We study how machine learning can improve education, with a particular focus on intermediate and advanced English.
The learner answers questions and completes activities. Their responses provide evidence about what they know.
A record of answers and learning activities. The models use this history to update estimates and choose the next task.
Books, videos and other content contain learning concepts. Identifying those concepts connects materials to a learner’s needs.
Identifies the learning concepts in content and student answers: word meanings, expressions, skills and misconceptions.
The units the system recognises and teaches. For English, these include individual word meanings and expressions.
The Knowledge Prediction model combines concepts with a learner’s history to estimate what that learner already knows.
Estimates between zero and one express uncertainty about a learner’s knowledge of each concept. They can change as new evidence arrives.
Selects the next concept and activity using learning history, knowledge estimates and the available tasks.
Organises the concepts that can be taught. Depending on the domain, it may include explicit relationships or a collection of concepts without them.
Pairs a learning concept with an activity. The learner’s response becomes part of the history for the next decision.
The available learning activities: different ways to practise, explain or assess a concept.
LLMs reach 95.6% word-sense accuracy on lexEN, versus about 85% for traditional systems. LLM-generated training labels improve smaller models by 2.4–3.4 F1 points on Raganato ALL.
A framework for turning human research hypotheses into parallel, verifiable experiments. Deterministic checks keep experiments isolated and completed work immutable.
How difficult is an English word for a learner with a particular first language? We combine holistic difficulty models with features describing words and learners’ language backgrounds.
Expert-reviewed labels for English word sense disambiguation.
A benchmark of accuracy and cost in English word sense disambiguation.
A bi-encoder for matching English words in context to their meanings.
LLM-generated replacement labels for the SemCor training corpus.
Individual English meanings, with examples, images, video, difficulty and frequency—and the research connecting them to personalised learning.
ERCT is our practical framework for reading educational randomised controlled trials: twelve questions about design, outcomes and reproducibility.
Four of the questions we ask
A selection from April–September 2026: classroom trials, virtual students, knowledge prediction and learning from teaching logs.
Pavel Katunin on incremental teaching, reinforcement learning and the three models behind Glite’s approach.
Finding learning concepts in text and student answers: how the models work, how accurate they are, and what they cost at scale.
Talks and paper reviews with researchers, educators and people building learning tools.
