arXiv subjects

Weixiong Zheng

Publications and source records attributed to Weixiong Zheng.

6 recordsLinked to original sources

arXiv2025

CLASS: Enhancing Cross-Modal Text-Molecule Retrieval Performance and Training Efficiency

Cross-modal text-molecule retrieval task bridges molecule structures and natural language descriptions. Existing methods predominantly focus on aligning text modality and molecule modality, yet they overlook adaptively adjusting the learning states at different training stages and enhancing training efficiency. To tackle these challenges, this paper proposes a Curriculum Learning-bAsed croSS-modal text-molecule training framework (CLASS), which can be integrated with any backbone to yield promising performance improvement. Specifically, we quantify the sample difficulty considering both text modality and molecule modality, and design a sample scheduler to introduce training samples via an easy-to-difficult paradigm as the training advances, remarkably reducing the scale of training samples at the early stage of training and improving training efficiency. Moreover, we introduce adaptive intensity learning to increase the training intensity as the training progresses, which adaptively controls the learning intensity across all curriculum stages. Experimental results on the ChEBI-20 dataset demonstrate that our proposed method gains superior performance, simultaneously achieving prominent time savings.

cs.CL↗

arXiv2024

Jailbreaking? One Step Is Enough!

Large language models (LLMs) excel in various tasks but remain vulnerable to jailbreak attacks, where adversaries manipulate prompts to generate harmful outputs. Examining jailbreak prompts helps uncover the shortcomings of LLMs. However, current jailbreak methods and the target model's defenses are engaged in an independent and adversarial process, resulting in the need for frequent attack iterations and redesigning attacks for different models. To address these gaps, we propose a Reverse Embedded Defense Attack (REDA) mechanism that disguises the attack intention as the "defense". intention against harmful content. Specifically, REDA starts from the target response, guiding the model to embed harmful content within its defensive measures, thereby relegating harmful content to a secondary role and making the model believe it is performing a defensive task. The attacking model considers that it is guiding the target model to deal with harmful content, while the target model thinks it is performing a defensive task, creating an illusion of cooperation between the two. Additionally, to enhance the model's confidence and guidance in "defensive" intentions, we adopt in-context learning (ICL) with a small number of attack examples and construct a corresponding dataset of attack examples. Extensive evaluations demonstrate that the REDA method enables cross-model attacks without the need to redesign attack strategies for different models, enables successful jailbreak in one iteration, and outperforms existing methods on both open-source and closed-source models.

cs.CL↗

arXiv2017

An Accurate Globally Conservative Subdomain Discontinuous Least-squares Scheme for Solving Neutron Transport Problems

In this work, we present a subdomain discontinuous least-squares (SDLS) scheme for neutronics problems. Least-squares (LS) methods are known to be inaccurate for problems with sharp total-cross section interfaces. In addition, the least-squares scheme is known not to be globally conservative in heterogeneous problems. In problems where global conservation is important, e.g. k-eigenvalue problems, conservative treatment must be applied. We, in this study, proposed a SDLS method that retains global conservation. Such a method resembles the LS formulation in each subdomain without material interface and differs from LS in that an additional least-squares interface term appears for each interface. Scalar flux is continuous in each subdomain with continuous finite element method (CFEM) while discontinuous on interfaces for every pair of contiguous subdomains. SDLS numerical results are compared with those obtained from other numerical methods with test problems having material interfaces. High accuracy of scalar flux in fixed-source problems and $k_\mathrm{eff}$ in eigenvalue problems are demonstrated.

physics.comp-ph↗

arXiv2017

Accurate Least-Squares P$_N$ Scaling based on Problem Optical Thickness for solving Neutron Transport Problems

In this paper, we present an accurate and robust scaling operator based on material optical thickness (OT) for the least-squares spherical harmonics (LSP$_N$) method for solving neutron transport problems. LSP$_N$ without proper scaling is known to be erroneous in highly scattering medium, if the optical thickness of the material is large. A previously presented scaling developed by Manteuffel, et al.\ does improve the accuracy of LSP$_N$, in problems where the material is optically thick. With the method, however, essentially no scaling is applied in optically thin materials, which can lead to an erroneous solution with presence of highly scattering medium. Another scaling approach, called the reciprocal-removal (RR) scaled LSP$_N$, which is equivalent to the self-adjoint angular flux (SAAF) equation, has numerical issues in highly-scattering materials due to a singular weighting. We propose a scaling based on optical thickness that improves the solution in optically thick media while avoiding the singularity in the SAAF formulation.

physics.comp-ph↗

arXiv2016

Effective Non-oscillatory Regularized L$_1$ Finite Elements for Particle Transport Simulations

In this work, we present a novel regularized L$_1$ (RL$_1$) finite element spatial discretization scheme for radiation transport problems. We review the recently developed least-squares finite element method in nuclear applications. We then derive an L$_1$ finite element by minimizing the L$_1$ norm of the transport residual. To ensure the stability on incident boundary, we newly develop a consistent L$_1$ boundary condition (BC). The numerical tests demonstrate such a method effectively prevents the oscillations which would occur to least-squares finite element when discontinuity exists such as void and absorber. Further, the RL$_1$ method is accurate in problems with scattering.

physics.comp-ph↗

arXiv2016

Moment Closures Based on Minimizing the Residual of the P$_N$ Angular Expansion in Radiation Transport

In this work we present two new closures for the spherical harmonics (PN) method in slab geometry transport problems. Our approach begins with an analysis of the squared-residual of the transport equation where we show that the standard truncation and diffusive closures do not minimize the residual of the PN expansion. Based on this analysis we derive two models, a moment-limited diffusive MLDN closure and a transient PN (TPN) closure that attempt to address shortcomings of common closures. The form of these closures is similar to flux-limiters for diffusion with the addition of a time-derivative in the definition of the closure. Numerical results on a pulsed plane source problem, the Gordian knot of slab-geometry transport problems, indicate that our new closure outperforms existing linear closures. Additionally, on a deep penetration problem we demonstrate that the TPN closure does not suffer from the artificial shocks that can arise in the MN entropy-based closure. Finally, results for Reed's problem demonstrate that the TPN solution is as accurate as the PN+3 solution. We further extend the TPN closure to 2D Cartesian geometry. The line source test problem demonstrates the model effectively damps oscillations and negative densities

physics.comp-ph↗