Latent Mechanisms of Language Control in Multilingual Language Models
研究解决多语言模型中不必要的代码转换问题,通过比较三种方法(ValSel、FreqSel和AnnSel)识别控制语言的潜在机制,实验表明这些方法有效。
研究解决多语言模型中不必要的代码转换问题,通过比较三种方法(ValSel、FreqSel和AnnSel)识别控制语言的潜在机制,实验表明这些方法有效。
针对口腔扫描图像分割中空间连续性和频率分布问题,FU-Mamba框架通过动态扫描和频域增强方法提高了分割精度。
This study addresses the limitations of static intrusion detection in vehicular networks, where adaptive adversaries and partial observability of the environment undermine defense efficacy. To tackle this challenge, the authors formulate security defense as a partially observable sequential attack-defense game and introduce a quantum-inspired, non-Bayesian amplitude-state belief representation mechanism. This approach effectively captures the defender’s uncertainty regarding attacker intent and integrates it into a Proximal Policy Optimization (PPO) framework to enable cost-aware dynamic defense decisions. Experimental results in a simulated environment demonstrate that the proposed method reduces cumulative average damage, damage variance, and attack success rate by 60.4%, 90.2%, and 50.0%, respectively, while improving system survival probability by 46.4% compared to baseline approaches, thereby significantly enhancing defensive effectiveness and robustness under partial observability.
This study addresses the limitations of traditional static defenses in vehicular networks, which fail to counter dynamic, multi-stage cyberattacks due to insufficient adaptability and high response latency. To overcome these challenges, the authors propose the QIRL framework, which uniquely integrates amplitude-phase quantum state encoding, a rotation gate-based exploration mechanism, and quantum interference-enhanced reward shaping into vehicular network security. The approach combines a lightweight deep Q-network with a cost-sensitive Markov decision process to model temporal attack dependencies and incorporates SMOTE oversampling to mitigate class imbalance. Evaluated on the CICIDS2017 and UNSW-NB15 datasets, QIRL achieves accuracy rates of 97.89% and 91.04%, F1 scores of 95.22% and 91.66%, and AUC-ROC values of 0.9945 and 0.9713, respectively, with per-sample inference latencies of only 32.5 and 45.7 microseconds—over 50 times faster than baseline methods.
This study addresses hope speech detection in English and German contexts, aiming to enhance the automatic identification of positive expressions in multilingual settings. For the first time, we apply RoBERTa (monolingual) and XLM-RoBERTa (English–German bilingual) to this task, leveraging their Transformer-based architectures and fine-tuning strategies for effective detection. Experimental results demonstrate that the proposed approach achieves an accuracy of 81.8% and a weighted F1-score of 0.818 on the English dataset, and 78.5% accuracy with a weighted F1-score of 0.786 in the multilingual English–German setting. These findings confirm the efficacy of pretrained language models in recognizing positively valenced text and highlight their strong potential for cross-lingual transfer in affective computing tasks.
研究解决多语言模型中不必要的代码转换问题,通过比较三种方法(ValSel、FreqSel和AnnSel)识别控制语言的潜在机制,实验表明这些方法有效。
针对口腔扫描图像分割中空间连续性和频率分布问题,FU-Mamba框架通过动态扫描和频域增强方法提高了分割精度。
This study addresses the limitations of static intrusion detection in vehicular networks, where adaptive adversaries and partial observability of the environment undermine defense efficacy. To tackle this challenge, the authors formulate security defense as a partially observable sequential attack-defense game and introduce a quantum-inspired, non-Bayesian amplitude-state belief representation mechanism. This approach effectively captures the defender’s uncertainty regarding attacker intent and integrates it into a Proximal Policy Optimization (PPO) framework to enable cost-aware dynamic defense decisions. Experimental results in a simulated environment demonstrate that the proposed method reduces cumulative average damage, damage variance, and attack success rate by 60.4%, 90.2%, and 50.0%, respectively, while improving system survival probability by 46.4% compared to baseline approaches, thereby significantly enhancing defensive effectiveness and robustness under partial observability.
This study addresses the limitations of traditional static defenses in vehicular networks, which fail to counter dynamic, multi-stage cyberattacks due to insufficient adaptability and high response latency. To overcome these challenges, the authors propose the QIRL framework, which uniquely integrates amplitude-phase quantum state encoding, a rotation gate-based exploration mechanism, and quantum interference-enhanced reward shaping into vehicular network security. The approach combines a lightweight deep Q-network with a cost-sensitive Markov decision process to model temporal attack dependencies and incorporates SMOTE oversampling to mitigate class imbalance. Evaluated on the CICIDS2017 and UNSW-NB15 datasets, QIRL achieves accuracy rates of 97.89% and 91.04%, F1 scores of 95.22% and 91.66%, and AUC-ROC values of 0.9945 and 0.9713, respectively, with per-sample inference latencies of only 32.5 and 45.7 microseconds—over 50 times faster than baseline methods.
This study addresses hope speech detection in English and German contexts, aiming to enhance the automatic identification of positive expressions in multilingual settings. For the first time, we apply RoBERTa (monolingual) and XLM-RoBERTa (English–German bilingual) to this task, leveraging their Transformer-based architectures and fine-tuning strategies for effective detection. Experimental results demonstrate that the proposed approach achieves an accuracy of 81.8% and a weighted F1-score of 0.818 on the English dataset, and 78.5% accuracy with a weighted F1-score of 0.786 in the multilingual English–German setting. These findings confirm the efficacy of pretrained language models in recognizing positively valenced text and highlight their strong potential for cross-lingual transfer in affective computing tasks.