Exact quantum splitting and the structure of finite algebras
本文提出了一种无条件的精确量子算法，用于解决有限代数结构中的因子分解问题，通过量子振幅放大技术确保每次测试成功。
0 citationsRead paper
Selected work
Representative Papers
Deployable Per-Instance Multi-Layer Activation Steering for Large Language Models
This work addresses the limitation of existing activation steering methods, which rely on fixed global injection layers and struggle to adapt to the optimal intervention location for diverse inputs. The authors propose a deployable, instance-wise dynamic multi-layer steering approach that predicts, based on prompt embeddings, the optimal steering layer and direction for each input and modulates steering depth via an adaptive gating mechanism. This method is the first to enable unsupervised, label-free, instance-specific multi-layer steering and introduces a “direction-over-magnitude” principle to explain and mitigate behavioral flipping and fluency collapse. Experiments on two 8B open-source models across six personality traits demonstrate that the approach closely approximates oracle performance, consistently matches or exceeds the non-steered baseline on average, and effectively avoids output degradation commonly induced by high-layer steering.
0 citationsRead paper
Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
This work proposes an efficient shallow physics-informed neural network (PINN) framework—employing only two hidden layers—to address the high computational cost and slow convergence of conventional deep PINNs in solving forward and inverse problems of nonlinear partial differential equations. By modeling the network as a nonlinear system and leveraging the Levenberg-Marquardt (LM) second-order optimization algorithm, the method constructs an exact Jacobian matrix using analytical derivatives of the input variables. Evaluated on benchmark problems including Burgers’, Schrödinger, Allen–Cahn, and three-dimensional Bratu equations, the approach consistently outperforms BFGS in terms of convergence speed, solution accuracy, and final loss, thereby challenging the prevailing notion that PINNs must be deep to be effective.
0 citationsRead paper
Preserving Fairness and Safety in Quantized LLMs Through Critical Weight Protection
This work investigates the adverse effects of quantization on fairness and safety in multilingual large language models, particularly under dynamic quantization and non-English contexts. While quantization reduces computational costs, it disproportionately exacerbates bias amplification and degrades safety alignment across diverse languages. The study systematically evaluates the differential impacts of static versus dynamic quantization on multilingual fairness and safety, and introduces a retraining-free critical weight protection mechanism. By identifying and preserving key parameters essential to ethical and secure behavior, the method significantly mitigates fairness and safety degradation without compromising inference efficiency. Extensive experiments across English, French, Dutch, Spanish, Turkish, Korean, and Arabic demonstrate that the proposed approach effectively balances model efficiency with trustworthiness in multilingual settings.
0 citationsRead paper
Stuttering-Aware Automatic Speech Recognition for Indonesian Language
This study addresses the significant performance degradation of automatic speech recognition (ASR) systems when processing stuttered speech, a challenge particularly acute in low-resource languages like Indonesian due to the scarcity of authentic stuttered speech data. To tackle this issue, the work proposes the first stutter-aware ASR system for Indonesian, introducing a synthetic data augmentation framework that does not require real stuttered recordings. The approach generates stuttered text—featuring repetitions and prolongations—through rule-based transformations and large language models, then synthesizes corresponding stuttered speech using text-to-speech systems. This synthetic data is used to fine-tune a pretrained Whisper model. Experimental results demonstrate that the method substantially reduces word error rates on stuttered speech while preserving recognition accuracy on fluent speech, thereby enhancing the inclusivity of ASR systems for low-resource languages.
0 citationsRead paper
Recent publications
Latest Papers
Exact quantum splitting and the structure of finite algebras
本文提出了一种无条件的精确量子算法，用于解决有限代数结构中的因子分解问题，通过量子振幅放大技术确保每次测试成功。
0 citationsRead paper
Deployable Per-Instance Multi-Layer Activation Steering for Large Language Models
This work addresses the limitation of existing activation steering methods, which rely on fixed global injection layers and struggle to adapt to the optimal intervention location for diverse inputs. The authors propose a deployable, instance-wise dynamic multi-layer steering approach that predicts, based on prompt embeddings, the optimal steering layer and direction for each input and modulates steering depth via an adaptive gating mechanism. This method is the first to enable unsupervised, label-free, instance-specific multi-layer steering and introduces a “direction-over-magnitude” principle to explain and mitigate behavioral flipping and fluency collapse. Experiments on two 8B open-source models across six personality traits demonstrate that the approach closely approximates oracle performance, consistently matches or exceeds the non-steered baseline on average, and effectively avoids output degradation commonly induced by high-layer steering.
0 citationsRead paper
Do physics-informed neural networks (PINNs) need to be deep? Shallow PINNs using the Levenberg-Marquardt algorithm
This work proposes an efficient shallow physics-informed neural network (PINN) framework—employing only two hidden layers—to address the high computational cost and slow convergence of conventional deep PINNs in solving forward and inverse problems of nonlinear partial differential equations. By modeling the network as a nonlinear system and leveraging the Levenberg-Marquardt (LM) second-order optimization algorithm, the method constructs an exact Jacobian matrix using analytical derivatives of the input variables. Evaluated on benchmark problems including Burgers’, Schrödinger, Allen–Cahn, and three-dimensional Bratu equations, the approach consistently outperforms BFGS in terms of convergence speed, solution accuracy, and final loss, thereby challenging the prevailing notion that PINNs must be deep to be effective.
0 citationsRead paper
Preserving Fairness and Safety in Quantized LLMs Through Critical Weight Protection
This work investigates the adverse effects of quantization on fairness and safety in multilingual large language models, particularly under dynamic quantization and non-English contexts. While quantization reduces computational costs, it disproportionately exacerbates bias amplification and degrades safety alignment across diverse languages. The study systematically evaluates the differential impacts of static versus dynamic quantization on multilingual fairness and safety, and introduces a retraining-free critical weight protection mechanism. By identifying and preserving key parameters essential to ethical and secure behavior, the method significantly mitigates fairness and safety degradation without compromising inference efficiency. Extensive experiments across English, French, Dutch, Spanish, Turkish, Korean, and Arabic demonstrate that the proposed approach effectively balances model efficiency with trustworthiness in multilingual settings.
0 citationsRead paper
Stuttering-Aware Automatic Speech Recognition for Indonesian Language
This study addresses the significant performance degradation of automatic speech recognition (ASR) systems when processing stuttered speech, a challenge particularly acute in low-resource languages like Indonesian due to the scarcity of authentic stuttered speech data. To tackle this issue, the work proposes the first stutter-aware ASR system for Indonesian, introducing a synthetic data augmentation framework that does not require real stuttered recordings. The approach generates stuttered text—featuring repetitions and prolongations—through rule-based transformations and large language models, then synthesizes corresponding stuttered speech using text-to-speech systems. This synthetic data is used to fine-tune a pretrained Whisper model. Experimental results demonstrate that the method substantially reduces word error rates on stuttered speech while preserving recognition accuracy on fluent speech, thereby enhancing the inclusivity of ASR systems for low-resource languages.
0 citationsRead paper