cs.CYMar 13, 2026

Thinking in Graphs with CoMAP: A Shared Visual Workspace for Designing Project-Based Learning

Authors: Ruijia Li, Bo Jiang

Organizations: School of Computer Science and Technology, East China Normal University · Shanghai Institute of Artificial Intelligence for Education, East China Normal University

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Designing project-based learning (PBL) demands managing highly interdependent components, a task that both traditional linear tools and purely conversational AI struggle with. Traditional tools fail to capture the non-linear nature of creative design, while conversational systems lack the persistent, shared context necessary for reflective collaboration. Grounded in theories of distributed cognition, we introduce CoMAP, a system that embodies a graph-based collaboration paradigm. By providing a shared visual workspace with dual-modality AI support, CoMAP transforms the human-AI relationship from a prompt-and-response loop into a transparent and equitable partnership. Our study with 30 educators shows CoMAP significantly improves teachers' design expression, divergent thinking, and iterative practice compared to a dialogue-only baseline. These findings demonstrate how a nonlinear, artifact-centric approach can foster trust, reduce cognitive load, and \textcolor{fix}{support} educators to take control of their creative process. Our contributions are available at: https://comap2025.github.io/.

Figures & tables

Figure 1. This figure conceptually contrasts two distinct paradigms for human-AI collaboration. (a) Represents the traditional paradigm, where the design process relies on linear tools and isolated conversational AI. This approach fails to capture the nonlinear nature of creative design. (b) Shows the CoMAP paradigm, which utilizes a shared visual workspace as a persistent cognitive artifact. This allows for the externalization of non-linear thoughts, offering users a form of human-AI collaboration that helps to mitigate the limitations of the traditional paradigm. Figure 1 is a conceptual diagram that contrasts two distinct paradigms for human-AI collaboration: the traditional, linear approach and a more modern, structured approach. Panel (a), titled "Restricted Thinking," shows a confused person facing a linear, chat-based interface with fragmented ideas, represented by text bubbles in a single flow, highlighting the limitations of traditional tools. In contrast, Panel (b), titled "Supported Thinking," depicts the same person with a confident expression and a light-bulb over their head, collaborating with two AI assistants on a shared digital canvas. This canvas shows a rich network of interconnected nodes and ideas, illustrating a non-linear, flexible workspace that allows for structured thinking and the externalization of complex thoughts, offering a helpful alternative to the constraints of the traditional paradigm.
Figure 2. The two-stage formative study design. In Stage 1, we conducted semi-structured interviews centered on participants’ past experiences with AI-assisted PBL design to identify their core challenges. In Stage 2, a static CoMAP prototype was used as a design probe to elicit specific design goals that address these challenges. Figure 2 is a two-stage diagram illustrating a formative study design. Stage 1: Semi-structured Interview is shown on the top, where a person is engaged in an interview on a tablet screen, discussing a list of materials related to "AI-assisted PBL design" and "Self-Esteem Bucket Protection," leading to the identification of three main challenges: C1 (Path Dependence and Limited Ideation Space), C2 (Difficulty in Maintaining a Holistic View), and C3 (Limitations in Contextual AI Support). Below, Stage 2: Think-aloud Study shows a person interacting with a desktop computer displaying a "CoMAP Static Prototype as Design Probe," which is a mind-map-like interface used to design curriculum. This stage is designed to elicit specific design goals, which are listed as DG1 (Support for Multi-Dimensional Ideation), DG2 (Holistic Visualization of Curriculum Structure), and DG3 (Contextual, Actionable, and Shared-Awareness AI Support).
Figure 3. Overview of the CoMAP system’s interactive components and the flow of distributed cognition. This architecture embodies a distributed cognitive framework, where the cognitive load is shared among three key components: the human designer, the Structured Graph Canvas serving as an external cognitive workspace, and a team of specialized AI agents. The human interacts with the canvas to externalize their ideas, while the agents operate on this shared workspace, providing targeted assistance that augments the human’s design process. Figure 3 presents a comprehensive overview of the CoMAP system's interactive components and the flow of distributed cognition. In the center, a human designer and a main AI assistant are collaborating on a digital canvas that is populated with interconnected nodes, representing a shared cognitive workspace. The human's thought process is represented by a bubble saying "I think……," while the AI assistant's suggestion is shown in a bubble saying "I suggest……," illustrating a collaborative dialogue. The cognitive load is shared among the human designer, the Structured Graph Canvas (serving as an external cognitive workspace), and three specialized AI agents on the right side. These agents are the Global Agent, which offers "structure advice" for macro-level insights; the Refine Agent, which provides "local focus" for precise editing; and the Split Agent, which helps in "deconstructing complex lesson components." The image highlights how the system supports the human's creative process by externalizing ideas and providing targeted, specialized assistance from multiple AI agents.
Figure 4. The visual grammar and interactive states of CoMAP nodes. (a) illustrates the six basic node types: Learner Analysis ( a.1 ), Learning Objectives ( a.2 ), Instructional Strategies ( a.3 ), Learning Activities ( a.4 ), Project Resources ( a.5 ), and Assessment & Evaluation ( a.6 ). Each type is distinguished by a unique icon and color to facilitate at-a-glance comprehension of the design structure. (b) demonstrates the different display modes and interactive states of an Objective node: Default ( b.1 ), Hovered ( b.2 ) to reveal options, Expanded ( b.3 ) for detailed content editing, Left Clicked ( b.4 ) exposing AI-powered refinement and splitting tools, and Connected ( b.5 ) to another node (e.g., Activity) to illustrate pedagogical relationships. These states support fluid, non-linear design exploration. Figure 4 provides a detailed breakdown of the visual grammar and interactive states of CoMAP nodes. Panel (a), labeled "Basic Node Type," illustrates the six distinct node types used for designing curriculum: Learner Analysis (a.1), Learning Objectives (a.2), Instructional Strategies (a.3), Learning Activities (a.4), Project Resources (a.5), and Assessment & Evaluation (a.6). Each node is represented by a unique icon and color to facilitate easy at-a-glance identification of the design structure. Panel (b), labeled "Different Display Mode," showcases the interactive states of an "Objective" node. It starts with the Default view (b.1), then shows the Hovered state (b.2), which reveals an option to "Show Details." The Expanded state (b.3) allows for detailed content editing and includes a description of the objective, such as "Master the ability to retrieve scientific data…" The Left Clicked state (b.4) exposes options like "Edit," "Delete," and AI-powered functions such as "Split" and "Refine," indicated by small icons. Finally, the Connected state (b.5) shows how a node can be linked to another, such as an "Objective" to an "Activity," to visually represent pedagogical relationships.
Node TypeCategory / LevelExample Title & Description
LearnerPrimary SchoolPrimary School Learners : High curiosity, short attention span, learn through play and concrete operations…
Junior HighJunior High Learners : Developing abstract thinking, sensitive to peer relationships and self-identity…
High SchoolHigh School Learners : Advanced abstract thinking, strong logical analysis and independent thought…
AdultAdult Learners : Goal-oriented, rich prior experience, value practicality and relevance…
StrategyConstructivismProject-Based Learning (PBL) : Centered on an authentic problem, guiding students in long-term inquiry…
CognitivismScaffolding : Providing structured support to help students complete tasks beyond their current ability…
Table 1. A selection of Preset Design Hints provided in CoMAP to scaffold the design process across different node types and categories.
Figure 5. Two export modes supported by CoMAP: (a) an editable linear lesson plan document for administrative or documentation purposes, and (b) printable cards for physical collaboration and hands-on brainstorming. These export options bridge the gap between digital design and traditional teaching workflows. Figure 5 illustrates two different export modes supported by the CoMAP system. Panel (a), titled "Editable Lesson Plan," shows a digital document in a linear, traditional lesson plan format. This document contains detailed text sections and headings, such as "Teaching Topic," "Lesson Duration," and "Instructional Strategies," representing a final, administrative or documentation-ready output. Panel (b), titled "Using Printed Card to Brainstorm," displays a physical set of brightly colored, printed cards scattered on a surface. These cards correspond to the node types shown in previous figures, such as "Strategy," "Objective," "Resource," and "Activity," each containing specific instructional content. This physical representation is designed to be used for hands-on, collaborative brainstorming, providing a tangible way to bridge the digital design process with traditional teaching and planning workflows.
Figure 6. The CoMAP Global Agent’s conversational workflow. (a) The Global Agent observes the entire design canvas and provides support through dialogue. The agent’s recommendations can be for either creation or modification. (b) In response to a teacher’s natural language query, the agent generates a new, structured card recommendation. Upon acceptance, this card is added to the canvas. (c) Alternatively, the agent provides a structured recommendation to modify an existing card. Upon acceptance, the targeted card on the canvas is updated to reflect the new information. This demonstrates the direct translation from a high-level dialogue intent to both generative and adaptive structured canvas updates. Figure 6 illustrates the conversational workflow of the CoMAP Global Agent. Panel (a) shows the Global Agent, a specialized AI, observing the entire canvas, which is filled with interconnected nodes. The agent's function is to "Observe the whole map" and "Provide structured recommendations in dialogue." Panel (b) demonstrates how the agent responds to a user's request for improvement. A dialogue bubble shows the agent's suggestion to improve a curriculum and focus on "evaluation," offering a new card recommendation for a "Process Evaluation Scale." Upon the user accepting this suggestion, the new card is added to the canvas, and a highlighted line directs the user's attention to the change. Panel (c) shows a different scenario where the agent suggests modifying an existing card to a "Process Evaluation Scale" with an emphasis on social skills, based on the user's need. Once accepted, the targeted card on the canvas is updated to reflect the new, more detailed information. This figure demonstrates the Global Agent's ability to provide high-level, context-aware suggestions that directly translate into both new card creations and the modification of existing content on the shared canvas.
Figure 7. The "before-and-after" comparison UI for CoMAP’s Local Agents. (Left) The Refine Agent for ’Content Optimization’ transforms ’Original Content’ into ’Optimized Content’ version, enhancing its quality and detail. (Right) The Splitting Agent for ’Content Splitting’ decomposes a complex ’Original Content’ into multiple, selectable ’Split Content’ components. This design explicitly preserves user agency by making AI suggestions transparent and providing teachers with fine-grained control to accept or reject changes. Figure 7 provides a side-by-side "before-and-after" comparison of the user interface for CoMAP's local AI agents. The panel on the left shows the Refine Agent for "Content Optimization." It compares an "Original Content" box with an "Optimized Content" box, which has been enhanced to be more detailed and structured, starting with an engaging question and outlining distinct phases of the lesson plan. The panel on the right shows the Split Agent for "Content Splitting." It takes a block of "Original Content" and decomposes it into multiple, smaller "Split Content" components, represented as individual, selectable check-boxes. This design allows the teacher to have granular control and accept or reject each part of the AI's suggestion, ensuring user agency is preserved.
DimensionM (CoMAP)SDM (Baseline)SDt/Wpholmp_{\text{holm}}pANCOVAp_{\text{ANCOVA}}Cohen’s d
Perceived Expression6.130.774.151.48t = 6.37<.001<.001<.001<.0011.1079
Perceived Understanding5.770.804.291.37W = 12<.001<.001<.001<.0011.1627
t/W: t-test or Wilcoxon test statistic; pholmp_{\text{holm}} : p-value from Holm’s correction
Table 2. RQ1: Comparison of user ratings for design expression and understanding between CoMAP and Baseline conditions, with ANCOVA controlling for TPACK.
DimensionM (CoMAP)SDM (Baseline)SDt/Wpholmp_{\text{holm}}pANCOVAp_{\text{ANCOVA}}Cohen’s d
Controllability5.021.134.161.66t = 2.870.0080.0380.5235
Transparency5.420.804.451.35W = 63.50.0030.0030.7226
Cognitive Load5.670.854.741.53t = 3.380.0060.0110.6176
Collaboration5.720.904.491.61W = 75.50.0070.0010.7161
Trust5.550.884.571.31W = 460.0010.0050.7750
t/W: t-test or Wilcoxon test statistic; pholmp_{\text{holm}} : p-value from Holm’s correction;
Table 3. RQ2: Comparison of user ratings for Human-AI Interaction Experience between CoMAP and Baseline conditions, with ANCOVA controlling for TPACK.
MetricM (CoMAP)SDM (Baseline)SDtprawp_{\text{raw}}pANCOVAp_{\text{ANCOVA}}Cohen’s d
Session Duration (mins)12.865.9914.455.84t = -0.890.3760.2890.26
Total Turns11.824.3918.647.15t = -3.91<.001<.001<.001<.0011.11
Avg. User Message Length63.0828.6374.2732.70t = -1.220.2280.2470.35
Negative Keywords2.142.674.162.54t = -2.600.0130.0080.76
t: t-test statistic; prawp_{\text{raw}} : p-value from raw t-test
Table 4. Comparison of efficiency and interaction process metrics between CoMAP and Baseline conditions, with ANCOVA controlling for TPACK.
More in this topic

Educational Technology

Educational technology studies digital tools and systems that support teaching, learning, assessment, and educational decision-making.

See all 118 papers

Explore similar work

CardsList
Apr 20, 2026

Choose Your Own Adventure: Non-Linear AI-Assisted Programming with EvoGraph

Current AI-assisted programming tools are predominantly linear and chat-based, which deviates from the iterative and branching nature of programming itself. Our preliminary study with developers using AI assistants suggested that they often struggle to explore alternatives, manage prompting sequences, and trace changes. Informed by these insights, we created EvoGraph, an IDE plugin that integrates AI interactions and code changes as a lightweight and interactive development graph. EvoGraph automatically records a branching AI-assisted coding history and allows developers to manipulate the graph to compare, merge, and revisit prior collaborative AI programming states. Our user study with 20 participants revealed that EvoGraph addressed developers' challenges identified in our preliminary study while imposing lower cognitive load. Participants also found the graph-based representation supported safe exploration, efficient iteration, and reflection on AI-generated changes. Our work highlights design opportunities for tools to help developers make sense of and act on their problem-solving progress in the emerging AI-mediated programming context.

Vassilios Exarhakos, Jinghui Cheng, Jin L. C. Guo

McGill University Montreal, Quebec, Canada · Polytechnique Montreal Montreal, Quebec, Canada

Human-AI InteractionHuman-AI CollaborationAI-Assisted Software Development
AskPDF
Sep 11, 2026

Generative AI and Extended Reality in Collaborative Architectural Design Education: An Exploratory Studio Study

Architectural design education relies heavily on visual ideation and representation to support collaborative learning in studio environments. Recent advances in generative artificial intelligence (GenAI) and extended reality (XR) offer new opportunities for rapid idea exploration and immersive spatial visualization. This exploratory mixed-methods classroom study investigated how GenAI-assisted multi-user XR influenced collaborative architectural conceptual design. We developed GenARch, a pipeline that integrates GenAI-based visual generation with collaborative XR environments, and deployed it in an undergraduate architectural design studio. Twenty-seven students formed seven self-selected design teams; four teams incorporated GenARch into their usual course workflow to support collaborative ideation and visualization, while three teams continued the same course workflow without GenARch. Pre- and post-intervention surveys assessed design self-efficacy, attitudes toward collaborative learning, and teamwork; a seven-member panel evaluated team design presentations; and GenARch teams participated in group interviews. The quantitative results showed larger relative declines in confidence and outcome expectancy for the GenARch condition and a positive difference-in-differences estimate for perceived conflict management, while panel-rated presentation outcomes were not significantly different between conditions. Interviews indicated complementary roles for the technologies: GenAI supported idea externalization and visual reference generation, whereas XR supported spatial, contextual, and scale-based evaluation. Students also reported challenges related to control, dimensional fidelity, shared attention, and motion comfort. These findings highlight both opportunities and limitations when GenAI and XR are incorporated into collaborative design education.

Yao Xiao, Max Chen, Yichen Li +5

Department of Civil, Environmental, and Architectural Engineering, Worcester Polytechnic Institute, Worcester, MA, USA · Interactive Media and Game Development, Worcester Polytechnic Institute, USA · Communication and Media Arts, Western Connecticut State University, USA

Generative AI in EducationHuman-AI Collaboration
AskPDF
Mar 2, 2026

"When to Hand Off, When to Work Together": Understanding Concurrent Human-Agent Interaction in Shared Co-Creative Workspaces

Shared co-creative workspaces allow users to contribute while agents execute. Yet how users adjust their participation as their work and the agent's execution shape one another remains less understood. We conducted two design probe studies with professional designers. In Study 1 (N=10), participants identified opportunities to contribute by observing the agent's progress, but avoided intervention when the agent misinterpreted their edits. Therefore, we developed CLEO, a second probe that tracks user actions and selectively updates its plan. In Study 2 (N=12), we analyzed 255 interaction turns and stimulated recall interviews, identifying five action categories, ten codes, six triggers, and four enabling factors. Participants engaged in concurrent interaction during 33.73% of turns, expressing newly recognized design ideas or preferences through edits and changing the division of work by completing pending subtasks or independently developing copies of unfinished outputs. We present a decision model and five design implications for concurrent human-agent collaboration.

Kihoon Son, Hyewon Lee, DaEun Choi +6

School of Computing, KAIST, Daejeon, Republic of Korea · Department of Computer Science, Purdue University, West Lafayette, Indiana, United States · Computer Science and Engineering, University of Michigan, Ann Arbor, Michigan, United States +3

Mixed-Initiative InteractionHuman-AI InteractionHuman-AI CollaborationHuman-AI Co-Creation
AskPDF