Large Language Models (LLMs) have demonstrated remarkable capabilities in capturing complex conceptual representations from textual data for a wide range of real-world applications. However, in Intelligent Fault Diagnosis (IFD), leveraging sensor data such as vibration signals is essential but remains a challenge due to the modality gap between time series and LLMs’ inputs. Existing efforts to bridge this gap often treat LLMs merely as classifiers, overlooking their potential for understanding and reasoning over vibration-based data. In this paper, we propose a novel LLM-based fault diagnosis framework (FD-LLM) that aligns vibration signals with LLMs by encoding the signals into textual representations. FD-LLM introduces a classification-oriented approach, which formulates fault diagnosis as a multi-class classification task for benchmarking LLMs’ performance, and a context-aware spectrum language modeling approach that enables explainable, reasoning-driven fault analysis. We evaluate four open-source LLMs using FD-LLM across multiple datasets and noise conditions, assessing their validity, adaptability, and robustness. The results demonstrate that models such as LLaMA models achieve robust diagnostic performance, strong zero-shot adaptability across operating conditions, and effective generalization in cross-dataset scenarios with few-shot learning. The results further indicate that explainable fault diagnosis can be achieved in LLMs.

Time series foundation models (TSFMs) have demonstrated increasing capabilities due to their extensive pretraining on large volumes of diverse time series data. Consequently, the quality of time series data is crucial to TSFM performance, rendering an accurate and efficient data valuation of time series for TSFMs indispensable. However, traditional data valuation methods, such as influence functions, face severe computational bottlenecks due to their poor scalability with growing TSFM model sizes and often fail to preserve temporal dependencies. In this paper, we propose LTSV, a Lightweight Time Series Valuation on TSFMS via in-context finetuning. Grounded in the theoretical evidence that in-context finetuning approximates the influence function, LTSV estimates a sample's contribution by measuring the change in context loss after in-context finetuning, leveraging the strong generalization capabilities of TSFMs to produce robust and transferable data valuations. To capture temporal dependencies, we introduce temporal block aggregation, which integrates per-block influence scores across overlapping time windows. Experiments across multiple time series datasets and models demonstrate that LTSV consistently provides reliable and strong valuation performance, while maintaining manageable computational requirements. Our results suggest that in-context finetuning on time series foundation models provides a practical and effective bridge between data attribution and model generalization in time series learning.

Time series is a pervasive data type across various application domains, rendering the reasonable solving of diverse time series tasks a long-standing goal. Recent advances in large language models (LLMs), especially their reasoning abilities unlocked through reinforcement learning (RL), have opened new opportunities for tackling tasks with long Chain-of-Thought (CoT) reasoning. However, leveraging LLM reasoning for time series remains infancy, hindered by the absence of carefully curated time series CoT data for training, limited data efficiency caused by underexplored data scheduling, and the lack of RL algorithms tailored for exploiting such time series CoT data. In this paper, we introduce VeriTime, a framework that tailors LLMs for time series reasoning through data synthesis, data scheduling, and RL training. First, we propose a data synthesis pipeline that constructs a time series–text multimodal dataset with process-verifiable annotations. Second, we design a data scheduling mechanism that arranges training samples according to a principled hierarchy of difficulty and task taxonomy. Third, we develop a two-stage reinforcement finetuning featuring fine-grained, multi-objective rewards that leverages verifiable process-level CoT data. Extensive experiments show that VeriTime substantially boosts LLM performance across diverse time series reasoning tasks. Notably, it enables compact 3B–4B models to achieve reasoning capabilities on par with or exceeding those of larger proprietary LLMs.

Ensuring the integrity of Multivariate Time Series (MTS) data is essential for critical tasks such as health monitoring, predictive maintenance, and real-time decision-making. In real-world scenarios, missing data are unavoidable due to power outages, network interruptions, and other technical issues. These disruptions create “missing gaps”, during which data from all channels is absent for a certain period. Additionally, time series data are often collected at varying frequencies, creating “missing intervals” in low-frequency variables compared to high-frequency ones. Existing methods primarily focus on scenarios with randomly distributed missing values within the tabular data while under-exploring more realistic scenarios involving missing gaps and intervals. Specifically, they assume missing data occur randomly and predict it based on adjacent observations, which under-represent the complex co-evolution of channel information and long-term temporal dependencies. In this study, we propose Restoring Missing gaps and intervals with global Channel-Temporal Consistency and local Contextual Coherence (C4T-RM), which ensures global consistency of the imputed data by modeling long-term dependency buried in non-missing channel and temporal dimensions. An inverse mapping scheme with adversarial training is adopted to ensure imputed reality by capturing the local contextual coherence of the observed non-missing portion. Experiments on several benchmark datasets show that C4T-RM outperforms other deep learning-based imputation methods, achieving State-Of-The-Art (SOTA) performance. The source code is available at https://github.com/huricha1/C4T-RM.

High-quality time series (TS) data are essential for ensuring TS model performance, rendering research on rating TS data quality indispensable. Existing methods have shown promising rating accuracy within individual domains, primarily by extending data quality rating techniques such as influence functions and Shapley values to account for temporal characteristics. However, they neglect the fact that real-world TS data can span vastly different domains and exhibit distinct properties, hampering the accurate and efficient rating of diverse TS data. In this paper, we propose TSRating, a novel and unified framework for rating the quality of time series data crawled from diverse domains. TSRating leverages LLMs' inherent ample knowledge, acquired during their extensive pretraining, to comprehend and discern quality differences in diverse TS data. We verify this by devising a series of prompts to elicit quality comparisons from LLMs for pairs of TS samples. We then fit a dedicated rating model, termed TSRater, to convert the LLMs' judgments into efficient quality predictions by inferring future TS samples through TSRater's inference. To ensure cross-domain adaptability, we develop a meta-learning scheme to train TSRater on quality comparisons collected from nine distinct domains. To improve training efficiency, we employ signSGD for inner-loop updates, thus circumventing the demanding computation of hypergradients. Extensive experimental results on eleven benchmark datasets across three time series tasks, each using both conventional TS models and TS foundation models, demonstrate that TSRating outperforms baselines in terms of estimation accuracy, efficiency, and domain adaptability.

Fine-tuning is an essential and pervasive functionality for applying large language models (LLMs) to downstream tasks. However, it has the potential to substantially degrade safety alignment, e.g., by greatly increasing susceptibility to jailbreak attacks, even when the fine-tuning data is entirely harmless. Despite garnering growing attention in defense efforts during the fine-tuning stage, existing methods struggle with a persistent safety-utility dilemma: emphasizing safety compromises task performance, whereas prioritizing utility typically requires deep fine-tuning that inevitably leads to steep safety declination. In this work, we address this dilemma by shedding new light on the geometric interaction between safety- and utility-oriented gradients in safety-aligned LLMs. Through systematic empirical analysis, we uncover three key insights: (I) safety gradients lie in a low-rank subspace, while utility gradients span a broader high-dimensional space; (II) these subspaces are often negatively correlated, causing directional conflicts during fine-tuning; and (III) the dominant safety direction can be efficiently estimated from a single sample. Building upon these novel insights, we propose safety-preserving fine-tuning (SPF), a lightweight approach that explicitly removes gradient components conflicting with the low-rank safety subspace. Theoretically, we show that SPF guarantees utility convergence while bounding safety drift. Empirically, SPF consistently maintains downstream task performance and recovers nearly all pre-trained safety alignment, even under adversarial fine-tuning scenarios. Furthermore, SPF exhibits robust resistance to both deep fine-tuning and dynamic jailbreak attacks. Together, our findings provide new mechanistic understanding and practical guidance toward always-aligned LLM fine-tuning.

Modern configurable systems offer customization via intricate configuration spaces, yet such flexibility introduces pervasive configuration-related issues such as misconfigurations and latent softwarebugs. Existing diagnosability supports focus on post-failure analysis of software behavior to identify configuration issues, but none of these approaches look into whether the software clue sufficient failure information for diagnosis. To fill in the blank, we propose the idea of configuration logging to enhance existing logging practices at the source code level. We develop ConfLogger, the first tool that unifies configuration-aware static taint analysis with LLM-based log generation to enhance software configuration diagnosability. Specifically, our method 1) identifies configuration-sensitive code segments by tracing configuration-related data flow in the whole project, and 2) generates diagnostic log statements by analyzing configuration code contexts. Evaluation results on eight popular software systems demonstrate the effectiveness of ConfLogger to enhance configuration diagnosability. Specifically, ConfLogger-enhanced logs successfully aid a log-based misconfiguration diagnosis tool to achieve 100

Large Language Models (LLMs) have showcased remarkable capabilities across various domains. Accompanying the evolving capabilities and expanding deployment scenarios of LLMs, their deployment challenges escalate due to their sheer scale and the advanced yet complex activation designs prevalent in notable model series, such as Llama, Gemma, Mistral. These challenges have become particularly pronounced in resource-constrained deployment scenarios, where mitigating inference bottlenecks is imperative. Among various recent efforts, activation approximation has emerged as a promising avenue for pursuing inference efficiency, sometimes considered indispensable in applications such as private inference. Despite achieving substantial speedups with minimal impact on utility, even appearing sound and practical for real-world deployment, the safety implications of activation approximations remain unclear. In this work, we fill this critical gap in LLM safety by conducting the first systematic safety evaluation of activation approximations. Our safety vetting spans seven state-of-the-art techniques across three popular categories (activation polynomialization, activation sparsification, and activation quantization), revealing consistent safety degradation across ten safety-aligned LLMs. To overcome the hurdle of devising a unified defense accounting for diverse activation approximation methods, we perform an in-depth analysis of their shared error patterns and uncover three key findings. We propose QuadA, a novel safety enhancement method tailored to mitigate the safety compromises introduced by activation approximations. Extensive experiments and ablation studies corroborate QuadA's effectiveness in enhancing the safety capabilities of LLMs after activation approximations.

Graph neural networks have shown great potential in these years, primarily owing to their remarkable capacity for neighborhood aggregation. However, compared with their great success on homophilic networks where the connected nodes have similar features or labels, traditional graph neural networks encounter significant performance degradation on heterophilic networks where dissimilar nodes tend to form connections. Although some heterophilic graph neural networks have been proposed to tackle this problem, most of them have not fully taken advantage of neighborhood label information. In particular, though nodes tend to connect with heterophilic neighbors in heterophilic networks, the neighborhood label distributions of nodes in the same class tend to be similar and vice versa, which helps in improving the distinguishability of node representations. In this work, by leveraging the soft label of nodes, we propose an end-to-end label-guided graph neural network to aggregate the neighborhood information from different classes individually. To further improve the generality, we incorporate an attention mechanism to determine the aggregation weights for each class adaptively. Comprehensive experiments demonstrate the superiority of the proposed model on both heterophilic and homophilic datasets.

Buildings are major energy consumers, and efficient management is crucial for reducing emissions, yet the lack of standardised data formats complicates this process. This paper presents a method for automating the classification of time-series building data, using a segmentation approach to enhance feature extraction and applying hierarchical differential features to capture temporal dynamics. The multi-label classification problem is reformulated into a unified single-label classification task by encoding multiple labels into a discrete 91-class label space. The model, trained with CatBoost and evaluated using 5-fold cross-validation, secured a top ranking in the Brick by Brick 2024 competition, demonstrating its effectiveness in promoting energy-efficient building operations.

Given their mission-critical nature, SCADA systems in Cyber-Physical Systems (CPS) are highly susceptible to cyber-attacks. While abnormal system states can be detected through analyzing streaming data, accurately identifying the precise location of these attacks remains challenging due to the compensatory mechanisms inherent in CPS. In industrial control, the development of high-performance controllers that require minimal data and incur low technical debt is especially appealing. Recently, foundation models have demonstrated their effectiveness in solving various problems with few or no demonstrations, leveraging the rich prior knowledge obtained through pre-training on Internet-scale datasets. Most of these models are based on the Transformer architecture, originally designed for natural language processing (NLP) tasks. However, Transformer-based models struggle to fully capture the spatial correlations embedded in industrial software systems. Previous approaches using Graph Convolution Networks (GCNs) have largely focused on learning static graph structures of CPSs, overlooking the potential variations in dynamic, short-term spatial correlations during the mechanical processes. This paper presents a novel approach to capturing hidden spatial-temporal correlations in large-scale industrial data, offering a fresh perspective on building industrial pre-trained foundation models. Specifically, we introduce a cyber-attack detection and localization framework that incorporates Multi-scale Graph Structure Learning (MGSL) within a multi-stream sequence reconstruction architecture. The approach combines Long Short-Term Memory (LSTM) for temporal information extraction with GCNs to capture spatial correlations across multiple scales, including long-term static and short-term dynamic relationships. Additionally, a self-learning mechanism is employed to facilitate the discovery of long-term static graph structures, while feature similarities are used to identify short-term dynamic structures. The proposed MGSL framework is validated using three publicly available real-world datasets: Secure Water Treatment (SWaT A4&A5) and the BATtle of Attack Detection Algorithm (BATADAL). Experimental results demonstrate the effectiveness of MGSL, outperforming state-of-the-art methods by successfully detecting cyber-attacks, pinpointing impacted areas, and analyzing the target points of detected attacks.

Log parsing, as a process of extracting structured information from semi-structured raw log data, is a crucial step in log analysis workflows. Rule-based parsing methods often overlook the rich semantic information contained in logs. Recently, LLM-based methods face three major challenges: 1) limited understanding of domain-specific logs due to lack of professional domain knowledge; 2) significant parsing costs and potential data privacy risks associated with commercial LLMs like ChatGPT; and 3) the Group Accuracy (GA) is highly susceptible to individual anomalous data. To address these challenges, we propose LogExpertSolver, a multi-agent framework for domain-specialized log parsing. This framework effectively reduces manual intervention in log parsing through inter-agent collaboration mechanisms. LogExpertSolver employs locally deployed medium-scale language models to construct multiple agents with diverse domain expertise. By decomposing complex log parsing tasks into a series of fine-grained subtasks, which are handled by corresponding expert agents, the framework achieves efficient log parsing. Through evaluation on the LogHub2.0 dataset, LogExpertSolver achieves an average Parsing Accuracy (PA) of 0.915, surpassing state-of-the-art parsers (LibreLog and LILAC) by 6.2% and 7.3% respectively.

The scarcity of high-quality public log datasets has become a critical bottleneck in advancing log-based anomaly detection techniques. Current datasets exhibit three fundamental limitations: (1) incomplete event coverage, (2) artificial patterns introduced by static analysis-based generation frameworks, and (3) insufficient semantic awareness. To address these challenges, we present AnomalyGen, the first automated log synthesis framework specifically designed for anomaly detection. Our framework introduces a novel four-phase architecture that integrates enhanced program analysis with Chain-of-Thought reasoning (CoT reasoning), enabling iterative log generation and anomaly annotation without requiring physical system execution. Evaluations on Hadoop and HDFS distributed systems demonstrate that AnomalyGen achieves substantially broader log event coverage (38-95 times improvement over existing datasets) while producing more operationally realistic log sequences compared to static analysis-based approaches. When augmenting benchmark datasets with synthesized logs, we observe maximum F1-score improvements of 3.7 1.8 work not only establishes a high-quality benchmarking resource for automated log analysis but also pioneers a new paradigm for applying large language models (LLMs) in software engineering workflows.

Code editing plays a vital role in software engineering, requiring developers to adjust existing code according to natural language instructions while keeping functionality intact and avoiding unnecessary modifications. However, commit-based datasets commonly used for this task are often noisy, lack diversity, and fail to reflect the style of real-world edit instructions. To address this, we introduce CanItEdit, an open-source pipeline that leverages multiple LLMs to synthesize realistic code-edit triplets. The pipeline produces both concise "lazy" instructions and more detailed "descriptive" ones, and applies filtering based on diffs and topics to guarantee data quality and variety. Using this process, we construct OCEDataFT, a curated dataset of 20K samples. Fine-tuning three advanced base models on OCEDataFT leads to significant performance boosts on the CanItEdit benchmark, with relative pass@1 improvements ranging from 4.50

The reasoning capabilities of large language models (LLMs) have significantly advanced their performance by enabling in-depth understanding of diverse tasks. With growing interest in applying LLMs to the time series domain, this has proven nontrivial, as evidenced by the limited efficacy of straightforwardly adapting text-domain reasoning techniques. Although recent work has shown promise in several time series tasks, further leveraging advancements in LLM reasoning remains under-explored for time series classification (TSC) tasks, despite their prevalence and significance in many real-world applications. In this paper, we propose ReasonTSC, a novel framework designed to effectively leverage LLM reasoning for time series classification through both a multi-turn reasoning and a fused decision-making strategy tailored to TSC. Rather than straightforwardly applying existing reasoning techniques or relying solely on LLMs' built-in reasoning capabilities, ReasonTSC first steers the model to think over the essential characteristics of time series data. Next, it integrates predictions and confidence scores from plug-in classifiers, e.g., domain-specific time series models, as in-context examples. Finally, ReasonTSC guides the LLM through a structured reasoning process: it evaluates the initial assessment, backtracks to consider alternative hypotheses, and compares their merits before arriving at a final classification. Extensive experiments and systematic ablation studies demonstrate that ReasonTSC consistently outperforms both existing time series reasoning baselines and plug-in models, and is even capable of identifying and correcting plug-in models' false predictions.

Prognostic and Health Management (PHM) are crucial ways to avoid unnecessary maintenance for Cyber-Physical Systems (CPS) and improve system reliability. Predicting the Remaining Useful Life (RUL) is one of the most challenging tasks for PHM. Existing methods require prior knowledge about the system, contrived assumptions, or temporal mining to model the life cycles of machine equipment/devices, resulting in diminished accuracy and limited applicability in real-world scenarios. This paper proposes a Bi-directional Adversarial network with Covariate Encoding for machine Remaining Useful Life (BACE-RUL) prediction, which only adopts sensor measurements from the current life cycle to predict RUL rather than relying on previous consecutive cycle recordings. The current sensor measurements of mechanical devices are encoded to a conditional space to better understand the implicit inner mechanical status. The predictor is trained as a conditional generative network with the encoded sensor measurements as its conditions. Various experiments on several real-world datasets, including the turbofan aircraft engine dataset and the dataset collected from degradation experiments of Li-Ion battery cells, show that the proposed model is a general framework and outperforms state-of-the-art methods.

Dataset license compliance is a critical yet complex aspect of developing commercial AI products, particularly with the increasing use of publicly available datasets. Ambiguities in dataset licenses pose significant legal risks, making it challenging even for software IP lawyers to accurately interpret rights and obligations. In this paper, we introduce LicenseGPT, a fine-tuned foundation model (FM) specifically designed for dataset license compliance analysis. We first evaluate existing legal FMs (i.e., FMs specialized in understanding and processing legal texts) and find that the best-performing model achieves a Prediction Agreement (PA) of only 43.75%. LicenseGPT, fine-tuned on a curated dataset of 500 licenses annotated by legal experts, significantly improves PA to 64.30%, outperforming both legal and general-purpose FMs. Through an A/B test and user study with software IP lawyers, we demonstrate that LicenseGPT reduces analysis time by 94.44%, from 108 seconds to 6 seconds per license, without compromising accuracy. Software IP lawyers perceive LicenseGPT as a valuable supplementary tool that enhances efficiency while acknowledging the need for human oversight in complex cases. Our work underscores the potential of specialized AI tools in legal practice and offers a publicly available resource for practitioners and researchers. Moreover, LicenseGPT has the potential to assist AI software developers in managing preliminary license checks before involving legal counsel, helping to avoid costly late-stage rework and ensuring AI software compliance.

Time series (TS) data are ubiquitous across various application areas, rendering time series forecasting (TSF) a fundamental task. With the astounding advances in large language models (LLMs), a variety of methods have been developed to adapt LLMs for time series forecasting. Despite unlocking the potential of LLMs in comprehending TS data, existing methods are inherently constrained by their shallow integration of TS information, wherein LLMs typically access TS representations at shallow layers, primarily at the input layer. This causes the influence of TS representations to progressively fade in deeper layers and eventually leads to ineffective adaptation between textual embeddings and TS representations. In this paper, we propose the Multi-layer Steerable Embedding Fusion (MSEF), a novel framework that enables LLMs to directly access time series patterns at all depths, thereby mitigating the progressive loss of TS information in deeper layers. Specifically, MSEF leverages off-the-shelf time series foundation models to extract semantically rich embeddings, which are fused with intermediate text representations across LLM layers via layer-specific steering vectors. These steering vectors are designed to continuously optimize the alignment between time series and textual modalities and facilitate a layer-specific adaptation mechanism that ensures efficient few-shot learning capabilities. Experimental results on seven benchmarks demonstrate significant performance improvements by MSEF compared with baselines, with an average reduction of 31.8% in terms of MSE. The code is available at https://github.com/One1sAll/MSEF.

This research explores strategies for using one-way anonymous Q&A software to enhance class participation and teaching effectiveness of undergraduate students. The study finds that undergraduate students, influenced by cultural factors, ed-ucational systems, and upbringing environments, tend to display introverted psychological traits, resulting in insufficient class participation. One-way anonymous Q&A software significantly improves student participation enthusiasm through mechanisms such as breaking psychological barriers, promoting deep thinking, enhancing classroom interaction, and providing diverse feedback. Case analysis shows that by introducing the proposed anonymous Q&A software to teaching activities, students asked questions more frequently, and the quality of questions also improved accordingly with a wider adoption of the software. Besides, it also shows that the learning interests of students have increased after using the software. The research suggests that future studies should deepen technological innovation, pro-mote teaching model reform, and strengthen interdisciplinary integration to better meet the needs of a modern educational environment in universities.

Limited by the scale and diversity of time series data, the neural networks trained on time series data often overfit and show unsatisfacotry performances. In comparison, large language models (LLMs) recently exhibit impressive generalization in diverse fields. Although massive LLM based approaches are proposed for time series tasks, these methods require to load the whole LLM in both training and reference. This high computational demands limit practical applications in resource-constrained settings, like edge-computing and IoT devices. To address this issue, we propose Knowledge Pruning (KP), a novel paradigm for time series learning in this paper. For a specific downstream task, we argue that the world knowledge learned by LLMs is much redundant and only the related knowledge termed as "pertinent knowledge" is useful. Unlike other methods, our KP targets to prune the redundant knowledge and only distill the pertinent knowledge into the target model. This reduces model size and computational costs significantly. Additionally, different from existing LLM based approaches, our KP does not require to load the LLM in the process of training and testing, further easing computational burdens. With our proposed KP, a lightweight network can effectively learn the pertinent knowledge, achieving satisfactory performances with a low computation cost. To verify the effectiveness of our KP, two fundamental tasks on edge-computing devices are investigated in our experiments, where eight diverse environments or benchmarks with different networks are used to verify the generalization of our KP. Through experiments, our KP demonstrates effective learning of pertinent knowledge, achieving notable performance improvements in regression (19.7 average) and classification (up to 13.7 results.