Sketches are widely used in conceptual design to externalize early ideas and communicate intent. With the rise of generative AI, sketch-to-design workflows have advanced rapidly. However, sketches are limited for organizing component-level structure and intent: parts, functions, and relations are often implicit, making systematic design space exploration difficult. We present SketchConcept, a sketch-to-design system that enables multimodal exploration through sketching and language. It allows designers to sketch out the form, then use voice or text to articulate and refine component functions and structural organization. This enables designers to explore not only satisfying appearances, but also functional and structural alternatives that are essential for design. To support this workflow, SketchConcept introduces a function-to-visual mapping mechanism that connects visual components to functional properties for component-wise iteration. We demonstrate the system through a set of representative use cases and evaluate its efficacy and usability in a two-session user study.

Real-time handwriting interactions between tutors and students —where tutors observe individual problem-solving processes, provide personalized annotations, and adapt explanations based on students’ work—are fundamental to effective STEM tutoring. However, scaling such personalized handwriting-based tutoring remains challenging—human tutors cannot be available to every student on demand, and current online platforms often fail to recreate equivalent learning experiences. As an initial step toward tackling this challenge, we present AmIWrite, an LLM-powered AI tutoring system for mathematical problem-solving that provides real-time co-speech handwriting interactions on tablet devices, instantiated here as a case study in linear algebra. We conducted a within-subjects study (N = 40) comparing AmIWrite to a text-based AI tutor on two linear algebra topics. Our case study demonstrates how a multimodal AI tutor can preserve the pedagogical benefits of handwriting-based math tutoring and offer a potential path toward more scalable one-on-one STEM tutoring.

Generative AI (GenAI) has significantly advanced the ease and flexibility of image creation. However, it remains a challenge to precisely control spatial compositions, including object arrangement and scene conditions. To bridge this gap, we propose Canvas3D, an interactive system leveraging a 3D engine to enable precise spatial manipulation for image generation. Upon user prompt, Canvas3D automatically converts textual descriptions into interactive objects within a 3D engine-driven virtual canvas, empowering direct and precise spatial configuration. These user-defined arrangements generate explicit spatial constraints that guide generative models in accurately reflecting user intentions in the resulting images. We conducted a closed-ended comparative study between Canvas3D and a baseline system, and an open-ended, free-form study to assess overall system usability. The results indicate that Canvas3D outperforms the baseline on spatial control, interactivity, and overall user experience.

Generated Design SpaceIteration-1 Iteration-2 Iteration-3Figure 1: Exploring the design space of a desk using DesignFromX.The process begins with the user selecting a component from a reference product image-here, the legs of a wooden chair.The system identifies and suggests design features of the selected component.The user then composes this feature, in this case, the structural form, into a designated part of the desk.Based on the user-defined composition, a Generative AI model generates a new design space for the desk.In subsequent iterations, the user can incorporate additional design features from other reference products to further explore the design space while retaining the features of their previous selections.

Deciphering coordinated movements is integral to understanding the daily activities and interactions between the nervous system and muscles, especially in robot-assisted rehabilitation. This study proposes a novel temporal-guided adaptive graph learning (TAGL) network to recognize coordinated movements from functional near-infrared spectroscopy (fNIRS) data. The temporal-guided node construction module is designed to build graph nodes while considering spatiotemporal and causal dependencies. Given the brain network's affinity for learning asymmetric structures, an adaptive edge learning module is devised, integrating a multihead attention mechanism for the tailored acquisition of directional edge connections among nodes. The TAGL model undergoes evaluation on both a proprietary fNIRS dataset featuring eight circular finger movements and a public fNIRS dataset involving three distinct actions. Comparative experiments with state-of-the-art methods reveal its superior performance, showcasing its potential in deciphering coordinated movements effectively.

HyperGraph Convolutional Neural Networks (HGCNNs) have demonstrated their potential in modeling high-order relations preserved in graph-structured data. However, most existing convolution filters are localized and determined by the pre-defined initial hypergraph topology, neglecting to explore implicit and long-range relations in real-world data. In this paper, we propose the first learning-based method tailored for constructing adaptive hypergraph structure, termed HypERgrAph Laplacian aDaptor (HERALD), which serves as a generic plug-and-play module for improving the representational power of HGCNNs. Specifically, HERALD adaptively optimizes the adjacency relationship between vertices and hyperedges in an end-to-end manner and thus the task-aware hypergraph is learned. Furthermore, HERALD employs the self-attention mechanism to capture the non-local paired-nodes relation. Extensive experiments on various popular hypergraph datasets for node classification and graph classification tasks demonstrate that our approach obtains consistent and considerable performance enhancement, proving its effectiveness and generalization ability.

Recently, the teacher-student knowledge distillation framework has demonstrated its potential in training Graph Neural Networks (GNNs). However, due to the difficulty of training over-parameterized GNN models, one may not easily obtain a satisfactory teacher model for distillation. Furthermore, the inefficient training process of teacher-student knowledge distillation also impedes its applications in GNN models. In this paper, we propose the first teacher-free knowledge distillation method for GNNs, termed GNN Self-Distillation (GNN-SD), that serves as a drop-in replacement of the standard training process. The method is built upon the proposed neighborhood discrepancy rate (NDR), which quantifies the non-smoothness of the embedded graph in an efficient way. Based on this metric, we propose the adaptive discrepancy retaining (ADR) regularizer to empower the transferability of knowledge that maintains high neighborhood discrepancy across GNN layers. We also summarize a generic GNN-SD framework that could be exploited to induce other distillation strategies. Experiments further prove the effectiveness and generalization of our approach, as it brings: 1) state-of-the-art GNN distillation performance with less training cost, 2) consistent and considerable performance enhancement for various popular backbones.

Deep learning has obtained remarkable success for image restoration. However, most existing deep image restoration models are trained by minimizing the pixel-level reconstruction error between restored images and target images (ground truth), while neglecting the rich information from the intermediate feature layers, thus hindering the representational power of networks. To address this problem, we propose a Supervised Target Feature Transfer (STFT) framework to enhance the power of feature expression of the deep image restoration models. Specifically, we introduce a self-supervised antoencoder-based target feature extractor to extract compact feature representation of target images, which serves as supervision signals to train the deep backbone models at the same time. With such feature-level supervised information, deep backbone model can be enhanced by transfer learning of such target features. Moreover, we theoretically analyze our STFT training strategies and demonstrate that it imposes learnable prior information on the backbone restoration model. Extensive experiments demonstrate the effectiveness of our proposed framework compared with the state-of-the-art image restoration models.

Image downscaling has become a classical problem in image processing and has recently connected to image super-resolution (SR), which restores high-quality images from low-resolution ones generated by predetermined downscaling kernels (e.g., bicubic). However, most existing image downscaling methods are deterministic and lose information during the downscaling process, while rarely designing specific downscaling methods for image SR. In this paper, we propose a novel learning-based image downscaling method, Hamiltonian Rescaling Network (HRNet). The design of HRNet is based on the discretization of Hamiltonian System, a pair of iterative updating equations, which formulate a mechanism of iterative correction of the error caused by information missing during image or feature downscaling. Extensive experiments demonstrate the effectiveness of our proposed method in terms of both quantitative and qualitative results.