Intrinsic functions are specialized functions provided by the compiler that efficiently operate on architecture-specific hardware, allowing programmers to write optimized code in a high-level language that fully exploits hardware features. Using intrinsics to vectorize core code blocks is a standard optimization method in high-performance libraries, often requiring specific vector optimization implementations for multiple mainstream architectures. The promising RISC-V software ecosystem has a significant demand for algorithm library migration and adaptation. Translating existing intrinsic functions to RISC-V Vector (RVV) intrinsic functions across architectures is currently a mainstream approach. Rule-based intrinsic mapping methods and LLM-based code generation can help developers address the code migration challenge. However, existing intrinsic code benchmarks focus on mainstream SIMD intrinsics and lack support for the emerging RISC-V architecture. There is currently no benchmark that comprehensively evaluates the intrinsic migration capabilities for the RVV extension. To fill this gap, we propose VecIntrinBench, the first intrinsic benchmark encompassing RVV extensions. It includes 50 function-level tasks from open source repositories, implemented as scalars, RVV intrinsics, Arm Neon intrinsics, and x86 intrinsics, along with comprehensive functional and performance test cases. We systematically evaluated various code migration approaches on VecIntrinBench, yielding a series of insightful findings. The results demonstrate that advanced Large Language Models (LLMs) achieve a similar effect as rule-based mapping approaches for RISC-V code migration, while also delivering superior performance. We further analyze the reasons and identify future directions for LLM development in the code migration field. The VecIntrinBench is open-sourced to benefit the broader community and developers.

The use of intrinsic functions to exploit hardware-specific capabilities is an important approach for optimizing library performance. Many mainstream libraries implement a large number of vectorized algorithms on Arm or x86 SIMD intrinsic functions. With the rapid expansion of the RISC-V hardware-software ecosystem, there is a growing demand for support of the RISC-V Vector (RVV) extension. Translating existing vectorized intrinsic code onto RVV intrinsics is a practical and effective approach. However, current cross-architecture translation largely relies on manual rewriting, which is time-consuming and error-prone. Furthermore, while some rule-based methods can reduce the need for manual intervention, their translation success rate is limited by incomplete rule coverage and syntactic constraints, and the performance suffers from inadequate utilization of RVV-specific features. We present IntrinTrans, a LLM-based multi-agent approach that utilizes compile-and-test feedback to translate intrinsic code across architectures automatically, and further optimizes the generated RVV intrinsics using register-usage information derived from liveness analysis. To evaluate the effectiveness of our approach, we collected 34 vectorized algorithm cases from open-source libraries. Each case includes an Arm Neon intrinsics implementation and a RVV intrinsics implementation contributed by the open-source community, together with correctness and performance tests. Our experiments show that advanced LLMs produce semantically correct RISC-V Vector intrinsics in most cases within a limited number of iterations, and in some cases achieve up to 5.93x the performance of the native implementation from the open-source community.

Software repositories such as PyPI and npm are vital for software development but expose users to serious security risks from malicious packages. The malicious packages often execute their payloads immediately upon installation, leading to rapid system compromise. Existing detection methods are heavily dependent on difficult-to-obtain explicit knowledge, rendering them susceptible to overlooking emergent malicious packages. In this paper, we present a lightweight and effective method, namely EMPHunter, to detect malicious packages without requiring any explicit prior knowledge. EMPHunter is founded upon two fundamental and insightful observations. First, malicious packages are considerably rarer than benign ones, and second, the functionality of installation scripts for malicious packages diverges significantly from those of benign packages, with the latter frequently forming clusters. Consequently, EMPHunter utilizes the clustering technique to group the unique installation scripts of new-uploaded packages and identifies outliers as candidate malicious packages. It then ranks the outliers according to their deviate degrees and the distance between each of them and known malicious instances, effectively highlighting potential malicious packages. With EMPHunter, we successfully identified 122 previously unknown malicious packages from a pool of 267,009 newly-uploaded PyPI and npm packages, achieving an mAP (Mean Average Precision) of 0.813 and an exceptional recall of 0.992 when auditing the top-10 rankings. All detected packages have been officially confirmed as genuine malicious package by PyPI and npm. We assert that EMPHunter offers a valuable and advantageous supplement to existing detection tools, augmenting the arsenal of software supply chain security analysis.

Using data mining to extract frequent code patterns for bug detection has proven effective. However, prior studies have overlooked the prevalence of infrequent ( rare ) patterns, even though violations of such patterns can also lead to bugs. In this paper, we present LTMiner, which mines rare patterns from large-scale projects and detects potential bugs by checking for violations of these patterns. In practice, rare patterns far outnumber frequent ones and lack strong statistical support. Consequently, we face a pattern explosion, and many rare patterns and their violations are uninteresting. LTMiner addresses this by using instance-based ranking and filtering to prioritize violations of rare patterns. It further employs a large language model (LLM) as a domain expert to audit top-ranked violations; mined information supports in-context learning, and task decomposition and self-reflection mitigate possible hallucinations. This pipeline effectively curbs pattern explosion and false positives, uncovering previously unknown bugs in large-scale projects at an acceptable cost. Applied to Linux kernel 6.12.1, LTMiner identified 42 previously unknown bugs, 27 of which have been confirmed by developers. These results indicate that, although rare-pattern bugs are sparse, a considerable number remain and exhibit a non-negligible long tail. We believe that rare-pattern bugs constitute a promising blue ocean for bug detection.

Measuring the function similarity to detect bugs is effective, but the statements unrelated to the bugs can impede the performance due to the noise interference. Suppressing the noise interference in existing works does not manage the tough job, i.e., eliminating the noise in the targets. In this paper, we propose MATUS to mitigate the target noise for precise bug detection based on similarity measurement. Feature slices are extracted from both the buggy query and the targets to represent the semantic feature of (potential) bug logics. In particular, MATUS guides the target slicing with the prior knowledge from the buggy code, in an end-to-end way to pinpoint the slicing criterion in the targets. All feature slices are embedded and compared based on the vector similarity. Buggy candidates are audited to confirm unknown bugs in the targets. Experiments show that MATUS holds advantages in bug detection for real-world projects with acceptable efficiency. In total, MATUS has spotted 31 unknown bugs in the Linux kernel. All of them have been confirmed by the kernel developers, and 11 have been assigned CVEs.

Database configuration tuning is critical for workload performance, but remains difficult on real deployments. Existing tuners may leave important regions of the configuration space unexplored, require costly workload replays, and provide limited support for safely applying and recovering configuration changes. To address these limitations, we propose AgenticDB, a self-evolving agentic framework for database workload reconfiguration. Specifically, AgenticDB implements a context-grounded harness over a broad yet safe cross-layer action space spanning DBMS and OS parameters. The harness operates through two coupled technical routes. Context-grounded bottleneck diagnosis interprets workload characteristics, deployment context, and runtime-state evidence to identify likely execution bottlenecks and prescribe targeted DBMS/OS actions. Closed-loop context evolution consolidates execution feedback and reusable experience to update subsequent diagnoses, reconfiguration focus, and loop decisions, reducing low-value exploration and unnecessary workload replays. These routes are supported by an execution-aware lifecycle that validates, applies, and verifies configuration changes and restores the last safe DBMS/OS state when failures occur. We evaluate AgenticDB on MySQL and PostgreSQL using YCSB, Sysbench, and TPC-H workloads. AgenticDB achieves the strongest observed performance across all evaluated workloads, improving over the strongest baseline by 118.1

Compiler auto-tuning faces a dichotomy between traditional black-box search methods, which lack semantic guidance, and recent Large Language Model (LLM) approaches, which often suffer from superficial pattern matching and causal opacity. In this paper, we introduce ECCO, a framework that bridges interpretable reasoning with combinatorial search. We first propose a reverse engineering methodology to construct a Chain-of-Thought dataset, explicitly mapping static code features to verifiable performance evidence. This enables the model to learn the causal logic governing optimization decisions rather than merely imitating sequences. Leveraging this interpretable prior, we design a collaborative inference mechanism where the LLM functions as a strategist, defining optimization intents that dynamically guide the mutation operations of a genetic algorithm. Experimental results on seven datasets demonstrate that ECCO outperforms the LLVM opt -O3 baseline, achieving an average 24.44% reduction in cycles. Our code is available at https://anonymous.4open.science/r/ECCO-Evidence-Driven-Causal-Reasoning-for-Compiler-Optimization-3AD2.

Convolutional neural networks (CNNs) deployed on unreliable hardware are vulnerable to soft errors that can severely degrade inference accuracy. Existing algorithm-level fault-tolerance methods often require retraining or offline profiling and may fail under high fault rates or deep architectures. We propose BN-Guard, a lightweight runtime fault-mitigation mechanism that leverages Batch Normalization (BN) statistics to define layer-wise activation bounds and clip abnormal activations during inference. BN-Guard requires no retraining, offline profiling, or architectural changes, making it directly applicable to pre-trained models. Experiments on various popular CNN architectures under activation and weight faults show that BN-Guard restores up to 95

Linux kernel tuning is essential for optimizing operating system (OS) performance, yet remains challenging due to the complex kernel space, sparse performance feedback, and strong workload sensitivity. We present TuneAgent, an agentic Linux kernel tuning framework powered by rule-based reinforcement learning (RL). TuneAgent formulates the kernel space as a constrained RL environment, enabling large language models (LLMs) to autonomously explore the kernel while enforcing valid and precise configuration modifications. To address sparse performance feedback, we design structured reward functions that jointly promote reasoning standardization, configuration correctness, and performance awareness. Furthermore, we propose a two-phase training strategy that first ensures format and semantic correctness and then transitions to performance-driven exploration, accelerating convergence and reducing overhead. Experimental results show that TuneAgent consistently outperforms existing baselines, achieving up to 5.6

Documentation has long guided computer system tuning by distilling expert knowledge into per-parameter recommendations. Yet such guides capture only what experts conclude, discarding how they reason. This fundamental gap manifests in three concrete deficiencies: documentation grows stale as software evolves, fails under heterogeneous workloads, and ignores inter-parameter dependencies. We propose shifting from static documentation to dynamic action for system tuning. We introduce PerfEvolve, which translates expert tuning methodologies into executable skills that equip LLM-based agents to perform version-consistency verification, workload-specific profiling, and multi-parameter joint optimization. Evaluated on PostgreSQL under TPC-C and TPC-H benchmarks, PerfEvolve outperforms state-of-the-art documentation-driven tuning baselines by up to 35.2

Learning effective numerical representations, or embeddings, of programs is a fundamental prerequisite for applying machine learning to automate and enhance compiler optimization. Prevailing paradigms, however, present a dilemma. Static representations, derived from source code or intermediate representation (IR), are efficient and deterministic but offer limited insight into how a program will behave or evolve under complex code transformations. Conversely, dynamic representations, which rely on runtime profiling, provide profound insights into performance bottlenecks but are often impractical for large-scale tasks due to prohibitive overhead and inherent non-determinism. This paper transcends this trade-off by proposing a novel quasi-dynamic framework for program representation. The core insight is to model a program's optimization sensitivity. We introduce the Program Behavior Spectrum, a new representation generated by probing a program's IR with a diverse set of optimization sequences and quantifying the resulting changes in its static features. To effectively encode this high-dimensional, continuous spectrum, we pioneer a compositional learning approach. Product Quantization is employed to discretize the continuous reaction vectors into structured, compositional sub-words. Subsequently, a multi-task Transformer model, termed PQ-BERT, is pre-trained to learn the deep contextual grammar of these behavioral codes. Comprehensive experiments on two representative compiler optimization tasks---Best Pass Prediction and -Oz Benefit Prediction---demonstrate that our method outperforms state-of-the-art static baselines. Our code is publicly available at https://anonymous.4open.science/r/PREP-311F/.

Single Instruction Multiple Data (SIMD) is a key technique for exploiting the processors’ data-level parallelism and improving computational performance. Many high-performance computing libraries therefore rely heavily on SIMD-based optimizations. However, existing SIMD-optimized libraries provide limited support for the RISC-V Vector Extension (RVV), hindering the deployment of high-performance software on RISC-V platforms. Translating mature ecosystems such as ARM NEON into RVV is a practical solution. However, current translation tools are not aware of the target hardware during the translation process and therefore adopt conservative vector configurations to preserve correctness, resulting in substantial performance loss.To address this limitation, we present a human–machine collaborative translation framework that integrates user-provided hardware information into the translation process. Guided by this hardware information, the translator balances data-level parallelism with register-pressure constraints during the translation from ARM NEON to RVV, generating efficient RVV code. On a set of test cases, our method achieves up to 21.34× speedup over compiler auto-vectorization, up to 1.79× over hand-optimized native RVV kernels, and up to 2.71× over traditional translation methods.

The Linux kernel offers nearly 20,000 configuration options, making it highly customizable but also extremely challenging to manually optimize for performance. The diversity of operating environments and workloads further limits the effectiveness of static or expert-crafted configurations. This paper introduces KconfigTune, an automated tuning system that jointly optimizes Linux kernel configuration options and system-level parameters (e.g., procfs entries). We propose extconfig to expand the parameter search space and, for the first time, address configuration dependencies that arise during automated tuning. To ensure valid and bootable kernels, we design a dependency automatic fix tool and a GRUB-based mechanism. KconfigTune models the tuning process as a machine learning problem and applies Bayesian optimization with a random forest model to efficiently explore the vast and interdependent configuration space. Unlike prior approaches that focus solely on either kernel options or system parameters, KconfigTune achieves deeper integration by jointly tuning both compile-time and runtime behaviors. Experimental results show significant performance improvements, with gains of 18.62% and 19.92% over the default configuration in UnixBench and LEBench tests, respectively. Compared to the state-of-the-art, KconfigTune outperforms by 2.9% and 12.69% in these benchmarks. Ablation studies further confirm that these gains primarily stem from the combined tuning of kernel configurations and system-level parameters.

Tensor program optimization is essential for modern machine learning systems, but its search space is enormous. Existing auto-schedulers reduce measurement cost with learned cost models, yet they usually evaluate each candidate as a static code snapshot, ignoring the schedule trajectory that produced it. This makes them insensitive to action dependencies and vulnerable to superficial code variations. We propose a world-model-inspired evaluator that models schedule evaluation as action-conditioned latent dynamics over program states. Starting from the initial program, it rolls out scheduling actions in a continuous latent space with a lightweight transition model, avoiding expensive AST mutation and repeated code encoding. The final dynamic representation is combined with action and hardware features to rank candidates. Implemented in TVM AutoScheduler, our method improves representative-subgraph latency over Ansor by 1.37× on GPU and 1.54× on CPU under the same 64-trial budget. It also matches Ansor-10K within 2.2

Developing high-performance GPU kernels is critical for AI and scientific computing, but remains challenging due to its reliance on expert crafting and poor portability. While large language models (LLMs) offer promise for automation, both general-purpose and finetuned LLMs suffer from two fundamental and conflicting limitations: correctness and efficiency. The key reason is that existing LLM-based approaches directly generate the entire optimized low-level programs, requiring exploration of an extremely vast space encompassing both optimization policies and implementation codes. To address the challenge of exploring an intractable space, we propose Macro Thinking Micro Coding (MTMC), a hierarchical framework inspired by the staged optimization strategy of human experts. It decouples optimization strategy from implementation details, ensuring efficiency through high-level strategy and correctness through low-level implementation. Specifically, Macro Thinking employs reinforcement learning to guide lightweight LLMs in efficiently exploring and learning semantic optimization strategies that maximize hardware utilization. Micro Coding leverages general-purpose LLMs to incrementally implement the stepwise optimization proposals from Macro Thinking, avoiding full-kernel generation errors. Together, they effectively navigate the vast optimization space and intricate implementation details, enabling LLMs for high-performance GPU kernel generation. Comprehensive results on widely adopted benchmarks demonstrate the superior performance of MTMC on GPU kernel generation in both accuracy and running time. On KernelBench, MTMC achieves near 100% and 70% accuracy at Levels 1-2 and 3, over 50% than SOTA general-purpose and domain-finetuned LLMs, with up to 7.3× speedup over LLMs, and 2.2× over expert-optimized PyTorch Eager kernels. On the more challenging TritonBench, MTMC attains up to 59.64% accuracy and 34× speedup. All models and datasets will be made publicly available.

Approximate nearest neighbor search (ANNS) on CPUs is increasingly constrained by candidate-vector movement and decoding rather than peak arithmetic throughput. Although the RISC-V Vector Extension (RVV) provides vector-length-agnostic execution and LMUL-based register grouping, generic low-precision decoding still incurs conversion overhead, while irregular graph traversal generates scattered accesses that degrade cache locality and memory-level parallelism. We present RVANNS, an RVV-oriented ANNS engine that jointly optimizes vector representation and graph locality. Its Mixed-Precision Multi-Layer Index (MPMI) represents each vector with a dense 8-bit affine base and sparse FP16/FP32 residuals, fusing reconstruction with distance accumulation and aligning widening with LMUL-sized register groups. ROrder co-locates likely co-visited graph nodes and sorts remapped adjacency lists, transforming scattered payload probes into denser, predominantly forward-moving address streams. Integrated into Milvus, RVANNS achieves 3.39x and 4.94x speedups over scalar execution on real 128-bit and 256-bit RVV processors, respectively. Under controlled HNSW configurations, it improves throughput by 2.27–2.76x over RVV SIMD+FP32 and by 1.18–1.59x over the corresponding AVX-512 and SVE baselines. On Cohere10M, it further delivers 1.82–2.27x higher QPS/W than the evaluated GPU baselines.

Object detection models can achieve excellent detection performance with fully annotated instances. However, requiring complete annotations for every dataset is impractical due to high labor and time costs, as well as the inevitable occurrence of missing annotations. As a result, the absence of annotations can potentially provide misleading supervision and harm the training process. Recent methodologies have achieved remarkable effectiveness through the application of Co-Mining. However, the independence of each branch in Co-Mining cannot be guaranteed, overlooking valuable information during multi-perspective training. To address this issue, we introduce an Independence Guided Co-Training Model (IGCDet) that leverages Image Independence Decomposition to ensure the independence of each co-training branch. This model aims to capture diverse perspectives from images as extensively as possible, identifying missing annotations and incorporating them as positive supervision in the training process. Additionally, we propose the use of Joint-Confidence, derived from the combination of classification and regression, as pseudo-label scores, effectively mitigating issues associated with pseudo-label bias. Extensive experiments have verified the effectiveness of the proposed method.

The \(k\)-core model has garnered widespread adoption for preserving essential cohesive subgraphs owing to its linear-time computability, making it particularly suitable for hypergraph analysis. However, considering the continuously evolving characteristics of real-world hypergraphs, recent research efforts have focused on developing efficient algorithms that can maintain the core value of each vertex amid structural alterations. Despite these efforts, frequent insertions and deletions in dynamic hypergraphs continue to pose significant inefficiencies, primarily due to the increased traversal overhead incurred by hyperedge insertion algorithms. This exacerbates performance disparities between handling hyperedge insertions and deletions, underscoring the persistent challenge of effective \(k\)-core analysis in hypergraphs. To effectively address these challenges, we have gained key insights that enable us to define a specific order, termed the hypergraph \(k\)-order, which significantly reduces redundant vertex traversal and narrows down the search space during hyperedge insertions. Based on the proposed hypergraph \(k\)-order, we define two indices, the order index and the pivotal index, aimed at minimizing traversal costs and expediting the hyperedge insertion algorithm. Moreover, it is essential to recognize that the recomputation of the support degree (\(sd\)) for all vertices following each hyperedge deletion can significantly diminish the performance efficiency of deletion algorithms. To address this, we introduce an optimized approach that leverages the incremental maintenance of the support degree (\(sd\)) value to expedite the hyperedge deletion process. By leveraging these optimizations, we introduce a novel Order-based Traversal core Maintenance methodology, designated as OTM, which markedly enhances the efficiency of core maintenance in dynamic hypergraphs. Our comprehensive evaluation, which covers 12 real-world hypergraph datasets and a synthetic dataset, reveals that OTM achieves staggering speedup, outperforming the state-of-the-art approach with a 41,420\(\times\) speedup in the insertion algorithm and 8,284\(\times\) speedup in the deletion algorithm, underscoring its remarkable efficiency and effectiveness.