CoRemix: Supporting Informal Learning in Scratch Community With Visual Graph and Generative AI

TL;DR

CoRemix addresses the lack of structured guidance in informal Scratch learning by combining a visual graph of events and computing concepts with a RA-LLM-based conversational agent and visual-textual scaffolding. The two-phase design guides learners first to decompose projects and then to remix by adding nodes and edges, supported by a community knowledge base. An empirical study with 16 elementary learners shows improved project understanding, computation concept learning, and remixing creativity compared to a Scratch baseline. The work demonstrates the viability and challenges of integrating retrieval-augmented generative AI into informal coding communities, and outlines future directions for scalable, safe, and inclusive AI-guided learning tools.

Abstract

Online programming communities provide a space for novices to engage with computing concepts, allowing them to learn and develop computing skills using user-generated projects. However, the lack of structured guidance in the informal learning environment often makes it difficult for novices to experience progressively challenging learning opportunities. Learners frequently struggle with understanding key project events and relations, grasping computing concepts, and remixing practices. This study introduces CoRemix, a generative AI-powered learning system that provides a visual graph to present key events and relations for project understanding. We propose a visual-textual scaffolding to help learners construct the visual graph and support remixing practice. Our user study demonstrates that CoRemix, compared to the baseline, effectively helps learners break down complex projects, enhances computing concept learning, and improves their experience with community resources for learning and remixing.

Figures (7)

• Figure 1: Overview of CoRemix. In the understanding phase, learners decompose project events, create event nodes and edges, and incorporate computing concept (CC) nodes to understand key computing concepts, guided by generative AI scaffolding. In the co-remixing phase, learners engage in remixing, adding new nodes and relationships to enhance their projects.
• Figure 2: Event and CC Nodes: In CoRemix, the child can use two types of nodes to build visual graph: Event nodes (Character, Behavior, and Result) and CC nodes (Condition, Boolean, Loop, and Variable).
• Figure 3: When using CoRemix, learners first create event and computing concept nodes from the node area (a.4) and build edges on the visual graph (a.1). Learners can also add corresponding event descriptions on the canvas. If they encounter difficulties while constructing the graph, they can get scaffolding support from the conversational agent equipped with generative AI (a.2). Then, learners can click "new event" (a.3) to create a new canvas and start remixing the project by generating new event nodes with images (b.1) and building relationships with existing nodes (b.2).
• Figure 4: Visual and Textual Scaffolding in the Constructive Loop: When a learner poses a question, CoRemix provides visual scaffolding to support them. If the learner clicks "Got it," CoRemix generates follow-up questions to further assess their understanding. However, if the Response Check of CoRemix is vague or the learner chooses "I don't know," CoRemix offers additional textual scaffolding until the learner achieves a clear comprehension.
• Figure 5: Workflow of community resource-driven retrieval-augmented generation and project analysis.