Sign languages are used by millions of individuals who are deaf or hard-of-hearing worldwide, yet learning them can be challenging due to limited access to qualified instructors and real-time feedback. Sign language learning is crucial for inclusion and accessible communication; however, existing e-learning solutions often fail to provide detailed or explainable feedback. To address this gap, we present a novel computer-assisted system that leverages computer vision and machine learning to teach sign languages. In our approach, students follow prerecorded videos of an instructor performing signs, and the system captures the students' own sign production through a conventional webcam. By analyzing both static handshapes and dynamic trajectories, our method provides detailed and explainable feedback at the level of keyframes, pointing out specific misalignments in hand, wrist, and motion path. This differs from many existing tools that offer only binary correct/incorrect evaluations without guidance for improvement. We implemented the system as a prototype with a full curriculum and conducted a pilot usability evaluation with 43 participants, ranging from beginners to skilled signers. Results indicate that users found the platform intuitive, the feedback mechanism informative, and the overall experience motivating for continued practice. The pilot assessed user perceptions of usability and feedback quality; learning outcomes were not measured in this initial evaluation. Notably, our design is language-agnostic, enabling broad application across different sign language communities. By offering instant fine-grained feedback, our system aims to improve self-guided learning of sign languages and promote greater accessibility, ultimately bridging communication gaps and fostering inclusion for the Deaf community.
As generative artificial intelligence (GenAI) rapidly evolves, K-12 education is introducing artificial intelligence (AI) courses whose interdisciplinary complexity exposes teachers' limited readiness. Research suggests that teachers can harness GenAI tools not only to enhance instruction but also to scaffold reflection that drives further teaching improvement. Therefore, this exploratory quasi-experimental study developed a customized GenAI agent system to support teachers' reflection on AI-course teaching and to enhance reflection self-efficacy, instructional design, and reflective thinking. A total of 60 in-service teachers were recruited and divided into two experimental groups, the self-reflection group (SRG) and the peer-reflection group (PRG), both supported by the customized GenAI agent system, and one control group (CG) with conventional technology-based reflection for a four-week AI course reflective practice experiment. The results revealed that the GenAI-agent-supported reflection approach could significantly promote teachers' self-reflection efficacy of two experimental groups compared with CG teachers. However, no significant differences between the SRG and PRG teachers' self-reflections efficacy could be found. Moreover, SRG and PRG teachers also outperformed the CG teachers on mutual-reflection self-efficacy. Furthermore, GenAI approach could significantly boost SRG and PRG teachers' instructional design reflection on methods and behavior; however, no significant differences were observed in instructional objectives or content among the three groups. To further explore the effects of GenAI agent support, epistemic network analysis was applied to examine the coded results of teachers' reflective journals. The findings indicated that SRG and PRG teachers demonstrated broader and higher order reflective thinking, integrating more dialogic and critical elements, whereas CG networks were predominantly descriptive. Overall, the study confirms that a customized GenAI agent can effectively deepen reflective thinking and practice, offering new insights into fostering teachers' professional development within K-12 AI education.
With the rapid development of large language models (LLMs), image-to-LaTeX conversion has shown great potential for applications such as automated grading and mathematical expression recognition in educational and assessment settings. However, real-world handwritten Science, Technology, Engineering, and Mathematics (STEM) manuscripts often contain strikeouts, overwrites, and other markings that disrupt handwritten content structure, reducing recognition accuracy and compilability. In this article, we present MathScrub, a single-pipeline multimodal framework that cleans and transcribes handwritten calculus expressions by grouping connected components with geometry-constrained clustering, detecting deletions via an early-fusion vision transformer, applying geometry-aware inpainting, and generating compilable LaTeX with a parameter-efficiently adapted open-source multimodal vision-language model (VLM). To support this task, we introduce Math-Strike, a dataset of handwritten calculus answers with real and synthetic strikeouts, component-level annotations, and aligned LaTeX. Experiments show that MathScrub outperforms open-source baselines on component detection and image-to-LaTeX tasks while remaining robust in real strikeout scenarios, providing a reliable preprocessing stage for LLM-based grading and formula recognition.
Generative large language models (LLMs) possess considerable potential for mitigating psychological distress among adolescents and advancing mental health education. Existing mental health-oriented large models primarily concentrate on acquiring empathetic capabilities, yet they still struggle to produce counseling responses that combine professional guidance with clear objectives for each dialogue turn. In addition, a significant gap exists between the linguistic expressions used by psychology experts and those of the general population, and this discrepancy directly affects users' willingness to express emotions. To tackle these challenges, this article introduces an iterative training framework that integrates dual perspectives from both counselors and adolescent clients. The framework first trains the model to identify core user needs and establish phased counseling goals based on user input. It then employs contrastive learning techniques to ensure the generated content maintains internal consistency in both linguistic style and counseling objectives. The study also provides an in-depth examination of the dynamic psychological evolution of adolescents after receiving multiple rounds of large model-based counseling. Experimental results demonstrate that the proposed framework achieves excellent performance across multiple key metrics, including credibility, empathy, respectfulness, and depth of guidance. Human-machine dialogue mental health LLMs are undergoing a paradigm shift from passive empathy to active goal-directed intervention.
Aesthetic perception, the cognitive process through which individuals interpret and evaluate the expressive and emotional qualities of visual art, is fundamental to students' creative and emotional development. Recent progress in artificial intelligence has enabled computational models to assist in aesthetic analysis by identifying patterns in visual composition and affective expression. However, such models often struggle to recognize abstract or context-dependent aesthetic dimensions, and improving these aspects through comprehensive annotation remains costly and time-consuming. This study presents a human-AI teaming framework designed to identify the aesthetic perception dimensions that AI models find most difficult to interpret and to allocate these dimensions to human experts for annotation. The framework employs a multiagent reinforcement learning mechanism, where each agent is assigned to a specific aesthetic dimension and learns a policy for determining whether expert annotation is required. Two complementary state representation strategies are introduced: a statistical representation that captures the model's predictive distribution across dimensions, and a graph-based attention module that models interdependencies among aesthetic attributes. A reward mechanism further guides agents to balance the improvement of model perception with the minimization of human annotation effort. Experiments conducted on two real-world datasets demonstrate that the proposed framework effectively identifies the challenging dimensions for AI models and strategically delegates them for human evaluation. This targeted collaboration significantly enhances annotation efficiency and model interpretability, providing a scalable approach for improving human-AI synergy in aesthetic perception analysis.
