Test-Time Adaptation (TTA) has emerged as an effective solution for adapting Vision Transformers (ViT) to distribution shifts without additional training data. However, existing TTA methods often incur substantial computational overhead, limiting their applicability in resource-constrained real-world scenarios. To reduce inference cost, plug-and-play token aggregation methods merge redundant tokens in ViTs to reduce total processed tokens. Albeit efficient, it suffers from significant performance degradation when directly integrated with existing TTA methods. We formalize this problem as Efficient Test-Time Adaptation (ETTA), seeking to preserve the adaptation capability of TTA while reducing inference latency. In this paper, we first provide a theoretical analysis from a novel mutual information perspective, showing that token aggregation inherently leads to information loss, which cannot be fully mitigated by conventional norm-tuning-based TTA methods. Guided by this insight, we propose to Neutralize Token Aggregation via Information Augmentation (NAVIA). Specifically, we directly augment the [CLS] token embedding and incorporate adaptive biases into the [CLS] token in shallow layers of ViTs. We theoretically demonstrate that these augmentations, when optimized via entropy minimization, recover the information lost due to token aggregation. Extensive experiments across various out-of-distribution benchmarks demonstrate that NAVIA significantly outperforms state-of-the-art methods by over 2.5%, while achieving an inference latency reduction of more than 20%, effectively addressing the ETTA challenge.

Test-time adaptation (TTA) has been widely explored in single-label recognition, effectively mitigating distribution shifts, especially when combined with vision-language models. However, real-world images often contain multiple objects, while the more practical multi-label test-time adaptation (MLTTA) has received little attention so far. Recent cache-based TTA methods have shown promising efficiency and effectiveness, yet directly extending them to multi-label scenarios suffers from a one-to-many mapping problem: a shared global representation entangling co-occurring objects is stored as class-wise cache prototypes, inducing dominant-label bias and compromised cache calibration. While introducing region-level cues helps isolate class-specific evidence, such regional evidence can also be unreliable under distribution shifts, making its identification and utilization non-trivial. To address these issues, we introduce PuRF, a novel PuRiFication-driven cache-based method for multi-label test-time adaptation of vision-language models. Specifically, PuRF first performs region purification to identify reliable regions, providing comprehensive regional cues for multi-label recognition and enabling fine-grained alignment. Based on these purified regions, PuRF conducts cache purification to enhance cache representation and adaptability, where episodic purification builds a discriminative region-based cache, and temporal refreshing further promotes long-term cache adaptability. Experiments demonstrate that PuRF consistently outperforms state-of-the-art methods, achieving a notable 4.05

In recent years， vision-language-action （VLA） models have attracted significant attention in the field of embodied intelligence. As model scale continues to grow， their ability to generalize across complex tasks has steadily improved. However， such performance improvements rely heavily on the availability of large-scale， high-quality training data. Unlike natural language processing and computer vision， which can directly leverage massive internet data， data collection in embodied intelligence typically involves physical interactions between real robots and their environments， leading to high collection costs and complex acquisition processes. Efficiently obtaining， processing， and organizing such data has therefore become a critical challenge for advancing embodied intelligence. To address this issue， this paper provides a systematic review of data collection and processing methods in embodied intelligence. First， we summarize the major data acquisition paradigms from the perspective of data sources and collection strategies， and analyze their characteristics and limitations in terms of data quality， scalability， and collection cost. Second， we present a standardized processing pipeline for embodied intelligence datasets， focusing on key technical components such as action representation alignment， multimodal temporal synchronization， language semantic normalization， and data quality control. Finally， we discuss the evolving data ecosystem in embodied intelligence， highlighting current challenges and potential future directions. The analysis presented in this paper aims to provide insights for dataset construction and large-scale robot learning research in embodied intelligence.

Anomaly Detection (AD) technology has received much attention recently, especially in industrial quality inspection applications. Most existing unsupervised AD methods assume that the training data contains only normal samples, which is difficult to satisfy in practice. When the training data are mixed with even a small number of defective samples, the AD methods, that use distillation learning, will be negatively affected, leading to significant performance drops. To tackle this issue, in this paper, we proposed an approach, namely AD^2 , to conduct anomaly detection during the training phase of an anomaly detection model. Specifically, we devise a Non-Major Feature Elimination (NMFE) module to eliminate the prominent anomaly-related discrepancy information and adopt an Anomaly Training Data Removal (ATDR) strategy to identify outliers in the training data, preventing abnormal information from affecting model training. During the inference phase, AD^2 does not introduce any extra computation overhead. Experiments demonstrate that AD^2 can successfully alleviate the performance deterioration caused by polluted training samples. On the MVTec LOCO dataset, when 10% of the training set is corrupted by anomalous samples, AD^2 can significantly improve the image-level AUROC from 0.793 to 0.865 compared to the ordinary AD method, without sacrificing any inference efficiency. AD^2 provides an effective solution for issues of data uncertainty in anomaly detection. The source code will be released.

The accurate classification of chondroblastoma and giant cell tumor of bone has been a major concern in the diagnosis of clinical pathology. The diagnosis of both neoplasms has traditionally relied on the experience and expertise of pathologists. However, a notable research gap persists in the field of computer-aided diagnosis of pathological slide images, particularly in the area of automated diagnosis and differentiation of chondroblastoma and giant cell tumor of bone. We propose an innovative, parallel, and efficient Paralleformer network to enhance the performance of the model for bone lesion images classification. Features are extracted from both the central and surrounding patches at the same time to better capture image details. The Focus Linear Attention module is employed to identify key information and discard irrelevant details, which helps to improve model accuracy. Additionally, the channel-wise module and the residual structure of the encoder are introduced to adequately capture the features. Consequently, our model attains an accuracy of 91.05 %, effectively addressing the limitations of traditional Transformer models in terms of computational complexity and accuracy. Tested by self-built and public datasets, the accuracy of our method is improved by over 4 % compared with traditional convolutional neural networks and Transformer models.

Vision Transformers (ViTs) have emerged as state-of-the-art models for various vision tasks recently. However, their heavy computation costs remain daunting for resource-limited devices. To address this, researchers have dedicated themselves to compressing redundant information in ViTs for acceleration. However, existing approaches generally sparsely drop redundant image tokens by token pruning or brutally remove channels by channel pruning, leading to a sub-optimal balance between model performance and inference speed. Moreover, they struggle when transferring compressed models to downstream vision tasks that require the spatial structure of images, such as semantic segmentation. To tackle these issues, we propose CAIT, a joint compression method for ViTs that achieves a harmonious blend of high accuracy, fast inference speed, and favorable transferability to downstream tasks. Specifically, we introduce an asymmetric token merging (ATME) strategy to effectively integrate neighboring tokens. It can successfully compress redundant token information while preserving the spatial structure of images. On top of it, we further design a consistent dynamic channel pruning (CDCP) strategy to dynamically prune unimportant channels in ViTs. Thanks to CDCP, insignificant channels in multi-head self-attention modules of ViTs can be pruned uniformly, significantly enhancing the model compression. Extensive experiments on multiple benchmark datasets show that our proposed method can achieve state-of-the-art performance across various ViTs.

Recent advancements in autonomous driving, augmented reality, robotics, and embodied intelligence have necessitated 3D perception algorithms. However, current 3D perception methods, especially specialized small models, exhibit poor generalization in open scenarios. On the other hand, multimodal large language models (MLLMs) excel in general capacity but underperform in 3D tasks, due to weak 3D local spatial object perception, poor text-based geometric numerical output, and inability to handle camera focal variations. To address these challenges, we propose the following solutions: Spatial-Enhanced Local Feature Mining for better spatial feature extraction, 3D Query Token-Derived Info Decoding for precise geometric regression, and Geometry Projection-Based 3D Reasoning for handling camera focal length variations. We employ parameter-efficient fine-tuning for a pre-trained MLLM and develop LLMI3D, a powerful 3D perception MLLM. Additionally, we have constructed the IG3D dataset, which provides fine-grained descriptions and question-answer annotations. Extensive experiments demonstrate that our LLMI3D achieves state-of-the-art performance, outperforming other methods by a large margin.

Vision-Language Models (VLMs) have shown strong promise on Optical Character Recognition (OCR), yet the sheer number of visual tokens required to encode dense documents incurs prohibitive inference cost. Existing pruning methods rely on physical eviction, e.g., permanently discarding visual tokens during the prefill stage. While effective for natural images, this strategy fundamentally breaks down on OCR, where virtually every visual token may correspond to a character or structural element, and any irreversible loss leads to catastrophic accuracy degradation. We observe that, although document images appear globally dense and seemingly unprunable, the model's attention to them is in fact temporally sparse: at each decoding step it concentrates on a small region that shifts gradually across steps, much as a human reader fixates on successive words rather than perceiving an entire page at once. Motivated by this Dynamic Visual Fixation phenomenon, we recast the intractable global pruning problem as a tractable local, dynamic one and propose FastOCR, a training-free framework with two complementary modules. Specifically, Focal-Guided Pruning identifies a small set of focal layers and selects the most task-relevant visual tokens from them at each step, while Cross-Step Fixation Reuse exploits the gradual shift of fixation to warm-start each step from the previous one. By dynamically adjusting which tokens are attended rather than evicting any from the cache, FastOCR avoids permanent information loss. Extensive experiments show that FastOCR serves as a plug-and-play acceleration module, generalizing consistently across five VLMs of varying sizes and architectures. On Qwen2.5-VL, FastOCR retains 98

Anomaly detection (AD) aims to distinguish anomalous samples from normal ones and localize the corresponding anomalous regions in real-world manufacturing environments. Benefiting from the strong visual priors of foundation models (FMs), few-shot FM-based methods can rapidly adapt to new manufacturing categories using only a small support set. However, existing methods still suffer from triple inconsistencies: inconsistency between synthetic and real anomalies before training, inconsistency in anomaly evidence across patch tokens from different FM layers during training, and inconsistency between image-level and pixel-level assessments during evaluation. Consequently, AD models may correctly distinguish anomalous samples at the image level while still mislocalizing the anomalous regions at the pixel level. Detection performance further degrades markedly when adapting to other manufacturing datasets under the sparse few-shot setting. To address these inconsistencies, we propose TCFMAD, a tri-consistency FM-based few-shot AD pipeline that enforces consistency in anomaly synthesis, cross-layer anomaly evidence, and finegrained label construction throughout training and evaluation. TCFMAD adapts the anomaly detector using a few normal support images together with an equally small class-specific anomalous-patch bank. Specifically, realistic anomaly synthesis (RAS) uses this anomalous-patch bank, matched to the 1/2/4-shot normal-support budget, to generate realistic synthetic anomalous samples and alleviate the inconsistency between synthetic and real anomalies. During training, cross-layer evidence agreement (CLEA) optimizes the anomaly predictor with a consistency regularization loss to learn stable anomaly evidence from multilayer patch tokens of DINOv3. CLEA strengthens the consistency of anomaly evidence extraction. For PKU-Market-PCB, weaklyconstrained bootstrapping (WCB) operates offline to refine weak bounding boxes into masks for anomalous-patch extraction and fixed fine-grained localization assessment; it is not part of the anomaly detector or its online inference path. TCFMAD achieves state-of-the-art performance on the human-mask-based public benchmarks and strong relative performance on PKU-Market-PCB under a shared WCB-refined pseudo-mask protocol. Code is available at: https://github.com/GingerCohle/TCFMAD.

With the rapid growth of large-scale AI, the demand for GPUs and other compute devices has surged, driving massive production of circuit-boards (e.g., PCBs). Consequently, accurate real-time defect detection has become more essential. Many recent methods adopt the vision foundation model DINO for defect detection due to its consistent feature representations across diverse environments and domains. However, image-level DINO features tend to emphasize semantics and global structure. In contrast, circuit-board defects are often visually similar across categories because they are small and texture-dominated. Meanwhile, many detection methods rely on global image-level feature aggregation. This weakens local instance-level defect cue modeling and degrades category discriminability. To enhance circuit-board defect detection, we propose DINO-PCB, which adopts a two-stage design comprising vision foundation model pretraining and dual-level knowledge distillation. In the pretraining stage, DINO is lightly fine-tuned together with a class projection head pretrained from scratch. Specifically, circuit-board images are fed into DINO to extract patch tokens. RoIAlign is then applied to obtain RoI features, which are passed through a class projection head to produce RoI embeddings. These RoI embeddings are optimized with an ArcFace loss to enhance category separability. In the distillation stage, the pretrained DINO and the class projection head are frozen as the teacher to supervise RT-DETR via dual-level knowledge distillation. The objective combines hierarchical image-level distillation between DINO patch tokens and RT-DETR encoder features, and contrastive instance-level distillation between RoI embeddings from the class projection head and RT-DETR decoder object queries. DINO-PCB enhances low-level texture and micro-structure modeling. Consequently, the RT-DETR detector achieves high accuracy across three circuit-board defect datasets without introducing additional inference latency.

Recently, Large language models (LLMs) have revolutionized Natural Language Processing (NLP). Pretrained LLMs, due to limited training context size, struggle with handling long token sequences, limiting their performance on various downstream tasks. Current solutions toward long context modeling often employ multi-stage continual pertaining, which progressively increases the effective context length through several continual pretraining stages. However, those approaches require extensive manual tuning and human expertise. In this paper, we introduce a novel single-stage continual pretraining method, Head-Adaptive Rotary Position Encoding (HARPE), to equip LLMs with long context modeling capabilities while simplifying the training process. Our HARPE leverages different Rotary Position Encoding (RoPE) base frequency values across different attention heads and directly trains LLMs on the target context length. Extensive experiments on 4 language modeling benchmarks, including the latest RULER benchmark, demonstrate that HARPE excels in understanding and integrating long-context tasks with single-stage training, matching and even outperforming existing multi-stage methods. Our results highlight that HARPE successfully breaks the stage barrier for training LLMs with long context modeling capabilities.

Large Language Models (LLMs) are increasingly deployed in both latency-sensitive online services and cost-sensitive offline workloads. Co-locating these workloads on shared serving instances can improve resource utilization, but directly applying this approach to Prefill/Decode (P/D) disaggregated systems introduces severe load imbalance, as fluctuating request mixes alter the intrinsic P/D ratio. Existing dynamic adjustment techniques cannot keep up with the bursty traffic patterns of online services. We propose a latency-constraint disaggregated architecture, which separates cluster resources into latency-strict and latency-relaxed pools based on task latency requirements. This design enables flexible placement of offline decode tasks, mitigating P/D imbalance while preserving online performance. To fully exploit this flexibility, we propose (1) a bottleneck-based scheduler guided by a Roofline-based performance model for performance bottleneck based scheduling, and (2) a fast preemption mechanism that strictly enforces Service Level Objectives (SLOs) for online requests. Experiments on real-world traces show that compared to existing offline system approaches, our method improves offline throughput by up to 3x, while maintaining online request SLOs.

We introduce xLLM, an intelligent and efficient Large Language Model (LLM) inference framework designed for high-performance, large-scale enterprise-grade serving, with deep optimizations for diverse AI accelerators. To address these challenges, xLLM builds a novel decoupled service-engine architecture. At the service layer, xLLM-Service features an intelligent scheduling module that efficiently processes multimodal requests and co-locates online and offline tasks through unified elastic scheduling to maximize cluster utilization. This module also relies on a workload-adaptive dynamic Prefill-Decode (PD) disaggregation policy and a novel Encode-Prefill-Decode (EPD) disaggregation policy designed for multimodal inputs. Furthermore, it incorporates a distributed architecture to provide global KV Cache management and robust fault-tolerant capabilities for high availability. At the engine layer, xLLM-Engine co-optimizes system and algorithm designs to fully saturate computing resources. This is achieved through comprehensive multi-layer execution pipeline optimizations, an adaptive graph mode and an xTensor memory management. xLLM-Engine also further integrates algorithmic enhancements such as optimized speculative decoding and dynamic EPLB, collectively serving to substantially boost throughput and inference efficiency. Extensive evaluations demonstrate that xLLM delivers significantly superior performance and resource efficiency. Under identical TPOT constraints, xLLM achieves throughput up to 1.7x that of MindIE and 2.2x that of vLLM-Ascend with Qwen-series models, while maintaining an average throughput of 1.7x that of MindIE with Deepseek-series models. xLLM framework is publicly available at https://github.com/jd-opensource/xllm and https://github.com/jd-opensource/xllm-service.

In order to boost the performance of a convolutional neural network (CNN), several approaches have shown the benefit of enhancing the spatial encoding of feature maps. However, few works paid attention to the positional properties of convolutional kernels. In this paper, we demonstrate that different kernel positions are of different importance, which depends on the task, dataset and architecture, and adaptively emphasizing the informative parts in convolutional kernels can lead to considerable improvement. Therefore, we propose a novel structural re-parameterization Position Boosting Convolution (PBConv) to exploit and enhance the position information in the convolutional kernel. PBConv consists of several concurrent small convolutional kernels, which can be equivalently converted to the original kernel and bring no extra inference cost. Different from existing structural re-parameterization methods, PB-conv searches for the optimal re-parameterized structure by a fast heuristic algorithm based on the dispersion of kernel weights. Such heuristic search is efficient yet effective, well adapting the varying kernel weight distribution. As a result, PBConv can significantly improve the representational power of a model, especially its ability to extract fine-grained low-level features. Importantly, PBConv is orthogonal to procedural re-parameterization methods and can further boost performance based on them.

Post-training is essential for adapting Large Language Models (LLMs) to real-world applications. Deploying post-trained models faces significant challenges due to substantial memory overhead and noticeable inference latency. Existing work has identified significant redundancies in LLMs and proposed efficient architectures, namely intra-layer KV sharing and cross-layer KV sharing. However, these methods still result in high inference time overhead, remaining suboptimal for post-training pre-trained LLMs. In this paper, we identify that the operation is a primary bottleneck for LLM inference and discover that it is actually highly redundant during post-training. We propose Softmax Unification in Attention (UniAttn), a novel post-training method that unifies Softmax activations across transformer blocks to reduce LLM inference costs. Additionally, UniAttn adopts a linear projection to compensate for the errors induced by Softmax unification. Experiments show that UniAttn matches the performance of standard post-training while significantly reducing inference costs, outperforming existing efficient architectures during post-training.

Video Large Language Models (Video LLMs) have achieved remarkable results in video understanding tasks. However, they often suffer from heavy computational overhead due to the large number of visual tokens generated from multiple video frames. Existing visual token compression methods often rely on attention scores from language models as guidance. However, these scores exhibit inherent biases: global bias reflects a tendency to focus on the two ends of the visual token sequence, while local bias leads to an over-concentration on the same spatial positions across different frames. To address the issue of attention bias, we propose Attention-Debiased Token Pruning for Video Large Language Models (AdaTP), a novel token pruning pipeline for Video LLMs. AdaTP integrates two dedicated debiasing modules into the pipeline, targeting global attention bias and local attention bias, respectively. Without the need for additional training, our method significantly reduces the computational overhead of Video LLMs while retaining the performance of vanilla models. Extensive evaluation shows that AdaTP achieves state-of-the-art performance in various commonly used video understanding benchmarks. In particular, on LLaVA-OneVision-7B, AdaTP maintains performance without degradation while using only up to 27.3% FLOPs compared to the vanilla model. Our code will be released soon.

Byte Pair Encoding (BPE) serves as a foundation method for text tokenization in the Natural Language Processing (NLP) field. Despite its wide adoption, the original BPE algorithm harbors an inherent flaw: it inadvertently introduces a frequency imbalance for tokens in the text corpus. Since BPE iteratively merges the most frequent token pair in the text corpus to generate a new token and keeps all generated tokens in the vocabulary, it unavoidably holds tokens that primarily act as components of a longer token and appear infrequently on their own. We term such tokens as Scaffold Tokens. Due to their infrequent occurrences in the text corpus, Scaffold Tokens pose a learning imbalance issue. To address that issue, we propose Scaffold-BPE, which incorporates a dynamic scaffold token removal mechanism by parameter-free, computation-light, and easy-to-implement modifications to the original BPE method. This novel approach ensures the exclusion of low-frequency Scaffold Tokens from the token representations for given texts, thereby mitigating the issue of frequency imbalance and facilitating model training. On extensive experiments across language modeling and even machine translation, Scaffold-BPE consistently outperforms the original BPE, well demonstrating its effectiveness.

Herein, we describe a nickel-catalyzed reductive decarboxylative difluoromethylation reaction of alkenes using inexpensive and easy-to-handle difluoroacetic anhydride (DFAA)/pyridine N-oxide reagent system. A variety of C(sp3)-CF2H containing compounds were prepared through a hydrodifluoromethylation process. Besides, various gem-difluoroalkenes bearing CF2H group were synthesized via defluorinative reductive cross-coupling process from trifluoromethyl-substituted alkenes using this new reaction system. Difluoroacetic anhydride has been then extended to other common alkyl anhydrides, and the corresponding hydroalkylation and defluoroalkylation processes have been successfully achieved. This method features broad substrate scope, good functional group tolerance as well as high efficiency.

Recently, Large Language Models (LLMs) have been widely adopted in a wide range of tasks, leading to increasing attention towards the research on how scaling LLMs affects their performance. Existing works, termed Scaling Laws, have discovered that the final test loss of LLMs scales as power-laws with model size, computational budget, and dataset size. However, the temporal change of the test loss of an LLM throughout its pre-training process remains unexplored, though it is valuable in many aspects, such as selecting better hyperparameters directly on the target LLM. In this paper, we propose the novel concept of Temporal Scaling Law, studying how the test loss of an LLM evolves as the training steps scale up. In contrast to modeling the test loss as a whole in a coarse-grained manner, we break it down and dive into the fine-grained test loss of each token position, and further develop a dynamic hyperbolic-law. Afterwards, we derive the much more precise temporal scaling law by studying the temporal patterns of the parameters in the dynamic hyperbolic-law. Results on both in-distribution (ID) and out-of-distribution (OOD) validation datasets demonstrate that our temporal scaling law accurately predicts the test loss of LLMs across training steps. Our temporal scaling law has broad practical applications. First, it enables direct and efficient hyperparameter selection on the target LLM, such as data mixture proportions. Secondly, viewing the LLM pre-training dynamics from the token position granularity provides some insights to enhance the understanding of LLM pre-training.