MyEdMaster Research Shows Self‑Assessment AI Tutor Boosts Health‑Concept Learning Scores Far Beyond Standard LLMs
Students using a personalized, self‑assessment‑driven chatbot scored the equivalent of a full letter grade higher than peers using Large Language Models (LLM) alone.
Leesburg, VA, January 22, 2026 --(PR.com)-- A new study spanning middle and high school learners has found that a personalized AI chatbot—one that incorporates a brief self‑assessment step before tutoring—significantly outperforms a standard large language model (LLM) in helping students learn complex health concepts. Participants using the self‑assessment chatbot scored 80.5 percent (B average) on a posttest, compared to 58.8 percent (F average) among those using ChatGPT‑5‑2 alone, a statistically significant difference.
The findings build on more than four decades of research showing that individualized instruction consistently outperforms whole‑class teaching, including Bloom’s landmark “2‑sigma” effect. While intelligent tutoring systems (ITS) have long demonstrated the power of adaptive instruction, mainstream LLMs lack a core feature of ITS: a model of what the learner already knows.
“This study shows that when AI tools are given even a small window into a learner’s actual knowledge—rather than treating every user the same—their effectiveness increases dramatically,” said Dr. John Leddo, whose decades of work on Cognitive Structure Analysis (CSA) and the INKS knowledge framework underpin the self‑assessment method used in the experiment.
The research team developed a custom personal agent built on the GPT‑5‑2 API. Before asking questions, students in the experimental group completed a short self‑assessment using the CSA method, which captures four types of knowledge: factual, strategic, procedural, and rationale‑based. The chatbot then used this profile to tailor its explanations and guidance.
This simple step—completed in about ten minutes—produced a dramatic improvement in learning outcomes.
“Students don’t need a complex diagnostic system,” said Leddo. “They just need a structured way to articulate what they know and don’t know. When the chatbot uses that information, it stops giving generic answers and starts teaching.”
Previous studies by Leddo and collaborators have shown that CSA‑based self‑assessment improves learning across subjects including algebra, calculus, chemistry, Spanish, reading comprehension, biology, and history. The new study is the first to test whether the same approach enhances learning in health education, specifically the biology and psychology of stress.
The results were clear: students using the personalized chatbot scored 21.7 percentage points higher than those using ChatGPT alone. The study highlights a critical gap in current AI tutoring tools: they deliver information but rarely check whether users understand it. Research shows that learners often misjudge their own comprehension, making self‑assessment and feedback loops essential.
“AI systems today answer questions, but they don’t ensure learning,” said Leddo. “This research suggests that adding a simple self‑assessment step could be a low‑cost, scalable way to dramatically improve educational outcomes.”
The authors note that future research will explore how personalization interacts with learner characteristics, content difficulty, and answer formats—and whether similar gains can be achieved across additional subject areas.
About the Research The research draws on Cognitive Structure Analysis (CSA) and the INKS knowledge framework, which assess four types of knowledge essential for mastery: factual, strategic, procedural, and rationale‑based understanding. Prior studies using CSA have shown strong correlations between assessed knowledge and problem‑solving performance, as well as significant gains when students remediate their own knowledge gaps. The present experiment extends this body of work by testing whether a lightweight, self‑assessment‑informed chatbot can outperform a general‑purpose large language model in teaching health‑related content. Results show that incorporating a brief self‑assessment step enables the chatbot to tailor explanations to each learner’s cognitive profile, producing substantially higher posttest scores than standard AI tools.
You can find additional resources, templates, and links to published studies at (https://www.myedmaster.com/ways-to-improve-learning).
About MyEdMaster, LLC MyEdMaster is an educational company specializing in tutoring, test preparation, and STEM enrichment. Its students have published research in scientific journals and developed commercial technologies, including projects that corrected Nobel Prize‑winning work in economics. MyEdMaster’s spin‑off company, METY Technology, Inc., continues to advance these innovations.
About Dr. John Leddo
Dr. Leddo is the founder of MyEdMaster, LLC. He holds a PhD in educational psychology from Yale University and has over 100 scientific publications. Raised by a single mom in New York City, he was fortunate to receive a full scholarship to Phillips Exeter Academy. He was impressed not only by the excellent education he received, but also by the sense of
empowerment that his fellow students had and their conviction that they would be highly successful in life. Seeing the Exeter's students' sense of empowerment sparked a lifelong quest to help all students get the types of education they need to succeed in life.
The findings build on more than four decades of research showing that individualized instruction consistently outperforms whole‑class teaching, including Bloom’s landmark “2‑sigma” effect. While intelligent tutoring systems (ITS) have long demonstrated the power of adaptive instruction, mainstream LLMs lack a core feature of ITS: a model of what the learner already knows.
“This study shows that when AI tools are given even a small window into a learner’s actual knowledge—rather than treating every user the same—their effectiveness increases dramatically,” said Dr. John Leddo, whose decades of work on Cognitive Structure Analysis (CSA) and the INKS knowledge framework underpin the self‑assessment method used in the experiment.
The research team developed a custom personal agent built on the GPT‑5‑2 API. Before asking questions, students in the experimental group completed a short self‑assessment using the CSA method, which captures four types of knowledge: factual, strategic, procedural, and rationale‑based. The chatbot then used this profile to tailor its explanations and guidance.
This simple step—completed in about ten minutes—produced a dramatic improvement in learning outcomes.
“Students don’t need a complex diagnostic system,” said Leddo. “They just need a structured way to articulate what they know and don’t know. When the chatbot uses that information, it stops giving generic answers and starts teaching.”
Previous studies by Leddo and collaborators have shown that CSA‑based self‑assessment improves learning across subjects including algebra, calculus, chemistry, Spanish, reading comprehension, biology, and history. The new study is the first to test whether the same approach enhances learning in health education, specifically the biology and psychology of stress.
The results were clear: students using the personalized chatbot scored 21.7 percentage points higher than those using ChatGPT alone. The study highlights a critical gap in current AI tutoring tools: they deliver information but rarely check whether users understand it. Research shows that learners often misjudge their own comprehension, making self‑assessment and feedback loops essential.
“AI systems today answer questions, but they don’t ensure learning,” said Leddo. “This research suggests that adding a simple self‑assessment step could be a low‑cost, scalable way to dramatically improve educational outcomes.”
The authors note that future research will explore how personalization interacts with learner characteristics, content difficulty, and answer formats—and whether similar gains can be achieved across additional subject areas.
About the Research The research draws on Cognitive Structure Analysis (CSA) and the INKS knowledge framework, which assess four types of knowledge essential for mastery: factual, strategic, procedural, and rationale‑based understanding. Prior studies using CSA have shown strong correlations between assessed knowledge and problem‑solving performance, as well as significant gains when students remediate their own knowledge gaps. The present experiment extends this body of work by testing whether a lightweight, self‑assessment‑informed chatbot can outperform a general‑purpose large language model in teaching health‑related content. Results show that incorporating a brief self‑assessment step enables the chatbot to tailor explanations to each learner’s cognitive profile, producing substantially higher posttest scores than standard AI tools.
You can find additional resources, templates, and links to published studies at (https://www.myedmaster.com/ways-to-improve-learning).
About MyEdMaster, LLC MyEdMaster is an educational company specializing in tutoring, test preparation, and STEM enrichment. Its students have published research in scientific journals and developed commercial technologies, including projects that corrected Nobel Prize‑winning work in economics. MyEdMaster’s spin‑off company, METY Technology, Inc., continues to advance these innovations.
About Dr. John Leddo
Dr. Leddo is the founder of MyEdMaster, LLC. He holds a PhD in educational psychology from Yale University and has over 100 scientific publications. Raised by a single mom in New York City, he was fortunate to receive a full scholarship to Phillips Exeter Academy. He was impressed not only by the excellent education he received, but also by the sense of
empowerment that his fellow students had and their conviction that they would be highly successful in life. Seeing the Exeter's students' sense of empowerment sparked a lifelong quest to help all students get the types of education they need to succeed in life.
Contact
METY
Anthony Berry
774-317-0422
www.myedmaster.com
Anthony Berry
774-317-0422
www.myedmaster.com
Contact
Categories