The rapid evolution of generative artificial intelligence (GenAI) is transforming science education by facilitating innovative pedagogical paradigms while raising substantial concerns about scholarly integrity. One particularly pressing issue is the growing risk of student use of GenAI tools to outsource assessment tasks, potentially compromising authentic learning and evaluations. Addressing these challenges requires reflection on existing assessment practices and features. This position paper advances a conceptual framework for science assessment through the lens of multimodality and interactivity. Multimodality emphasizes the use of diverse, organized semiotic resources for meaning making, while interactivity characterizes assessment environments where outcomes are shaped by students' actions. With the two dimensions, our multimodal interactive framework classifies assessments into four categories, with varying degrees of modality and interactivity. We argue that tasks with higher modality and interactivity can potentially overcome the concerns of GenAI on academic integrity. To further articulate this point, we provide concrete assessment examples for each category and explain how the prompt and response affordances in each assessment category help gauge students' understandings of key science constructs and identify tasks that are resistant or susceptible to AI‐based outsourcing. We conclude by discussing how the framework serves as a meaningful analytical tool for educational researchers and practitioners.

This study investigated how one’s problem-solving style impacts his/her problem-solving performance in technology-rich environments. Drawing upon experiential learning theory, we extracted two behavioral indicators (i.e., planning duration for problem solving and human–computer interaction frequency) to model problem-solving styles in technology-rich environments. We employed an existing data set in which 7516 participants responded to 14 technology-based tasks of the Programme for the International Assessment of Adult Competencies (PIAAC) 2012. Clustering analyses revealed three problem-solving styles: Acting indicates a preference for active explorations; Reflecting represents a tendency to observe; and Shirking shows an inclination toward scarce tryouts and few observations. Explanatory item response modeling analyses disclosed that individuals with the Acting style outperformed those with the Reflecting or the Shirking style, and this superiority persisted across tasks with different difficulties.

Learning outcome modeling is a technical underpinning for the successful evaluation of learners' learning outcomes through computer-based assessments. In recent years, collaborative filtering approaches have gained popularity as a technique to model learners' item responses. However, how to model the temporal dependencies between item responses using a collaborative filtering approach for learning outcome modeling is still under investigation. Leveraging the advantages of deep learning, this study proposes a novel deep learning-based collaborative filtering approach for learning outcome modeling. Unlike conventional collaborative filtering approaches, the proposed model, utilizing recurrent neural networks, is capable of sequentially predicting learners' future learning outcomes based on their history learning records. Moreover, the proposed model has the capacity to discover item-skill associations from the scratch without expert input based on attentive modeling. The experimental results demonstrate that the proposed model outperforms a popular deep-learning approach and can be successfully used to discover item-skill associations for both real-world and synthetic datasets.

Background: Process data captured by computer‐based assessments provide valuable insight into respondents' cognitive processes during problem‐solving tasks. Although previous studies have utilized process data to analyse behavioural patterns or strategies in problem‐solving tasks, the connection between latent cognitive states and their theoretical interpretation in problem solving remains unclear.Objectives: This research aims to investigate the connections between similar hidden response states and unfold respondents' transition paths in problem‐solving processes. Analysing process data from the 2012 United States Programme for the International Assessment of Adult Competencies (PIAAC), this study seeks to discern patterns in problem solving among participants.Methods: The hidden Markov model was first used to uncover the hidden states based on a sequence of observed actions. Next, Gaussian graphical network analysis was employed to analyse the relationships between hidden response states.Results and Conclusions: Results indicated that correct responders had simpler, clearer state relationships, while incorrect responders displayed more complex connections. Respondents who solved the tasks correctly had clearer thoughts about the problem‐solving process, whereas incorrect respondents struggled to understand the problem and failed to figure out solutions. Cognitive state changes during problem solving also varied between groups. The correct groups showed cohesive, logical transitions, in contrast to the emerged isolated, erratic patterns of the incorrect groups.