This paper presents a theoretical analysis of a fourth-order exponential time differencing Runge–Kutta rational approximation scheme for the Allen–Cahn equation based on real and distinct poles, aiming to establish a theoretical framework for the method combining exponential time differencing with rational approximation. Starting from the scalar linearized equation, we rigorously prove that the scheme is L-stable and that its matrix norm is stable with three types of boundary conditions, revealing that the L-acceptable property of the RDP rational approximation is the key to ensuring stability. Regarding the error analysis, we establish that the fully discrete scheme attains fourth-order temporal convergence, which theoretically confirms that the scheme achieves fourth-order accuracy. Numerical examples in both 2D and 3D verify the accuracy and effectiveness of the method.
Visual Question Answering (VQA) models are highly vulnerable to multimodal shortcuts induced by answer imbalance and language priors. Existing debiasing methods mainly rely on bias branches, answer-frequency statistics, or sample difficulty, but they cannot explicitly determine whether a prediction is produced by genuine multimodal reasoning or by shortcut reliance. To address this limitation, we propose GSP-VQA, a generative shortcut probing framework for robust VQA. We design a Generative Question-Image Bias (GQIB) module that performs generated missing-modality probing to estimate instance-level shortcut reliance without altering the original training distribution. Based on this signal, we introduce a shortcut-aware Multi-Level Margin (MLM) mechanism that combines frequency-based, difficulty-based, and shortcut probing margins in the hyperspherical space. Experiments show that GSP-VQA achieves a competitive performance of 62.22% accuracy on VQA-CP v2 and a competitive 35.16% on the VQA-CE Counterexamples split, demonstrating effective robustness against shortcut-driven predictions. Our code is available at https://anonymous.4open.science/r/GSP-VQA-R-9C38/.
This quasi-experimental study integrates embodied cognition, constructionism, and culturally sustaining teaching to address the three-dimensional needs of cognitive development, practical competence, and cultural understanding, operationalized through the Child-Centered Guided Scaffolding AI Framework (CGS-AI). Based on this framework, we developed an AI-supported Intelligent Learning Companion called “Little Bean Companion,” which guides elementary students in learning about traditional cultural architecture and building miniature physical architectural models. The system integrates Vygotsky’s dynamic scaffolding support, Constructionism, and child-centered interaction design principles. A total of 38 fourth-grade students participated in this quasi-experimental study designed to quantitatively evaluate the effectiveness of our developed system. Based on existing class arrangements, students were assigned to an experimental group (n = 20) and a control group (n = 18). The experimental group used the Intelligent Learning Companion for architecture-themed learning over a one-month period totaling 7 class hours, while the control group received traditional teacher-led instruction. Results showed that the experimental group performed significantly better than the control group in the post-test of the Course Knowledge Test (Adj. M = 88.32 vs. 80.59, p<.001), with a particularly pronounced difference in multiple-choice questions requiring knowledge transfer and application. In the Architectural Model Evaluation, the experimental group not only had a higher median score (92 vs. 86) but also exhibited a “high-score clustering” phenomenon. Structural analysis of learning motivation revealed that while all four dimensions showed improvement, the effect size for mastery-goal orientation exceeded that of performance-goal orientation, perceived enjoyment, and self-efficacy. These students were able to externalize abstract knowledge principles into concrete physical operations when constructing miniature physical architectural models, and through heuristic dialogue with the Intelligent Learning Companion, they established deep connections between cultural concepts and engineering practice. The study demonstrates that when AI transforms from an “information tutor” to a “constructionist partner,” it can effectively promote children’s embodied learning, cultural understanding, and learning motivation.
