Yu Guo, Shenghao Ye, Shuangwu Chen, Zijian Wen, Tao Zhang, Bai Qirui, Dong Jin, Yunpeng Hou, Huasen He, Jianyang, Xiaobin Tan. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shenghao Ye, Yu Guo, Dong Jin, Yuxiang Wang, Yikai Shen, Yunpeng Hou, Shuangwu Chen, Jianyang, Xiaofeng Jiang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Logical anomalies occur when a product's assembly violates prescribed logical rules, which widely exist in industrial assembly and packaging processes. Due to the difficulty in comprehending such complex logical relationships, a paucity of research has focused on the industrial logical anomaly diagnosis (LAD). Recently, large language models (LLMs) have demonstrated strong semantic understanding and zero-shot reasoning capabilities, making them a promising tool for LAD. However, directly applying LLMs to LAD still face two critical challenges: 1) the scarcity of abnormal samples in real-world settings, and 2) the propensity of LLMs to generate hallucinated or unreliable diagnostic conclusions. To address these challenges, we propose LogiDiag, a novel diagnostic reasoning method for LAD, to pinpoint where and why a product fails to comply with the logical rules, thereby elevating product quality and reducing remedial intervention cost. We design a visual descriptor that identifies product component attributes even in out-of-distribution abnormal images and organizes them into component descriptions for LLMs' comprehension. To mitigate hallucinations, we devise a planner to guide LLMs in diagnostic reasoning through rule orchestration, tool allocation, and chain-of-diagnosis generation. Experimental results on multiple benchmark datasets validate the competitive performance of LogiDiag.

Fault detection (FD) is essential for wafer quality control in semiconductor manufacturing (SM), as it can identify abnormal wafers in the early stages. However, frequent chamber discrepancies leads to distribution shifts in the data from different chambers, which causes performance degradation of the existing model. In this paper, we conceive a robust cross-chamber fault detection method in SM, which to the best of our knowledge is the first work that employs domain generalization to address the issue of cross-chamber fault detection in SM. Our basic idea is to map the samples from the source chambers and target chambers to the same hypersphere space, making normal samples cluster around a specific feature center while faulty samples stay away from it. Due to the scarcity of faulty samples, we propose a one-class classification-based fault detection method relying on normal samples to establish classification boundaries. To generalize the model to unseen target chambers, we design a meta-learning-based one-class domain generalization approach. We also devise a strategy to enhance distribution alignment within hypersphere, making the classification boundaries of various chambers be close to each other. The evaluations on real-world data collected from a wet process equipment in SM verify the high robustness of the model to chamber discrepancies with limited faulty samples.

Website Fingerprinting (WF) attacks have posed a serious threat to the anonymity of the onion router (Tor) communication system, as attackers can passively pry into the encrypted traffic and infer the website visited by users. To defend against WF, recent studies focus on adversarial perturbations. However, most of them suffer from a high bandwidth overhead and a low defense performance. To address this problem, our basic idea is to generate perturbation only on the sensitive regions, which can effectively mask the website’s fingerprint, thus misleading the WF attack models and reducing the bandwidth overhead. In this paper, we formulate a joint optimization problem of perturbation position and magnitude by confining the perturbations within sensitive regions, which is rarely considered in the literature. We propose a robust low-overhead WF defense method based on reinforcement learning (RL), named RLpatch. RLpatch identifies the common sensitive regions of various surrogate models and adjusts perturbation according to the query result from a query WF model. It further employs the positional frequency of perturbations to generate a common perturbation paradigm for different traces of a same website. Experimental results show that RLpatch achieves higher defense performance, lower bandwidth overhead and better robustness against adversarial training compared to the state-of-the-art methods.

Anomaly diagnosis relying on system logs to record runtime events is essential for improving the service quality of distributed systems and reducing economic losses. However, most existing log-based anomaly detection approaches depend on the assumption of the fixed quantitative or sequential patterns of a normal log event sequence. This assumption is challenged in the context of practical distributed and parallel systems due to the dynamic pattern of the log sequence, log data noise as well as concurrency of multiple anomalies. Against these challenges, this paper aims to perform the robust log-based anomaly diagnosis by capturing the event context information of the log event graph, called LogGraph, instead of straightforwardly employing the fixed quantitative or sequential patterns of the log records, thus reducing its sensitivity to the log flaws and the concurrency of multiple anomalies. Specifically, in order to handle multiple anomalies concurrency, LogGraph invokes the association rule to decouple log sequences. It further reinterprets a log record sequence into a log event graph modeled by event semantic embedding and event adjacency matrix. An attention-based Gated Graph Neural Network (GGNN) model is developed to capture the semantic information of the log graph, which enables the fine-grained and robust anomaly identification of the proposed scheme. We use real log data sets collected from Hadoop systems and network switches to verify the effectiveness of the proposed LogGraph in log data scenarios that contain noise and multiple anomalies concurrency problems. The experimental results show that the proposed LogGraph achieves high performance and strong robustness in anomaly diagnosis.

Self-piercing riveting (SPR) is widely used for joining lightweight and dissimilar materials in automotive body manufacturing, the quality of which directly affects the safety of vehicles. However, there is still no reliable method that can be used for SPR quality control without destructive test and manual intervention. This article presents an online nondestructive SPR defect detection method based on deep learning. By learning the temporal dependencies of punch force varying with rivet displacement under different joint combinations, the proposed method can provide real-time defect alarms and avoid the enormous cost of joint dissection. We develop an SPR parameter selection mechanism to rule out the irrelevant parameters, which enhances the learning performance. For the problem of model overfitting caused by the savage imbalance of SPR data, we design a conditional generative adversarial network based data generation model. In order to accommodate the difference in defect patterns between factory and laboratory, we devise a transfer learning based model migration method, which substantially reduces the amount of labeled factory data for model training. The evaluations on real SPR data collected from two car assembly lines of Audi and NIO verify that the proposed method achieves a high detection accuracy and a low missing rate in SPR defect detection.