Skip to content

Estonian researchers develop AI to personalize student learning without big language models

What if AI understood students' emotions as well as their math skills? Estonia's groundbreaking project rethinks digital learning from the ground up.

The image shows a whiteboard with the words "Learning Analytics Personalization" written on it,...
The image shows a whiteboard with the words "Learning Analytics Personalization" written on it, along with diagrams and text that illustrate the concept of personalization and ethics.

Estonian researchers develop AI to personalize student learning without big language models

Researchers at Tallinn University are building an AI algorithm designed to support students’ learning in a more personalised way. Unlike many current tools, this model will not rely on existing large language models. Instead, it will focus on Estonian students’ emotions, motivation, and cognitive abilities to improve digital learning.

The project begins with eighth-grade mathematics, aiming to test the system in schools over the next two years. The research team, led by Professor Danial Hooshyar, has already studied the effects of digital learning tools on nearly 700 Estonian eighth-graders. Three versions of the Opiq platform were tested to understand how different approaches influence student performance.

The new algorithm will support self-regulated learning through three distinct modes, blending AI analysis with direct input from learners. Early results show it makes fewer errors in predicting students’ knowledge levels compared to traditional large language models. Hooshyar stresses the need for transparency and collaboration with social scientists and teachers to ensure the AI meets real classroom needs.

Over the next two years, the algorithm will be trialled in schools. The goal is to prove that AI can enhance digital learning environments without depending on pre-built large language models. The project marks a shift toward AI tools that adapt to individual students rather than relying on generic models. If successful, the algorithm could offer schools a more precise way to support learning in subjects like mathematics. Testing in real classrooms will determine its practical benefits over the next two years.

Read also:

Latest