Large Language Models (LLMs) power numerous AI applications, yet updating their knowledge remains costly. Model editing provides a lightweight alternative through targeted parameter modifications, with meta-learning-based model editing (MLME) demonstrating strong effectiveness and efficiency. However, we find that MLME struggles in low-data regimes and incurs high training costs due to the use of KL divergence. To address these issues, we propose Efficient Multi-Step Edit (EMSEdit), which leverages multi-step backpropagation (MSBP) to effectively capture gradient-activation mapping patterns within editing samples, performs multi-step edits per sample to enhance editing performance under limited data, and introduces norm-based regularization to preserve unedited knowledge while improving training efficiency. Experiments on two datasets and three LLMs show that EMSEdit consistently outperforms state-of-the-art methods in both sequential and batch editing. Moreover, MSBP can be seamlessly integrated into existing approaches to yield additional performance gains. Further experiments on a multi-hop reasoning editing task demonstrate EMSEdit's robustness in handling complex edits, while ablation studies validate the contribution of each design component. Our code is available at https://github.com/xpq-tech/emsedit.

WiFi fingerprinting remains a cornerstone of indoor localization, yet conventional KNN-based approaches suffer from high-dimensional noise and suboptimal parameter selection. While metaheuristic algorithms like GSA have been applied to optimize access point (AP) selection and K values, they often face high computational complexity in large-scale environments. This paper proposes SA-RLS-GSA-KNN, a novel two-stage optimization framework that integrates sparse-aware recursive least squares (SA-RLS) with the gravitational search algorithm (GSA) to optimize the parameters of KNN. First, SA-RLS with L1 regularization pre-screens critical APs by modeling RSS-position relationships, significantly reducing the search space. Second, GSA fine-tunes the final AP subset and optimal K value within this reduced space using a multi-objective fitness function balancing positioning error, feature count, and computational cost. The evaluation was conducted on 15 public datasets such as UJIIndoorLoc. The average positioning error achieved by SA-RLS-GSA was more than 20% lower than that of the baseline KNN algorithm.

Multimodal Large Language Models (MLLMs) have achieved remarkable progress in vision–language understanding, yet how they internally integrate visual and textual information remains poorly understood. To bridge this gap, we perform a systematic layer-wise masking analysis across multiple architectures, revealing how visual–text fusion evolves within MLLMs. The results show that fusion emerges at several specific layers rather than being uniformly distributed across the network, and certain models exhibit a late-stage “review” phenomenon where visual signals are reactivated before output generation. Besides, we further analyze layer-wise attention evolution and observe persistent high-attention noise on irrelevant regions, along with gradually increasing attention on text-aligned areas. Guided by these insights, we introduce a training-free contrastive attention framework that models the transformation between early fusion and final layers to highlight meaningful attention shifts. Extensive experiments across various MLLMs and benchmarks validate our analysis and demonstrate that the proposed approach improves multimodal reasoning performance.Code will be released.

As large language model agents gain access to increasingly large skill libraries, retrieving the right skill becomes critical to reliable capability selection and execution. Existing retrievers often treat skill contents as ordinary documents, overlooking their highly regular structure: shared descriptive patterns recur across many skills while providing little evidence for distinguishing the required capability. We show that this shared descriptive background is reflected in dense relevance scores, induces a pronounced energy gap between queries and skill documents, and obscures discriminative signals, especially for structurally similar hard negatives. Based on this observation, we propose SkillSight, a training-free retrieval framework that calibrates shared background in both semantic and lexical spaces. Semantic Background Calibration estimates a background subspace from generic tokens identified by IDF, reducing similarity induced by shared descriptive patterns, while Lexical Evidence Calibration downweights shared background tokens to recover discriminative token-level evidence. Experiments on SRA-Bench and SkillBench-Supp demonstrate consistent improvements across retrieval metrics, with SkillSight improving Recall@10 by up to 20.21 percentage points over the original dense retriever. It is up to 1,248 times faster than the Dense + Reranker baseline. In end-to-end evaluation, SkillSight achieves the best overall performance across three agent models and outperforms LLM Selection by up to 4.97 percentage points. These results identify shared descriptive background as a source of ranking interference in skill retrieval and demonstrate that calibrating it enables accurate and efficient skill selection without additional training. Our code can be found at https://github.com/xiaojinying/SkillSight

The Mixture-of-Experts (MoE) architecture enhances model scalability through sparse activation, but exacerbates VRAM bottlenecks in local deployments. Existing research commonly adopts offloading techniques to mitigate this issue by dynamically transferring model parameters between GPU and CPU memory. Existing studies typically improve inference efficiency by introducing layer-level global greedy orchestration to optimize offloading decisions, but the orchestration generally suffers from a collapse of heterogeneous parallelism gains, which introduce redundant overhead in single-request decoding. Moreover, rigid adherence to the optimal top- k routing forces a sub-optimal choice between slow CPU computation and PCIe weight transfers upon GPU cache misses. To address these issues, we propose SPAR, Step-wise Path Dispatching and Asymmetric Re-routing for efficient MoE Inference. Specifically, it retains global orchestration in prefill to exploit batch parallelism, while introducing a step-wise deterministic path for dispatching in decoding, which alleviates the redundant overhead. Moreover, SPAR introduces an asymmetric re-routing strategy that leverages semantic redundancy to avoid both CPU computation latency and PCIe activation transfers. Experimental results on Mixtral-8 × 7B demonstrate that SPAR achieves up to 1.75× average throughput speedup across three NVIDIA GPUs. Additionally, extensive analyses demonstrate that SPAR maintains high generation quality. Ablation studies further validate the effectiveness of SPAR.

Aligning natural language with operating system (OS) commands allows users to perform complex computer tasks through simple natural language descriptions. However, due to the complex nature of natural language, it still remains challenging to achieve precise alignment. In this paper, we present ComAlign, a Chinese benchmark dataset that pairs Chinese natural language descriptions with corresponding OS commands. ComAlign covers a broad range of 82 distinct OS command types with a total of 1811 natural language descriptions. We elaborate on the construction of ComAlign and construct three baselines to evaluate the alignment accuracy on ComAlign. Experimental results show that even advanced large language models struggle with certain ambiguously phrased OS commands. Specifically, the best performing baseline achieves 46.9% alignment accuracy. We demonstrate that ComAlign is collected from real-world application scenarios, making it particularly suitable for developing and benchmarking intelligent OS and agent systems that support user-machine interactions through natural language.

To prevent Large Language Models (LLMs) from being misused by bad actors, model developers delve into designing safety mechanisms to guarantee LLMs generate helpful, honest, and harmless responses. To disclose flaws in the safety mechanisms, researchers conduct adversarial attacks against LLMs to circumvent the safety mechanisms, which are called “jailbreak attack”. Existing jailbreak attacks solely focus on engineering toxic queries and adversarial prompts to induce LLMs to generate toxic responses, losing sight of the study starting from engineering non-toxic queries and adversarial prompts, which also plays an important role in disclosing flaws in safety mechanisms. To fill the research gap, we propose SEAttack, a self-evolving jailbreak attack method to induce LLMs to generate toxic responses for non-toxic queries. Given a non-toxic query, SEAttack initially generates a response, which is more likely to be non-toxic due to the safety mechanisms. Then, it uses multiple iterations of self-evolving to evolve the non-toxic response to a toxic one. To evaluate SEAttack, we construct JailChat, a dataset containing 3000 non-toxic queries. We drive SEAttack to attack eighteen state-of-the-art LLMs, including five closed-source and thirteen open-source LLMs. Experimental results demonstrate that SEAttack achieves up to 89.07 % attack success rate, revealing non-negligible flaws in LLMs’ safety mechanisms. Moreover, we track the changes in the safety mechanisms of four ChatGPT variants. Extensive analyses and human evaluation further validate the effectiveness and rationality of SEAttack.

Automated vulnerability detection is crucial for enhancing software security by identifying potential flaws that attackers could exploit, thereby reducing the reliance on labor-intensive manual code audits. Recent advancements have shifted towards leveraging large language models (LLMs) for vulnerability detection, with techniques like Vul-RAG and VulnSage demonstrating progress through structured prompting and external knowledge integration. However, these approaches typically rely on a single reasoning paradigm, limiting their ability to address the complex and diverse nature of real-world vulnerabilities. To overcome these limitations, we propose ReasonVul, a novel multi-perspective reasoning framework that harnesses cognitive synergy among three specialized LLM agents, each embodying a distinct reasoning mode. The framework begins with independent analyses of the source code, followed by a structured debate mechanism to resolve conflicts through iterative rebuttal and revision, ultimately converging on a collaborative judgment. Evaluated on the PrimeVul dataset, ReasonVul achieves a PairAcc of 40.00

Deobfuscation is critical to reverse engineering and security analysis because it restores the readability and analyzability of obfuscated code. However, existing research primarily focuses on source-code deobfuscation, while binary-level deobfuscation remains largely underexplored despite its practical importance when source code is unavailable. Existing binary deobfuscation methods typically decompile binaries into pseudocode and then apply structural transformations. However, because compilation discards high-level semantics such as precise type information and source-level structures, this decompilation-based paradigm often produces low-quality code and provides limited assurance that the recovered code preserves the runtime behavior of the original program. To address these limitations, we propose a paradigm shift from structural transformation to behavior-driven synthesis. Our core insight is that although obfuscation distorts a program's internal structure, semantics-preserving transformations must retain its observable execution behavior. Based on this insight, we introduce BinMirror, an approach that reformulates binary deobfuscation as a behavior-specification-guided program synthesis task. By treating dynamic execution traces and interaction snapshots as behavioral specifications, BinMirror synthesizes high-quality source code and validates it against runtime observations collected from heavily obfuscated binaries. Extensive evaluations on 1.5 million synthetically obfuscated binaries show that BinMirror significantly outperforms state-of-the-art baselines, achieving a unit-test Pass@1 of 74.5

Large language and vision models have demonstrated remarkable performance, but their high pre-training and fine-tuning costs have led to growing interest in accelerating training through dataset pruning. Traditional pruning approaches typically rely on sample loss to identify and retain the most "difficult" samples. However, as the pruning rate increases, the training frequency of each sample becomes more uniform, which can result in the underfitting of critical or general samples. We identify this phenomenon as Low-Frequency Learning (LFL), where the model fails to retain knowledge of the majority of samples. In this work, we decompose the scoring function of LFL, providing a theoretical analysis of its inefficiencies. To counteract this issue, we propose a novel enhancement to the scoring function by introducing a memory term designed to strengthen the model’s ability to memorize essential data. We also offer an approximation for this memory term. Additionally, we extend this concept to Self-Supervised Learning (SSL), marking the first investigation into the role of memory in SSL. By leveraging contrastive learning, we derive the memory term both theoretically and experimentally. Based on these insights, we introduce Enhance Memory Pruning (EMP), a technique that mitigates memory loss under high pruning rates by improving the model’s data retention. EMP has been evaluated across various tasks, including image classification, natural language understanding, and model pre-training. Our experiments demonstrate that EMP significantly enhances model performance, particularly under extreme pruning conditions. For instance, in the CIFAR100-ResNet50 pre-training task with a 70

Large language models (LLMs) are popping up all over the place, and they have been gaining prominence due to their exceptional abilities in conducting various tasks. Although extensive LLM evaluation has been explored on natural language understanding tasks like text classification and sentiment analysis, evaluating LLMs on named entity recognition (NER) still remains under-explored. To fill this gap, we evaluate twenty-eight representative LLMs on thirteen datasets across five domains, whose parameters range from 3 billion to 175billion, from four perspectives, that is, supervised fine-tuning (SFT), parameter scales, hallucinations, and prompt designs. We propose an LLM-based NER framework (LLM-NER) for the evaluation, which consists of a Recognition phase and a Check phase. Specifically, the Check guides LLMs to examine the correctness of recognized entities, which is designed to mitigate hallucinations in the NER scenario. Qualitative and quantitative evaluation analyses demonstrate that in the NER scenario: 1) SFT empowers LLMs to understand and follow human instructions; 2) LLMs’ ability generally improves as their parameter scales consistently increase; 3) hallucinations exist in all evaluated LLMs, and guiding LLMs to check their outputs is a feasible way to alleviate hallucinations; and 4) all evaluated LLMs are sensitive to prompt designs. Based on the analyses, we highlight a number of promising directions for future study. Moreover, our evaluation shows high consistency with two LLM evaluation leaderboards, which evaluate LLMs on other tasks, demonstrating the rationality of our evaluation design.

Entity Linking (EL) plays a crucial role in mapping textual mentions to corresponding entities in structured knowledge bases, while Multi-modal Entity Linking (MEL) extends this task by incorporating both textual and visual information. A key challenge in MEL is effectively utilizing multi-modal contextual data to improve entity disambiguation, particularly when candidate entities are highly similar. In this paper, we propose a novel approach that leverages the cross-modal attention and reasoning capabilities of multi-modal large language models (MLLM) to enhance MEL in an unsupervised setting. Our model consists of an optimized re-ranking stage that reduces computational cost by narrowing down the number of candidate entities, and a comparative selection stage that improves entity mention identification accuracy by fully leveraging indirect and latent information to distinguish highly similar candidates. Experimental results on three benchmark datasets show significant improvements in top-1 accuracy, with our model achieving 94.83% , 94.19% , and 91.25% on the Wiki-MEL, Richpedia-MEL, and WikiDiverse datasets, respectively, surpassing state-of-the-art approaches by up to 4.95% or more.

The Go programming language has been widely used in many cloud infrastructures. However, Go's concurrency features will introduce channel-related concurrency bugs which are caused by misuse of channels and locks. Several concurrency testing tools try to detect channel-related concurrency bugs by exploring the interleaving space during runtime. However, these tools naively control the interleavings by delay injection and do not consider equivalence relation, thus can not achieve efficiency. To address these limitations, we propose GoAway, a testing approach to detect Go concurrency bugs through a coverage-guided scheduler. GoAway uses statebased scheduler to achieve deterministic multi-threading and execution graph as equivalence coverage to guide scheduling. Experiments show that GoAway can trigger concurrency bugs with 54 % fewer tests on average.

The Go programming language has been used in many cloud native systems. However, Go programs suffer from channel-related concurrency bugs, such as goroutines blocking indefinitely due to unmatched channel operations or mixed use of channels and locks, which can lead to resource leaks and system failures. Several concurrency testing tools for Go have been designed to detect such concurrency bugs via exploring interleaving space. However, these tools do not consider equivalent interleavings and naively choose to-be-tested interleaving. To address these limitations, we propose EPCT, a testing approach to trigger Go concurrency bugs through equivalent interleaving priorities. EPCT uses read-write relation to represent equivalent interleavings and then uses the equivalence difference to guide the priorities of to-be-tested interleavings. Experiments show that EPCT can trigger more bugs while reducing test runs by at least 30%.

The software development process is characterized by an iterative cycle of continuous functionality implementation and debugging, essential for the enhancement of software quality and adaptability to changing requirements. This process incorporates two isolatedly studied tasks: Code Search (CS), which retrieves reference code from a code corpus to aid in code implementation, and Fault Localization (FL), which identifies code entities responsible for bugs within the software project to boost software debugging. The basic observation of this study is that these two tasks exhibit similarities since they both address search problems. Notably, CS techniques have demonstrated greater effectiveness than FL ones, possibly because of the precise semantic details of the required code offered by natural language queries, which are not readily accessible to FL methods. Drawing inspiration from this, we hypothesize that a fault localizer could achieve greater proficiency if semantic information about the buggy methods were made available. Based on this idea, we propose \(\mathtt{CosFL}\), an FL approach that decomposes the FL task into two steps: query generation, which describes the functionality of the problematic code in natural language, and fault retrieval, which uses CS to find program elements semantically related to the query, allowing for finishing the FL task from a CS perspective. Specifically, to depict the buggy functionalities and generate high-quality queries, \(\mathtt{CosFL}\) extensively harnesses the code analysis, semantic comprehension, text generation, and decision-making capabilities of LLMs. Moreover, to enhance the accuracy of CS, \(\mathtt{CosFL}\) captures varying levels of context information and employs a multi-granularity code search strategy, which facilitates a more precise identification of buggy methods from a holistic view. The evaluation on 835 real bugs from 23 Java projects shows that \(\mathtt{CosFL}\) successfully localizes 324 bugs within Top-1, which significantly outperforms the state-of-the-art approaches by 26.6%-57.3%. The ablation study and sensitivity analysis further validate the importance of different components and the robustness of \(\mathtt{CosFL}\) across different backend models.

Large Language Models (LLMs) quantization facilitates deploying LLMs in resource-limited settings, but existing methods that combine incompatible gradient optimization and quantization truncation lead to serious convergence pathology. This prolongs quantization time and degrades LLMs' task performance. Our studies confirm that Straight-Through Estimator (STE) on Stiefel manifolds introduce non-smoothness and gradient noise, obstructing optimization convergence and blocking high-fidelity quantized LLM development despite extensive training. To tackle the above limitations, we propose SingleQuant, a single-pass quantization framework that decouples from quantization truncation, thereby eliminating the above non-smoothness and gradient noise factors. Specifically, SingleQuant constructs Alignment Rotation Transformation (ART) and Uniformity Rotation Transformation (URT) targeting distinct activation outliers, where ART achieves smoothing of outlier values via closed-form optimal rotations, and URT reshapes distributions through geometric mapping. Both matrices comprise strictly formulated Givens rotations with predetermined dimensions and rotation angles, enabling promising LLMs task performance within a short time. Experimental results demonstrate SingleQuant's superiority over the selected baselines across diverse tasks on 7B-70B LLMs. To be more precise, SingleQuant enables quantized LLMs to achieve higher task performance while necessitating less time for quantization. For example, when quantizing LLaMA-2-13B, SingleQuant achieves 1,400× quantization speedup and increases +0.57% average task performance compared to the selected best baseline.

As the foundation of modern software ecosystems, a distribution of Operating System (OS) is usually composed of a certain number of software packages among which exist complex dependency relationships. The refactoring of package dependency structures is an important form of evolution between different distributions that has been key factor affecting the compatibility of OS. Existing studies primarily focus on static dependencies or specific software subsets, lacking systemlevel modeling and structural analysis of cross-version evolution of the OS. To address this gap, we propose a unified framework for system-level modeling and analysis based on graph embedding techniques, designed to characterize structural changes in OS package dependencies over time. Our method constructs a homogeneous dependency graph spanning multiple OS versions and applies the HOPE embedding algorithm to extract evolution features and detect changes in dependency structure. Using this framework, we identify a novel evolution pattern -intermediate-layer phenomena - where functional roles remain stable despite significant structural refactoring. We further classify these phenomena into two types: refactored intermediate layers, which optimize architecture by reorganizing dependency topologies, and transition intermediate layers, which preserve interface compatibility through proxy packages. We validate our approach on six Ubuntu LTS versions (14.04-24.04), achieving 85.5% accuracy in detecting refactored intermediate layers and successfully uncovering unmarked transition proxy packages. This work offers a new perspective and technical means for analyzing OS structural evolution and assessing compatibility risks.

Large language models (LLMs) generate human-aligned content under certain safety constraints. However, the current known technique “jailbreak prompt” can circumvent safety-aligned measures and induce LLMs to output malicious content. Research on Jailbreaking can help identify vulnerabilities in LLMs and guide the development of robust security frameworks. To circumvent the issue of attack templates becoming obsolete as models evolve, existing methods adopt iterative mutation and dynamic optimization to facilitate more automated jailbreak attacks. However, these methods face two challenges: inefficiency and repetitive optimization, as they overlook the value of past attack experiences. To better integrate past attack experiences to assist current jailbreak attempts, we propose the JailExpert , an automated jailbreak framework, which is the first to achieve a formal representation of experience structure, group experiences based on semantic drift, and support the dynamic updating of the experience pool. Extensive experiments demonstrate that JailExpert significantly improves both attack effectiveness and efficiency. Compared to the current state-of-the-art black-box jailbreak method, JailExpert achieves an average increase of 24% in attack success rate and 2.7 times improvement in attack efficiency.