Author: Sebastian Gombert

Language learners make the fastest progress when reading texts that match their proficiency level. But most real-world texts are too hard—and manually adapting them is time-consuming. So the big question is: Can AI automatically simplify texts to a specific CEFR level without losing meaning? We explored exactly this in the TSAR 2025 Shared Task, where systems had to rewrite advanced English texts (B2+) to easier levels like A2 or B1. Our team submitted two different approaches: EZ-SCALAR and SAGA. EZ-SCALAR works like an expert panel of AI models. Two large language models (GPT-5 and Claude) each produce their own simplification, critique each other, refine their versions, and then a final “judge” model picks the best result. An extended version, EZ-SCALAR Lex, adds something extra: a vocabulary check using EFLLex, a…

Helping students apply energy ideas to everyday situations is a core goal in physics education. But not all students get there—and it’s not just about who knows the most content. In a 10-week classroom study with 165 students, we tracked both their energy understanding and their affective and metacognitive factors (like emotions, cognitive load, and self-regulation). Using k-means clustering on their learning trajectories, we identified three distinct student groups that differed in the coherence of their energy knowledge development. The key insight: Students who learned the most also felt more positive, experienced lower cognitive load, and used stronger metacognitive strategies. Those who struggled often felt overwhelmed or disengaged. The takeaway is clear: supporting emotions and self-regulation is just as important as teaching physics content. Instruction that addresses these factors can…

In the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-Powered Tutors, the goal was to evaluate how well LLM-based math tutors support students. The task focused on four aspects of feedback: spotting mistakes identifying where the mistake happens giving guidance providing actionable suggestions For our submission, we built on FLAN-T5 models with a multi-step training pipeline. In addition to standard fine-tuning, we used model merging (DARE-TIES) to leverage information across all four labels – and saw clear improvements over plain fine-tuning. Our models achieved F1 scores between 52 and 69 and accuracies between 62% and 87%, ranking 11th, 8th, 11th, and 9th across the four tracks. Link to the paper: Link

At the recent 19th Joint Summer School on Technology-Enhanced Learning located in Rhetimno, Greece, I (Sebastian Gombert) gave an introductory workshop on language technology and language modeling and their various use cases of in education. This included use cases such as short answer scoring, essay scoring, classification of texts according to the CEFR framework, and group communication analysis in CSCL. Overall, the workshop was well-received and well-attended.