Controlling restricted knowledge in large language models is essential for model alignment and safe deployment. Test-time unlearning avoids costly retraining and parameter updates by intervening only during inference. However, existing activation-editing methods apply isolated pointwise corrections, overlooking how autoregressive generation continually reconstructs hidden states from the prompt, cache, and generated prefix. Consequently, later states may return to restricted knowledge regions after a locally successful correction, causing restricted knowledge re-entry. In this work, we propose Stateful Test-Time Unlearning via restricted knowledge boundary control (ST^2U), which formulates test-time unlearning as trajectory-wide boundary control. ST^2U first models restricted knowledge boundaries in low-dimensional invertible coordinates while leaving orthogonal non-target components unchanged. During inference, ST^2U monitors risk along the trajectory, applies minimal boundary corrections with contextual anchoring, and propagates historical correction states across tokens to mitigate knowledge re-entry. This trajectory-wide control enables more persistent forgetting while preserving non-target capabilities and limiting inference overhead. Across three benchmarks and three model families, ST^2U delivers the strongest overall balance, combining best or second-best retention with competitive forgetting and substantially less restricted-knowledge re-entry than test-time baselines (13.76

Transformer architectures have become increasingly prominent in medical image segmentation due to their self-attention mechanisms that effectively model global dependencies and long-range spatial relationships. However, the standard dot-product attention that underlies these architectures suffers from interconnected limitations that impair clinical performance. Specifically, dot-product attention exhibits high vulnerability to imaging noise and artifacts common in medical images. Additionally, its fixed scaling parameters cannot adapt to the diverse discrimination requirements of different anatomical structures, from fine-grained tumor boundaries to homogeneous organ regions. Furthermore, conventional skip connections in encoder-decoder architectures create semantic inconsistencies that further compromise segmentation accuracy. To address these challenges, we propose APFormer, a novel encoder-decoder architecture that incorporates Adaptive Probabilistic Attention (APA) and Triple-Domain Skip Connections (TDSC). The APA mechanism replaces standard dot-product attention with a probabilistic kernel based on exponential negative squared Euclidean distance. This enables dynamic modulation of discrimination sensitivity across different anatomical regions while incorporating entropy-guided sparse key selection to suppress noise and enhance signal-to-noise ratio. The TDSC module addresses semantic gaps through integrated spatial, channel, and feature domain attention mechanisms. Extensive experiments on Synapse, ACDC, SegTHOR, ISIC2017, ISIC2018, and PH2 datasets demonstrate that our method effectively balances precise boundary delineation with noise robustness, achieving superior segmentation performance across diverse anatomical structures.

This paper introduces SimPoint+, an enhanced sampled simulation methodology that addresses key limitations of the widely-used SimPoint approach. SimPoint+ achieves greater stability, accuracy, and efficiency in program analysis through three major improvements: (1) UMAP-based dimensionality reduction for Basic Block Vectors, (2) a two-stage clustering approach utilizing HDBSCAN, and (3) a lightweight cycle calibration method. Furthermore, an automated hyperparameter tuning strategy accommodates diverse program characteristics for the first two models. Evaluation of SPEC CPU 2006 benchmarks demonstrates that SimPoint+ significantly outperforms SimPoint by yielding more consistent results across runs, reducing cycle error rates by 3–5 orders of magnitude, and decreasing required simulation time by 25

Graphics processing units (GPUs) underpin high-performance computing, artificial intelligence, and scientific simulation, yet rising architectural complexity complicates timely and accurate performance evaluation. Cycle-accurate simulation is precise but prohibitively slow for design exploration; sampled-based approaches improve efficiency, but BBV-based characterizations capture only execution frequencies and insufficiently reflect instruction semantics and control-flow dynamics, especially in multi-kernel workloads.To address these limitations, this paper introduces SAGE, a semantic-aware GPU kernel representation method leveraging pretrained language models. SAGE semantically encodes warp-level instruction sequences with localized context and aggregates them hierarchically to produce kernel-level embeddings. Compared with conventional BBV-based approaches, these embeddings better capture program semantics, improving prediction accuracy and generalization on complex applications. Experimental results show that SAGE achieves approximation accuracy comparable to Photon—with relative errors of 19.70% and 19.88%, respectively, while delivering an average simulation speedup of 3.31×, representing an improvement over Photon’s 2.69×.

Large Language Models (LLMs) have demonstrated exceptional performance on natural language processing tasks. However, these models are computationally intensive and require substantial hardware resources for deployment. Quantization has emerged as a popular technique for LLM deployment, reducing memory requirements, but it results in accuracy degradation, particularly when using low-bit quantization. To mitigate this accuracy loss, we introduce Low-Rank Compensation (LoRC), a novel compensation mechanism that aims to recover the performance drop caused by quantization. Additionally, we propose Low-Rank Quantization (LoRQ), which further reduces the quantization-induced loss by adaptively adjusting weights at the element-wise level to help LLMs accommodate quantized computations. LoRC focuses on compensating for accuracy loss during inference, LoRQ integrates low-rank compensation directly into the quantization process, and they do not need end-to-end fine-tuning with LLM. Furthermore, we propose the Rank-alpha Addition Strategy (R alpha AS) to combine LoRC into the inference framework, which improves inference accuracy without increasing inference latency. Experimental results show that our method outperforms the state-of-the-art OmniQuant by 1.89% on several common zero-shot datasets under the W4A4 setting of the widely-used LLaMA. Through the joint design of algorithms and systems, our techniques can be easily integrated into the FlexGen inference framework without introducing additional inference latency, thereby maintaining high throughput while improving accuracy.

The widespread application of large language models(LLMs) has made distributed training increasingly important, especially pipeline parallelism, which is a fundamental technique for ultra-large-scale LLMs. Current research in this field mainly employs combinatorial optimization algorithms such as dynamic programming. However, as the problem size increases, these methods become difficult to solve quickly in large-scale scenarios due to their high search time. Online optimization algorithms that combine neural networks with reinforcement learning require real-time interaction with the cluster environment to obtain feedback, resulting in high resource overhead and low search efficiency. Moreover, current research lacks studies on heterogeneous computing environments, which are frequently used by small research teams. To address these issues, we designed a novel Neural Network-based Pipeline Parallel strategy solver (NN-Piper) for heterogeneous environments. NN-Piper can perceive computational and communication costs, the number of stages to be divided, and the number of micro-batches. In addition, it can directly provide the strategy for allocating specific devices to each pipeline stage. To avoid an online training process that requires interaction with the cluster environment, we propose the Virtual Contrastive Training Algorithm (VCTA) to enable efficient training of NN-Piper without collecting large amounts of real data. After training, NN-Piper can be transferred to many different scenarios without further training or fine-tuning, and it can search for strategies within a few dozen milliseconds. Compared with the state-of-the-art method, NN-Piper can improve the training speed on average by 16-25% in different environments for the transformer-based models.

Despite recent advances in lane detection methods, scenarios with limited- or no-visual-clue of lanes due to factors such as lighting conditions and occlusion remain challenging and crucial for automated driving. Moreover, current lane representations require complex post-processing and struggle with specific instances. Inspired by the DETR architecture, we propose LDTR, a transformer-based model to address these issues. Lanes are modeled with a novel anchor-chain, regarding a lane as a whole from the beginning, which enables LDTR to handle special lanes inherently. To enhance lane instance perception, LDTR incorporates a novel multi-referenced deformable attention module to distribute attention around the object. Additionally, LDTR incorporates two line IoU algorithms to improve convergence efficiency and employs a Gaussian heatmap auxiliary branch to enhance model representation capability during training. To evaluate lane detection models, we rely on Fréchet distance, parameterized Fl-score, and additional synthetic metrics. Experimental results demonstrate that LDTR achieves state-of-the-art performance on well-known datasets.

As the quantity and complexity of information processed by software systems increase, large-scale software systems have an increasing requirement for high-performance distributed computing systems. With the acceleration of the Internet in Web 2.0, Cloud computing as a paradigm to provide dynamic, uncertain and elastic services has shown superiorities to meet the computing needs dynamically. Without an appropriate scheduling approach, extensive Cloud computing may cause high energy consumptions and high cost, in addition that high energy consumption will cause massive carbon dioxide emissions. Moreover, inappropriate scheduling will reduce the service life of physical devices as well as increase response time to users' request. Hence, efficient scheduling of resource or optimal allocation of request, that usually a NP-hard problem, is one of the prominent issues in emerging trends of Cloud computing. Focusing on improving quality of service (QoS), reducing cost and abating contamination, researchers have conducted extensive work on resource scheduling problems of Cloud computing over years. Nevertheless, growing complexity of Cloud computing, that the super-massive distributed system, is limiting the application of scheduling approaches. Machine learning, a utility method to tackle problems in complex scenes, is used to resolve the resource scheduling of Cloud computing as an innovative idea in recent years. Deep reinforcement learning (DRL), a combination of deep learning (DL) and reinforcement learning (RL), is one branch of the machine learning and has a considerable prospect in resource scheduling of Cloud computing. This paper surveys the methods of resource scheduling with focus on DRL-based scheduling approaches in Cloud computing, also reviews the application of DRL as well as discusses challenges and future directions of DRL in scheduling of Cloud computing.

Existing facial expression recognition (FER) methods are mainly devoted to learning discriminative features from high-resolution images. However, when applied to low-resolution images, their performance drops rapidly. This paper proposes a unified learning framework (namely SR-FER) by cascading the image super-resolution (SR) task and FER task to alleviate the low-resolution challenge. It effectively feeds back expression-related information from the FER network to the SR network, and returns the quality-enhanced expression images via a SR network. Specifically, a multi-stage attention-aware consistency loss module is introduced to help the SR network achieve discriminative feature restoration guided by attention information. Furthermore, a prediction consistency loss module is also developed to encourage the SR network to restore discriminative features by reducing the difference in prediction information between the restored and original normal-resolution images. Therefore, more accurate results are obtained by performing FER on the restored images. We conduct extensive experiments to demonstrate that the proposed low-resolution FER solution can help SR methods restore features favorable for FER while maintaining acceptable FER performance in various resolution degradation scenarios. The proposed method effectively improves the FER challenge under resolution degradation conditions, which is of good reference value for real-world applications.

While encrypted proxy services like Shadowsocks and Vmess safeguard user privacy by encrypting traffic data, they also provide anonymity tools for cybercriminals, increasing the difficulty of regulating and combating activities such as online fraud and cyberattacks. This presents significant challenges for network management, as these services can obfuscate and disguise traffic with continually updated tools, leading to dynamic traffic features that traditional detection methods struggle to identify. However, the feature engineering and model designing of existing methods is based on previously commonly used encrypted proxy traffic and cannot cope with new types, which means that existing methods are poor in generalization ability. To overcome these challenges, this paper introduces a novel method for identifying encrypted proxy traffic. Initially, by analyzing encrypted proxy traffic, we extract features related to the information entropy, TLS protocol, and length of traffic packets. In this way, we obtain a comprehensive feature vector of length 325. Second, we propose a new identification model, called BiAE-MLP, consisting of an encoding layer with two autoencoders and a classification layer with a MLP classifier. We evaluate our method with four experiments. As a result, our method achieves a 98% F1-score on a test dataset with the same distribution as the training dataset and a 91% F1-score on a dataset with a different distribution. It indicates the effectiveness and generalization ability of our method. Additionally, in Experiment 4, we demonstrate that the encoding layer of the BiAE-MLP plays a crucial role in enhancing the model's generalization ability.

Pattern counting is a crucial task in graph pattern mining. Accurate counting is not affordable as the datasets grow larger and larger, and approximate counting is getting popular to provide an estimated answer quickly. However, current approximate counting approaches are still time-consuming and not scalable for extra-large graphs. This paper proposes SPAC, a fast and flexible pattern approximate counting method, based on the observation that pattern number distribution to degrees also follows power-law as the vertices, the common feature in graph datasets. By leveraging the distribution, SPAC can efficiently choose a small number of degrees as samples, fit the coefficients, and then calculate the pattern frequency directly. To provide flexibility for different use-cases, SPAC supports both accurate and approximate counting in the sampling phase. Moreover, edge weighting and interpolation techniques are adopted to emphasize the sample tail to improve fitting accuracy. The prototype of SPAC is implemented with GraphX on Spark, and is evaluated against various well-known graphs. The experimental results show that SPAC is up to 10x faster than accurate counting, keeping the same error level below 10%. Compared to existing approximate counting, SPAC is 1.4x–9x faster in general, while the error could be reduced to 20% of the current systems.

Pattern counting is a fundamental computational graph mining task. Recently, extensive research has been conducted on approximate pattern counting, but existing algorithms are not scalable, especially for large-scale static graphs. To address this challenge, we propose FAPC- a fast and efficient method for approximate pattern counting by exploiting the fact that the distribution of pattern numbers to degrees follows the power-law function as well. Unlike existing sampling processes, our approach can quickly fit the formula coefficients and calculate directly the pattern frequency. The prototype of FAPC has been evaluated against multiple public datasets and experimental results demonstrate that it outperforms current approaches by up to 50x speed.

Blockchains have been widely adopted to track critical data in complicated applications recently, thus it is necessary to provide efficient infrastructure to analyze large-scale chain data. However, the current blockchain clients are mostly designed for a single node with limited storage space and the separated data analysis systems are not suitable for huge chains due to redundant data transformation and scalability. To address these challenges, we propose BigChain, a distributed blockchain client with built-in bigdata processing capability. BigChain is designed on top of HBase by mapping blockchain entities to HBase’s KV storage to guarantee the rapid growth of on-chain data. By integrating the MapReduce framework, BigChain could leverage the Hadoop big data ecosystem directly to analyze blockchain data in a scalable manner. Additionally, an indexing service is devised accordingly to speed up BigChain. Finally, comprehensive experiments were conducted against Ethereum, and the results indicate that BigChain can outperform typical clients by 10 times in terms of comprehensive scenarios.

Lane detection is challenging due to the complicated onroad scenarios and line deformation from different camera perspectives. Lots of solutions were proposed, but can not deal with "corner lanes" well. To address this problem, this paper proposes a new top-down deep learning lane detection approach, CANet. A lane instance is first responded by the heatmap on the U-shaped "curved guide line" at global semantic level, thus the corresponding features of each lane are aggregated at the response point. Then CANet obtains the heatmap response of the entire lane through conditional convolution, and finally decodes the point set to describe lanes via adaptive decoder. The prototype is implemented with Pytorch, and evaluated against 3 well-known datasets extensively. The experimental results show that CANet reaches SOTA in different metrics.

Person re-identification (ReID) is a challenging cross-camera retrieval task to identify pedestrians. Many complex network structures are proposed recently and many of them concentrate on multi-branch features to achieve high performance. However, they are too heavy-weight to deploy in real-world applications. Additionally, pedestrian images are often captured by different surveillance cameras, so the varied lights, perspectives and resolutions result in inevitable multi-camera domain gaps for ReID. To address these issues, this paper proposes ATCN, a simple but effective angular triplet loss-based camera network, which is able to achieve compelling performance with only global features. In ATCN, a novel angular distance is introduced to learn a more discriminative feature representation in the embedding space. Meanwhile, a lightweight camera network is designed to transfer global features to more discriminative features. ATCN is designed to be simple and flexible so it can be easily deployed in practice. The experiment results on various benchmark datasets show that ATCN outperforms many SOTA approaches.

Distributed storage systems usually adopt replication for reliability and fast access. However, as the data volume grows, many large-scale storage systems are tending to employ erasure coding to reduce the storage overhead of replication while deliver the same reliability. Unfortunately, erasure coding could result in performance degradation due to less data locality and degraded reads. To trade off among reliability, performance and storage overhead at the same time, we propose FlexBM, a flexible dual-scheme block management approach. FlexBM supports both replication and erasure coding simultaneously, and applies them dynamically according to the recent data temperature. Erase coding is for cold data to reduce storage, while replication is for hot files so that applications can leverage data locality. To guarantee the same reliability as replication with fewer replicas, FlexBM models block placement with bipartite graphs. The prototype of FlexBM is implemented based on HDFS. The experimental results show that FlexBM succeeds in reducing the storage overhead even for a scenario with many small files without reliability compromising, and meanwhile, providing better data locality for frequently accessed datasets.

Graph has been widely used in complex network applications modeling, and the asynchronous graph processing model is superceding the BSP model because of its better convergence speed. However, the asynchronous GAS model proposed by PowerGraph usually results in irregular and unpredictable communication patterns as well as vertex-scale barriers, so it is difficult for programmers to optimize codes. To address these challenges, we propose LMCC, an improved message management approach including lazy pull-message model and vertex-oriented centralized cache, which can reduce communication cost in terms of message quantity, and reduce the number of computation iterations in turn, without compromising the accuracy of application results. Based on the deep investigation of the GAS phases, LMCC is designed to be totally transparent to user applications. Experimental results show that LMCC can deliver speedup for various types of graph computing benchmarks ranging from 129% to 271%.

In the big data era, the distributed file system is getting more and more significant due to the characteristics of its scale-out capability, high availability, and high performance. Different distributed file systems may have different design goals. For example, some of them are designed to have good performance for small file operations, such as GlusterFS, while some of them are designed for large file operations, such as Hadoop distributed file system. With the divergence of big data applications, a distributed file system may provide good performance for some applications but fails for some other applications, that is, there has no universal distributed file system that can produce good performance for all applications. In this paper, we propose a hybrid file system framework, HybridFS, which can deliver satisfactory performance for all applications. HybridFS is composed of multiple distributed file systems with the integration of advantages of these distributed file systems. In HybridFS, on top of multiple distributed file systems, we have designed a metadata management server to perform three functions: file placement, partial metadata store, and dynamic file migration. The file placement is performed based on a decision tree. The partial metadata store is performed for files whose size is less than a few hundred Bytes to increase throughput. The dynamic file migration is performed to balance the storage usage of distributed file systems without throttling performance. We have implemented HybridFS in java on eight nodes and choose Ceph, HDFS, and GlusterFS as designated distributed file systems. The experimental results show that, in the best case, HybridFS can have up to 30% performance improvement of read/write operations over a single distributed file system. In addition, if the difference of storage usage among multiple distributed file systems is less than 40%, the performance of HybridFS is guaranteed, that is, no performance degradation.

Metadata scalability is critical for distributed systems as the storage scale is growing rapidly. Because of the strict consistency requirement of metadata, many existing metadata services utilize a fundamentally unscalable design for the sake of easy management, while others provide improved scalability but lead to unacceptable latency and management complexity. Without delivering scalable performance, metadata will be the bottleneck of the entire system. Based on the observation that real file dependencies are few, and there are usually more idempotent than non-idempotent operations, we propose a practical strategy, Replichard, allowing a tradeoff between metadata consistency and scalable performance. Replichard provides metadata services through a cluster of metadata servers, in which a flexible consistency scheme is adopted: strict consistency for non-idempotent operations with dynamic write-lock sharding, and relaxed consistency with accuracy estimations of return values where consistency for idempotent requests is relaxed to achieve high throughput. Write-locks are dynamically created at subtree-level and designated to independent metadata servers in an application-oriented manner. A subtree metadata update that occurs on a particular server is replicated to all metadata servers conforming to the application "start-end" semantics, resulting in an eventually consistent namespace. An asynchronous notification mechanism is also devised to enable users to deal with potential stale reads from operations of relaxed consistency. A prototype was implemented based on HDFS, and the experimental results show promising scalability and performance for both micro benchmarks and various real-world applications written in Pig, Hive and MapReduce.