Students' interest pathways are fashioned not only from existing educational resources and opportunities, but importantly also by extending beyond them. In this study, we conceptualise these extensions as productive deviations and engage in a comparative analysis of various productive deviations we have identified in our ethnographic study of seven 5th and 6th grade FUSE Studios - an alternative learning infrastructure for schools. Our analysis shows that the deviations can vary from short term excursions to semester-long projects and can also become the focus of other students' interests.

Artificial intelligence (AI) is a growing field in both global job markets and educational spaces. This non-experimental quantitative study aims to explore how the educational program A Fresh Squeeze on Data affects students’ self-efficacy and career choices and whether gender will differentiate the learning outcomes. Under the social cognitive career theory framework, this study designs questionnaires as data collection instrument. The results suggest that the program significantly improves students’ comfortability with AI-related subjects but not for career interest or other measurements in self-efficacy. Unexpectedly, the program’s effect is not divided by gender. Nevertheless, this study opens up conversations about assisting students from underrepresented backgrounds to envision success in AI courses and career pathways through an activity-driven curriculum. The paper also informs educators and researchers to devise culturally responsive pedagogy in teaching AI that empowers young girls before they develop a gendered career view. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

While speech-enabled teachable agents have some advantages over typing-based ones, they are vulnerable to errors stemming from misrecognition by automatic speech recognition (ASR). These errors may propagate, resulting in unexpected changes in the flow of conversation. We analyzed how such changes are linked with learning gains and learners’ rapport with the agents. Our results show they are not related to learning gains or rapport, regardless of the types of responses the agents should have returned given the correct input from learners without ASR errors. We also discuss the implications for optimal error-recovery policies for teachable agents that can be drawn from these findings. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Productive learning of algebra is supported when students reflect on multiple strategies, compare them and discuss the rationale behind and relative merits of particular strategies. Comparison and Discussion of Multiple Strategies (CDMS) is an instructional approach designed to support these processes in math classrooms. In the current study, 16 Algebra I teachers received professional development and supplemental materials to support CDMS when teaching a unit on linear equation solving and 475 of their students completed assessments of their linear equation solving knowledge before and after the unit. Thirteen Algebra I teachers and their 359 students were the business-as-usual control group. CDMS increased how often teachers engaged their students in comparison of multiple strategies, sustained small group work, and sustained mathematical discussions. Students in CDMS classrooms also had higher knowledge of linear equations on the posttest, particularly procedural flexibility, even after controlling for pretest knowledge and school demographic differences. Thus, encouraging teachers to regularly compare and discuss multiple strategies increases students' algebra learning. Findings highlight the need to expand theories of algebra learning to include attention to procedural flexibility, illustrate an instructional theory and method to promote broader learning about algebra, and provide evidence for effective instructional practices.

Although designed to prepare students for future coursework or to fulfill basic degree requirements, introductory math courses often serve as barriers to student success. In two double-blind randomized field experiments, we tested the efficacy of a utility-value intervention on improving community college students' perceived math relevance and achievement in introductory math courses. Building upon prior research, we examined whether the intervention particularly benefited first-generation and racially marginalized students. Study 1 (N = 696) was conducted before the COVID-19 pandemic and within in-person classrooms, whereas Study 2 (N = 1,318) was conducted during the pandemic and within virtual learning environments. Across Studies 1 and 2, students in the utility-value condition benefited more in terms of their perceived relevance compared to their peers in the control condition. Additionally, in both studies, math relevance mediated the effects of the intervention on math grades. In Study 2, with a larger sample, the positive effect of the intervention on math relevance was more pronounced for first-generation students. Our findings imply that community colleges could significantly improve students' academic experiences by investing in motivation-enhancing activities such as utility-value interventions in introductory math courses. This strategy could especially help first-generation students' academic achievement and retention rates.

Intelligent Tutoring Systems (ITSs) leverage AI to adapt to individual students, and employ pedagogical policies to decide what instructional action to take next. A number of researchers applied Reinforcement Learning (RL) and Deep RL (DRL) to induce effective pedagogical policies. Most prior work, however, has been developed independently for a specific ITS and cannot directly be applied to another. In this work, we propose a Multi-Task Learning framework that combines Deep BIsimulation Metrics and DRL, named MTL-BIM, to induce a unified pedagogical policy for two different ITSs across different domains: logic and probability. Based on empirical classroom results, our unified RL policy performed significantly better than the expert-crafted policies and independently induced DQN policies on both ITSs. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Hyper-realistic human video generation (also referred to as “deepfake”) is a recent development in artificial intelligence that may be applied to create pedagogical agents (PAs) for education purposes. Traditionally, PA research has been using 2D or 3D virtual figures to examine how their design may affect student learning or perceptions. In this study, we employed the latest human video generation technology to create PAs and investigated how students’ perceived stereotypes of competent teachers would affect their preferences of a PA. It was found that students prefer to learn Japanese with Asian looking PAs over Caucasian or Black agents, supporting the hypothesis that real-world stereotype ideas can persist in the virtual world. Findings of the study offer references for how to better design PAs with generated human videos to boost learning motivation and enhance student perceptions. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

Errors are inevitable in most learning contexts, but under the right conditions, they can be beneficial for learning. Prior research indicates that generating and learning from errors can promote retention of knowledge, higher-level learning, and self-regulation. The present review proposes an integrated theoretical model to explain two major phases of learning from self-generated errors: the Generating Errors (GE) phase, which contributes to learning via semantically related prior knowledge activation, and the Detecting and Correcting Errors (DCE) phase, which contributes to learning via self-explanation when processing and comparing one's responses with provided reference information to promote high-quality internal feedback. Our model identifies general design principles that support each phase based on prior empirical research. We conclude by identifying research gaps and future directions regarding specific design features of the GE and DCE phases and the role of students' emotion, motivation, and individual differences in learning from errors.

Simulations of human learning can be used as computational models for evaluating theories of learning. They can also be taught interactively to author intelligent tutoring systems. Prior simulated learner systems have learned inductively from worked examples and correctness feedback. This work introduces a mechanism where simulated learners can also learn from natural language. Using a neural grammar parser with additional symbolic processing steps, we simulate the production of loose interpretations of verbal instructions. These interpretations can be combined with worked examples to resolve the ambiguities of either form of instruction alone. We find that our system has practical benefits over an alternative method using github Copilot and slightly better accuracy. © 2023, The Author(s), under exclusive license to Springer Nature Switzerland AG.

[No abstract available]