware paradigm can effectively reduce probe guidance errors compared to other advanced baseline methods.

Chenchu Xu, Meng Li, Xue Wu

The accurate segmentation of Coronary Computed Tomography Angiography (CCTA) images holds substantial clinical value for the early detection and treatment of Coronary Heart Disease (CHD). The Transformer, utilizing a self-attention mechanism, has demonstrated commendable performance in the realm of medical image processing. However, challenges persist in coronary segmentation tasks due to (1) the damage to target local structures caused by fixed-size image patch embedding, and (2) the critical role of both global and local features in medical image segmentation tasks.To address these challenges, we propose a deep learning framework, TransCC, that effectively amalgamates the Transformer and convolutional neural networks for CCTA segmentation. Firstly, we introduce a Feature Interaction Extraction (FIE) module designed to capture the characteristics of image patches, thereby circumventing the loss of semantic information inherent in the original method. Secondly, we devise a Multilayer Enhanced Perceptron (MEP) to augment attention to local information within spatial dimensions, serving as a complement to the self-attention mechanism. Experimental results indicate that TransCC outperforms existing methods in segmentation performance, boasting an average Dice coefficient of 0.730 and an average Intersection over Union (IoU) of 0.582. These results underscore the effectiveness of TransCC in CCTA image segmentation.

Fangyuan Liu, Sirui Zhao, Zeyu Zhang et al.

Automated depression estimation is highly vulnerable to signal corruption and ambient noise in real-world deployment. Prevailing deterministic methods produce uncalibrated point estimates, exposing safety-critical clinical systems to the severe risk of overconfident misdiagnoses. To establish a highly resilient and trustworthy assessment paradigm, we propose EviDep, an evidential learning framework that jointly quantifies depression severity alongside aleatoric and epistemic uncertainties via a Normal-Inverse-Gamma distribution. A fundamental vulnerability in multimodal evidential fusion is the uncontrolled accumulation of cross-modal redundancies. This structural flaw artificially inflates diagnostic confidence by double-counting overlapping evidence. To guarantee robust evidence synthesis, EviDep enforces strict information integrity. First, a Frequency-aware Feature Extraction module leverages a wavelet-based Mixture-of-Experts to dynamically isolate task-irrelevant noise, preserving the fidelity of diagnostic signals. Subsequently, a Disentangled Evidential Learning strategy separates the shared consensus from modality-specific nuances. By explicitly decoupling these representations before Bayesian fusion, EviDep systematically mitigates evidence redundancy. Extensive experiments on AVEC 2013, 2014, DAIC-WOZ, and E-DAIC confirm that EviDep achieves state-of-the-art predictive accuracy and superior uncertainty calibration, delivering a robust fail-safe mechanism for trustworthy clinical screening.

Weikai Xu, Meng Li, Shuzhang Zhong et al.

The Mixture-of-Experts (MoE) models have emerged as the state-of-the-art paradigm for scaling up large language models (LLMs) without proportionally increased computational cost. However, its on-device deployment faces a critical challenge due to the large memory requirement for storing all expert parameters. 3D NAND-based computing-in-memory (CIM) architectures uniquely offer high storage capacity and reduced data movement, while they are ill-suited for MoE models with dynamically sparse expert activation, leading to a degradation of effective computational parallelism, along with underutilization of multibit storage capability of Flash cells. In this work, we proposed a 3D NAND-based content addressable-selected CIM architecture, dubbed as NASiC, which is tailored to MoE models. By leveraging the intrinsic string structure of 3D NAND technology, NASiC fuses the dynamical expert selection through CAM-based masking mechanism and activated expert computation through CIM into a single computation cycle, eradicating redundant computation and enhancing computational parallelism. Moreover, circuit-level optimizations and multibit CIM cell are co-designed with proposed NASiC architecture, featuring block-wise parallel computation with in-situ signed multibit input and weight expansion, substantially improving the throughput and energy-efficiency of NAND CIM array, as well as the utilization of high-density 3D NAND technology for MoE models. With extensive experimental results, we demonstrate NASiC achieves 4-114.8x improved performance and 3.9-70x improved energy efficiency over state-of-the-art designs, along with high accuracy, showing its great potential for efficient on-device MoE LLM inference.

Caleb Skinner, Yihan Guo, Meng Li

Large language models (LLMs) offer a scalable mechanism to elicit domain-informed prior information for high-dimensional variable selection. However, existing methods such as LLM-Lasso are sensitive to weight quality, with performance degrading substantially when LLM-generated weights are inaccurate. To address this challenge, we first introduce a framework for quantifying the quality of LLM-generated weights, enabling rigorous evaluation of LLM-informed methods across varying weight regimes. We then propose the LLM Sparsity Prior (LSP), which integrates LLM-generated weights into the prior inclusion probabilities of Spike-and-Slab and Spike-and-Slab Lasso models via two interpretable hyperparameters governing global sparsity and weight concentration. Hierarchical hyperpriors on these parameters allow the model to dynamically discount uninformative or misleading weights, improving robustness without sacrificing gains when weights are accurate. Finally, we develop principled prompt engineering strategies and validate the method on a private medical dataset studying Acute Kidney Injury. LSP improves prediction accuracy and identifies clinically relevant features missed by the baselines, with robustness to prompt variation and particular effectiveness in low-data regimes.

DeepSeek-AI, Aixin Liu, Aoxue Mei et al.

We introduce DeepSeek-V3.2, a model that harmonizes high computational efficiency with superior reasoning and agent performance. The key technical breakthroughs of DeepSeek-V3.2 are as follows: (1) DeepSeek Sparse Attention (DSA): We introduce DSA, an efficient attention mechanism that substantially reduces computational complexity while preserving model performance in long-context scenarios. (2) Scalable Reinforcement Learning Framework: By implementing a robust reinforcement learning protocol and scaling post-training compute, DeepSeek-V3.2 performs comparably to GPT-5. Notably, our high-compute variant, DeepSeek-V3.2-Speciale, surpasses GPT-5 and exhibits reasoning proficiency on par with Gemini-3.0-Pro, achieving gold-medal performance in both the 2025 International Mathematical Olympiad (IMO) and the International Olympiad in Informatics (IOI). (3) Large-Scale Agentic Task Synthesis Pipeline: To integrate reasoning into tool-use scenarios, we developed a novel synthesis pipeline that systematically generates training data at scale. This methodology facilitates scalable agentic post-training, yielding substantial improvements in generalization and instruction-following robustness within complex, interactive environments.

Yushuai Song, Weize Quan, Weining Wang et al.

Recent advances in generative super-resolution (SR) have greatly improved visual realism, yet existing evaluation and optimization frameworks remain misaligned with human perception. Full-Reference and No-Reference metrics often fail to reflect perceptual preference, either penalizing semantically plausible details due to pixel misalignment or favoring visually sharp but inconsistent artifacts. Moreover, most SR methods rely on ground-truth (GT)-dependent distribution matching, which does not necessarily correspond to human judgments. In this work, we propose RefReward-SR, a low-resolution (LR) reference-aware reward model for preference-aligned SR. Instead of relying on GT supervision or NR evaluation, RefReward-SR assesses high-resolution (HR) reconstructions conditioned on their LR inputs, treating the LR image as a semantic anchor. Leveraging the visual-linguistic priors of a Multimodal Large Language Models (MLLM), it evaluates semantic consistency and plausibility in a reasoning-aware manner. To support this paradigm, we construct RefSR-18K, the first large-scale LR-conditioned preference dataset for SR, providing pairwise rankings based on LR-HR consistency and HR naturalness. We fine-tune the MLLM with Group Relative Policy Optimization (GRPO) using LR-conditioned ranking rewards, and further integrate GRPO into SR model training with RefReward-SR as the core reward signal for preference-aligned generation. Extensive experiments show that our framework achieves substantially better alignment with human judgments, producing reconstructions that preserve semantic consistency while enhancing perceptual plausibility and visual naturalness. Code, models, and datasets will be released upon paper acceptance.

Ziwen Liu, Meng Li

Discrete probability laws underpin statistical modeling, yet the catalog of interpretable distributions has expanded only gradually through centuries of case-by-case mathematical derivations. We introduce symbolic density estimation (SDE), an unsupervised framework that automatically recovers closed-form probability mass functions by composing elementary analytic operations within a structured search space. Our method integrates domain-specific structural priors with evolutionary search and a validity-aware inference stage, and it extends to richer distribution families such as zero inflation and finite mixtures. To support systematic evaluation and future research, we contribute a benchmark dataset spanning a broad collection of commonly used discrete distributions. The proposed algorithm recovers all benchmark families with accurate parameter estimates. A real data application shows that it identifies concise and interpretable mixture models that improve goodness-of-fit over standard models.

Meng Li, Peisong Wang, Yuantian Shao et al.

Large Language Models (LLMs) achieve strong performance across diverse tasks but face deployment challenges due to their massive size. Structured pruning offers acceleration benefits but leads to significant performance degradation. Recent PCA-based pruning methods have alleviated this issue by retaining key activation components, but are only applied between modules in order to fuse the transformation matrix, which introduces extra parameters and severely disrupts activation distributions due to residual connections. To address these issues, we propose IntraSlice, a framework that applies block-wise module-intra PCA compression pruning. By leveraging the structural characteristics of Transformer modules, we design an approximate PCA method whose transformation matrices can be fully fused into the model without additional parameters. We also introduce a PCA-based global pruning ratio estimator that further considers the distribution of compressed activations, building on conventional module importance. We validate our method on Llama2, Llama3, and Phi series across various language benchmarks. Experimental results demonstrate that our approach achieves superior compression performance compared to recent baselines at the same compression ratio or inference speed.

Xiangpeng Tian, Xiangyu Liao, Xiao Liu et al.

All-in-one image restoration aims to recover clear images from various degradation types and levels with a unified model. Nonetheless, the significant variations among degradation types present challenges for training a universal model, often resulting in task interference, where the gradient update directions of different tasks may diverge due to shared parameters. To address this issue, motivated by the routing strategy, we propose DFPIR, a novel all-in-one image restorer that introduces Degradation-aware Feature Perturbations(DFP) to adjust the feature space to align with the unified parameter space. In this paper, the feature perturbations primarily include channel-wise perturbations and attention-wise perturbations. Specifically, channel-wise perturbations are implemented by shuffling the channels in high-dimensional space guided by degradation types, while attention-wise perturbations are achieved through selective masking in the attention space. To achieve these goals, we propose a Degradation-Guided Perturbation Block (DGPB) to implement these two functions, positioned between the encoding and decoding stages of the encoder-decoder architecture. Extensive experimental results demonstrate that DFPIR achieves state-of-the-art performance on several all-in-one image restoration tasks including image denoising, image dehazing, image deraining, motion deblurring, and low-light image enhancement. Our codes are available at https://github.com/TxpHome/DFPIR.

Qingyu Guo, Jiayong Wan, Songqiang Xu et al.

Vision Transformer (ViT) acceleration with field programmable gate array (FPGA) is promising but challenging. Existing FPGA-based ViT accelerators mainly rely on temporal architectures, which process different operators by reusing the same hardware blocks and suffer from extensive memory access overhead. Pipelined architectures, either coarse-grained or fine-grained, unroll the ViT computation spatially for memory access efficiency. However, they usually suffer from significant hardware resource constraints and pipeline bubbles induced by the global computation dependency of ViT. In this paper, we introduce HG-PIPE, a pipelined FPGA accelerator for high-throughput and low-latency ViT processing. HG-PIPE features a hybrid-grained pipeline architecture to reduce on-chip buffer cost and couples the computation dataflow and parallelism design to eliminate the pipeline bubbles. HG-PIPE further introduces careful approximations to implement both linear and non-linear operators with abundant Lookup Tables (LUTs), thus alleviating resource constraints. On a ZCU102 FPGA, HG-PIPE achieves 2.78 times better throughput and 2.52 times better resource efficiency than the prior-art accelerators, e.g., AutoViTAcc. With a VCK190 FPGA, HG-PIPE realizes end-to-end ViT acceleration on a single device and achieves 7118 images/s, which is 2.81 times faster than a V100 GPU.

Tong Xie, Zuodong Zhang, Chao Yang et al.

Deep neural networks (DNNs) have showcased remarkable performance across various tasks and are widely deployed on AI accelerators fabricated in advanced technology nodes for efficiency. As aging effects become more pronounced, timing and voltage guardbands are increasingly applied. Aging-aware adaptive voltage scaling (AVS), which adjusts supply voltage based on on-chip aging scenarios, has emerged as a promising solution to avoid excessive guardbanding. However, conventional AVS techniques overlook the inherent resilience of DNNs and frequently raise the supply voltage unnecessarily, thereby exacerbating aging and increasing power consumption. To enable reliable and efficient AI inference with AVS, in this paper, we develop an accurate aging prediction framework that incorporates historical effects and iterative extrapolation for full-lifetime modeling. Building on this framework, we propose a fault-tolerant voltage scaling policy that exploits DNN resilience and defers voltage increases accordingly. Experiments show that our framework mitigates the pessimism of maximum-voltage baselines, reducing predicted threshold voltage shift (ΔVth) by 19.4% for PMOS and 19.1% for NMOS, respectively. Furthermore, evaluation on representative DNN workloads demonstrates that our optimization reduces aging degradation by up to 45.8% (NMOS) and 30.6% (PMOS) while achieving 14.0% average lifetime power savings compared to resilience-agnostic methods.

Renjie Wei, Songqiang Xu, Linfeng Zhong et al.

State space models (SSMs) like Mamba have recently attracted much attention. Compared to Transformer-based large language models (LLMs), Mamba achieves linear computation complexity with the sequence length and demonstrates superior performance. However, Mamba is hard to accelerate due to the scattered activation outliers and the complex computation dependency, rendering existing LLM accelerators inefficient. In this paper, we propose LightMamba that co-designs the quantization algorithm and FPGA accelerator architecture for efficient Mamba inference. We first propose an FPGA-friendly post-training quantization algorithm that features rotation-assisted quantization and power-of-two SSM quantization to reduce the majority of computation to 4-bit. We further design an FPGA accelerator that partially unrolls the Mamba computation to balance the efficiency and hardware costs. Through computation reordering as well as fine-grained tiling and fusion, the hardware utilization and memory efficiency of the accelerator get drastically improved. We implement LightMamba on Xilinx Versal VCK190 FPGA and achieve 4.65x to 6.06x higher energy efficiency over the GPU baseline. When evaluated on Alveo U280 FPGA, LightMamba reaches 93 tokens/s, which is 1.43x that of the GPU baseline. Our code is available at https://github.com/PKU-SEC-Lab/LightMamba.

Jinze Zhang, Jian Zhong, Li Lin et al.

Optical coherence tomography (OCT) has revolutionized retinal disease diagnosis with its high-resolution and three-dimensional imaging nature, yet its full diagnostic automation in clinical practices remains constrained by multi-stage workflows and conventional single-slice single-task AI models. We present Full-process OCT-based Clinical Utility System (FOCUS), a foundation model-driven framework enabling end-to-end automation of 3D OCT retinal disease diagnosis. FOCUS sequentially performs image quality assessment with EfficientNetV2-S, followed by abnormality detection and multi-disease classification using a fine-tuned Vision Foundation Model. Crucially, FOCUS leverages a unified adaptive aggregation method to intelligently integrate 2D slices-level predictions into comprehensive 3D patient-level diagnosis. Trained and tested on 3,300 patients (40,672 slices), and externally validated on 1,345 patients (18,498 slices) across four different-tier centers and diverse OCT devices, FOCUS achieved high F1 scores for quality assessment (99.01%), abnormally detection (97.46%), and patient-level diagnosis (94.39%). Real-world validation across centers also showed stable performance (F1: 90.22%-95.24%). In human-machine comparisons, FOCUS matched expert performance in abnormality detection (F1: 95.47% vs 90.91%) and multi-disease diagnosis (F1: 93.49% vs 91.35%), while demonstrating better efficiency. FOCUS automates the image-to-diagnosis pipeline, representing a critical advance towards unmanned ophthalmology with a validated blueprint for autonomous screening to enhance population scale retinal care accessibility and efficiency.

Linye Wei, Zixiang Luo, Pingzhi Tang et al.

Diffusion large language models (dLLMs) have recently gained significant attention due to their inherent support for parallel decoding. Building on this paradigm, Mixture-of-Experts (MoE) dLLMs with autoregressive (AR) initialization have further demonstrated strong performance competitive with mainstream AR models. However, we identify a fundamental mismatch between MoE architectures and diffusion-based decoding. Specifically, a large number of experts are activated at each denoising step, while only a small subset of tokens is ultimately accepted, resulting in substantial inference overhead and limiting their deployment in latency-sensitive applications. In this work, we propose TEAM, a plug-and-play framework that accelerates MoE dLLMs by enabling more accepted tokens with fewer activated experts. TEAM is motivated by the observation that expert routing decisions exhibit strong temporal consistency across denoising levels as well as spatial consistency across token positions. Leveraging these properties, TEAM employs three complementary expert activation and decoding strategies, conservatively selecting necessary experts for decoded and masked tokens and simultaneously performing aggressive speculative exploration across multiple candidates. Experimental results demonstrate that TEAM achieves up to 2.2x speedup over vanilla MoE dLLM, with negligible performance degradation. Code is released at https://github.com/PKU-SEC-Lab/TEAM-MoE-dLLM.

Shuzhang Zhong, Yanfan Sun, Ling Liang et al.

The Mixture of Experts (MoE) architecture has demonstrated significant advantages as it enables to increase the model capacity without a proportional increase in computation. However, the large MoE model size still introduces substantial memory demands, which usually requires expert offloading on resource-constrained platforms and incurs significant overhead. Hybrid CPU-GPU inference has been proposed to leverage CPU computation to reduce expert loading overhead but faces major challenges: on one hand, the expert activation patterns of MoE models are highly unstable, rendering the fixed mapping strategies in existing works inefficient; on the other hand, the hybrid CPU-GPU schedule for MoE is inherently complex due to the diverse expert sizes, structures, uneven workload distribution, etc. To address these challenges, in this paper, we propose HybriMoE, a hybrid CPU-GPU inference framework that improves resource utilization through a novel CPU-GPU scheduling and cache management system. HybriMoE introduces (i) a dynamic intra-layer scheduling strategy to balance workloads across CPU and GPU, (ii) an impact-driven inter-layer prefetching algorithm, and (iii) a score-based caching algorithm to mitigate expert activation instability. We implement HybriMoE on top of the kTransformers framework and evaluate it on three widely used MoE-based LLMs. Experimental results demonstrate that HybriMoE achieves an average speedup of 1.33$\times$ in the prefill stage and 1.70$\times$ in the decode stage compared to state-of-the-art hybrid MoE inference framework. Our code is available at: https://github.com/PKU-SEC-Lab/HybriMoE.

Meng Li, Xiaohua Yang, Jie Liu et al.

Context. Metamorphic Testing addresses the test-oracle problem in scientific computing, but classical Mutation Score operates on syntactic AST mutations and misses domain semantics. Objective. We propose the Semantic Mutation Score (SMS), built on five domain-semantic operators (Conservation Erosion, Operator Substitution, Hyperparameter, Trajectory Flip, Structural Injection). SMS degenerates almost everywhere to MS in a characterised limit, so any SMS-based conclusion remains consistent with prior mutation-testing literature in the classical regime. Method. A 12-PUT x 5-MP design over four single-output float-to-float classes (numeric, probabilistic, surrogate, machine-learning) is paired with a three-layer attribution classifier separating true semantic faults from tolerance, OOD, statistical, and artefact categories. A same-source / cross-source ablation under an identical prompt isolates the LLM-source-diversity contribution. LLM-generated mutants are compared against a default-configuration cosmic-ray syntactic pool at the AST-normalised level. Results. The pre-registered large-effect threshold for Cliff's delta is not met under the point-estimate criterion; the observed effect lies in the medium-effect range. Cross-source pooling under an identical prompt does not appreciably shift delta, indicating that LLM identity is not the lever within this design. AST-level overlap between LLM-generated and default cosmic-ray syntactic mutants is small; the Hyperparameter, Structural Injection, and Trajectory Flip classes are unreachable under default first-order syntactic configurations. Conclusion. SMS is a backward-compatible adequacy metric for domain-semantic metamorphic-relation sets in scientific computing. The first-order unreachability evidence is independent of the effect-size question.