💻 computer science

Explainable Dual-view Workflow Discovery for GenAI-supported Learning in Higher Education

This study introduces an Explainable Dual-view Workflow Discovery framework that analyzes log data from 278 students to reveal that effective GenAI-supported learning is characterized not merely by chat frequency, but by the integration of AI feedback into writing, revision, and cross-resource regulation.

Original authors: Fan Zhang, Wanli Xing

Published 2026-09-23

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Fan Zhang, Wanli Xing

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

In the modern classroom, a new kind of tutor has arrived, one that can answer questions, draft paragraphs, and summarize complex texts in seconds. This technology, known as generative artificial intelligence, is rapidly becoming a standard tool for university students. However, a simple question remains unanswered: does simply using this tool help a student learn, or does it merely create the illusion of activity? For decades, researchers in learning analytics have understood that the story of learning is not found in a final grade or a simple count of how many times a tool was opened. Instead, the true story lies in the sequence of actions—the rhythm of reading, thinking, writing, and revising over time. Just as a doctor looks at the flow of blood rather than a single snapshot to understand health, educators need to see how a student moves through a task to understand how they are actually learning.

A team of researchers at the University of Florida and the University of Miami set out to solve a specific puzzle in this landscape. They wanted to know how students actually organize their work when they use generative AI to solve complex information problems. The study focused on 278 students enrolled in a course where they had to research a topic and write a proposal. The researchers did not just look at how many times a student asked the AI a question. Instead, they built a detailed map of every action the students took, from reading a source document to typing a sentence, highlighting a passage, or chatting with the AI. They treated these actions as a timeline, looking for patterns that repeated across different students.

The researchers discovered that looking at the timeline of actions alone was not enough. A student might spend a long time reading and asking questions, while another might spend the same amount of time writing and revising. To get the full picture, the team combined two different ways of looking at the data. The first view tracked the order of events, like a movie of the student's work. The second view looked at the overall habits of the student, such as how much they wrote after receiving AI feedback or how often they revised their work. By merging these two perspectives, the researchers could identify distinct types of work habits that were not visible when looking at either view in isolation.

The analysis revealed two very different ways students approached their work with the help of AI. The first group, comprising 145 students, followed a path of exploration. These students spent a significant amount of time reading materials, navigating the platform, and chatting with the AI. They asked many questions and gathered a lot of information. However, the data showed that they struggled to turn this information into written work. Their writing growth was low, and they did not absorb the AI's suggestions into their drafts. They were active, but their activity remained separate from the final product.

The second group, consisting of 133 students, followed a path of integration. These students also used the AI, but their workflow showed a different pattern. They used the AI to help them write, revise, and improve their arguments. When they received a response from the AI, they did not just read it; they incorporated it into their writing, deleted weak sentences, and built stronger claims. The data showed that this group had a much higher rate of writing growth and revision. They treated the AI as a partner in the writing process rather than just a source of information.

The most important finding of the study is that the frequency of using the AI does not determine how well a student learns. A student who chats with the AI many times is not necessarily learning more than a student who chats less. The key difference is what happens after the chat. Learning occurs when the feedback from the AI is absorbed into the writing process. The researchers found that the students who successfully integrated AI into their writing demonstrated a distinct workflow focused on writing growth and revision, while the exploratory group showed a different pattern of high engagement with reading and chat but lower absorption. Notably, the data indicated that the exploratory group actually achieved a slightly higher mean proposal score than the integrated group, suggesting that high scores can emerge from different workflow strategies and that the relationship between workflow type and final grade is complex.

This discovery challenges the common assumption that more interaction with a tool equals better learning. The study suggests that the educational value of generative AI emerges only when it becomes part of a regulated cycle of evidence use, feedback interpretation, and writing. For students who are stuck in the exploration phase, simply encouraging them to use the AI more will not help. Instead, they need support that helps them convert their research and questions into written arguments. For students who are already integrating the tool, the focus should shift to ensuring they are evaluating the quality of the AI's suggestions and refining their own arguments.

The researchers used a method that allowed them to see these patterns clearly without relying on complex mathematical models that are hard to interpret. They focused on the actual behaviors recorded by the system, such as the rate of writing, the number of revisions, and the timing of interactions. They found that the most important factors for a successful proposal were the rate of writing actions, the volume of work produced, and the frequency of revision. These factors mattered more than the specific content of the chat or the number of times a student clicked a button.

This study provides a new way for educators to understand how students learn with AI. It moves beyond simple metrics like "number of chats" to look at the quality of the interaction. The findings suggest that effective teaching with AI requires helping students bridge the gap between gathering information and producing writing. It is not enough to have the tool; students must learn to weave the tool's output into their own thinking and writing. By understanding these different workflow patterns, educators can design better support systems that guide students from passive exploration to active, integrated learning. The study concludes that the future of AI in education depends not on the technology itself, but on how well students can absorb its feedback into their own growth.

Technical Summary: Explainable Dual-view Workflow Discovery for GenAI-supported Learning in Higher Education

Problem Statement
While Generative AI (GenAI) is increasingly integrated into higher education, empirical understanding of how students organize their interactions with AI remains limited. Existing research often relies on static indicators such as usage frequency, prompt quality, or final product scores. These metrics fail to capture the temporal dynamics of learning, specifically how AI interaction is embedded within broader processes of reading, writing, annotation, and feedback uptake. A critical gap exists in distinguishing between students who may have similar chat counts but vastly different learning trajectories: one student might use AI to clarify concepts and subsequently revise their work, while another might engage in frequent questioning without converting responses into writing growth. Current learning analytics methods often struggle to simultaneously preserve the path structure (temporal sequence of actions) and learning mechanisms (aggregated indicators of absorption and regulation) required to identify educationally productive workflows.

Methodology
The study introduces Explainable Dual-view Workflow Discovery, a framework designed to identify and interpret student workflows in a GenAI-supported information problem-solving (IPS) course. The analysis utilized a dataset of 278 students from the FLoRA Open Data IPS records, containing complete log, dialogue, writing, and outcome data.

The methodology proceeds through four main stages:

Dual-View Representation:
- Sequence Path View: Encodes student behavior as chronological sequences of behavioral tokens (READ, WRITE, ANNOTATE, CHAT, RESPONSE). Two granularities were tested: "family" tokens (broad action categories) and "full" tokens (fine-grained events). Five sequence algorithms were compared, including TF-IDF KMeans, LSA sequence embedding, Markov transition models, and multiview KMeans.
- Student Profile View: Represents aggregated, mechanism-oriented features such as chat intensity, reading/writing ratios, writing growth, revision rates, and "AI absorption" (writing growth following chat). Five clustering algorithms were compared, with a focus on PCA-based dimensionality reduction to handle correlated variables.
Dual-view Workflow Consensus Embedding (DWCE):
To integrate the temporal structure of the sequence view with the mechanism insights of the profile view, the authors proposed DWCE. Two fusion strategies were evaluated:
- Weighted Concatenation: Combining latent sequence and profile embeddings with varying weights.
- Consensus Similarity: A conservative approach where two students are considered similar only if both views indicate similarity. This prevents masking different mechanisms with similar path orders or vice versa.
Clustering and Selection:
The fused representations were clustered using KMeans, Gaussian Mixture Models (GMM), agglomerative clustering, and spectral clustering. Model selection prioritized not just statistical separation (e.g., silhouette scores) but also cluster balance, minimum cluster size, and, crucially, educational interpretability.
Explainable AI (XAI) Analysis:
To translate computational clusters into educational evidence, three XAI techniques were applied:
- Cluster Contrast: Comparing cluster features against the global mean to identify distinguishing characteristics.
- Surrogate Tree: Fitting a shallow decision tree to approximate cluster assignments, yielding readable rules.
- Outcome XAI: Using a random forest model with permutation importance to identify workflow features most relevant to proposal scores, while acknowledging that logs alone may not fully predict grades.

Key Results

Single-View Performance: The sequence path view (using LSA sequence embedding with KMeans) produced the clearest path taxonomy with a cosine silhouette of 0.552, identifying three distinct workflow shapes (Read-Write Integrated, Reading-Centered, and Annotation-Heavy). The student profile view (PCA feature KMeans) yielded a lower silhouette (0.348) but provided valuable mechanistic insights regarding writing growth and AI absorption.
DWCE Fusion: The consensus similarity embedding with GMM produced the selected fusion solution ( $k=2$ ) with a cosine silhouette of 0.517 and a balanced cluster distribution ( $n=145$ and $n=133$ ).
Workflow Taxonomy: The DWCE identified two primary workflow types:
1. Reading-Chat Exploratory Workflow: Characterized by high reading, chat, and navigation rates but lower writing growth and AI absorption.
2. AI-Integrated Writing Absorption Workflow: Characterized by higher writing rates, writing growth, and strong AI absorption (converting chat into writing and revision).
Feature Importance: Explainable analyses revealed that essay action rate, write token rate, trace events, chat token rate, navigation rate, annotation rate, sequence length, and revision rate were the most outcome-relevant features. Notably, chat frequency alone was not the primary differentiator; the critical factor was whether AI interaction led to writing activity.

Significance and Claims
The paper claims three primary contributions:

Stable Taxonomy: It demonstrates that latent sequence workflow clustering can reveal a stable taxonomy of GenAI-supported learning paths.
Mechanism Explanation: It shows that student profile clustering explains whether AI interaction becomes part of writing production and regulation.
Integrated Interpretation: It establishes that DWCE can effectively connect path taxonomy with mechanism profiling, offering a more educationally meaningful interpretation of effective versus ineffective GenAI interaction.

The central educational conclusion is that the value of GenAI interaction is not determined by the frequency of use but by absorption. Educational value emerges when AI feedback is absorbed into writing growth, revision, and cross-resource regulation. The study argues that effective interaction is a regulated cycle of evidence use, feedback interpretation, and writing, rather than isolated prompting.

Limitations
The authors modestly note that the study is observational and does not establish causal effects. The outcome variable (proposal score) is a coarse measure, and the analysis relied on source-provided chat labels rather than GPT-generated reclassification to avoid circularity. Additionally, the negative cross-validated $R^2$ in the outcome model indicates that log data alone cannot fully explain proposal quality, suggesting a need for future integration with text quality measures and qualitative data.

Technical Summary: Explainable Dual-view Workflow Discovery for GenAI-supported Learning in Higher Education

More like this