Recent research has experimented with using Large Language Models (LLMs) for simulating student responses to exam questions. This approach, known as virtual pretesting, potentially offers a scalable alternative to traditional pretesting, which is costly and time-intensive, by enabling the creation of datasets of virtual students’ responses. Prior studies focused on zero-shot role-playing, prompting one LLM to imitate students of different levels, but showed limited alignment with response patterns of real students. This work introduces a framework that improves the alignment of LLM-based student simulations through in-context learning (ICL), leveraging previous question-answer records to provide the model with richer information about students’ skills and misconceptions. Our experiments show that not all models can leverage the additional contextual information. However, a multi-model approach, which combines simulations from several models, significantly improves alignment of the simulated responses when provided with relevant context: we observe a reduction of up to 30% in difficulty estimation RMSE with respect to the non contextual and individual contextual models. Overall, our findings indicate that LLMs can be used with ICL to create synthetic datasets of student responses approximating some patterns of learner behavior, however their ability to align with authentic student performance remains limited.

This paper details a system developed for the SlavicNLP 2025 Shared Task on the Detection and Classification of Persuasion Techniques in Texts for Slavic Languages (Bulgarian, Croatian, Polish, Russian and Slovene). The shared task comprises two subtasks: binary detection of persuasive content within text fragments and multi-class, multi-label identification of specific persuasion techniques at the token level. Our primary approach for both subtasks involved fine-tuning pre-trained multilingual Transformer models. For Subtask 1 (paragraph‐level binary detection) we fine‐tuned a multilingual Transformer sequence classifier, its training augmented by a set of additional labelled data. For Subtask 2 (token‐level multi‐label classification) we re‐cast the problem as named‐entity recognition. The resulting systems reached F1 score of 0.92 in paragraph‐level detection (ranked third on average). We present our system architecture, data handling, training procedures, and official results, alongside areas for future improvement.

To tailor a learning system to the student’s level and needs, we must consider the characteristics of the learning content, such as its difficulty. While natural language processing allows us to represent text efficiently, the meaningful representation of mathematical formulas in an educational context is still understudied. This paper adopts structural embeddings as a possible way to bridge this gap. Our experiments validate the approach using publicly available datasets to show that incorporating syntactic information can improve performance in predicting the exercise difficulty.

Being able to accurately perform Question Difficulty Estimation (QDE) can improve the accuracy of students’ assessment and better their learning experience. Traditional approaches to QDE are either subjective or introduce a long delay before new questions can be used to assess students. Thus, recent work proposed machine learning-based approaches to overcome these limitations. They use questions of known difficulty to train models capable of inferring the difficulty of questions from their text. Once trained, they can be used to perform QDE of newly created questions. Existing approaches employ supervised models which are domain-dependent and require a large dataset of questions of known difficulty for training. Therefore, they cannot be used if such a dataset is not available ( for new courses on an e-learning platform). In this work, we experiment with the possibility of performing QDE from text in an unsupervised manner. Specifically, we use the uncertainty of calibrated question answering models as a proxy of human-perceived difficulty. Our experiments show promising results, suggesting that model uncertainty could be successfully leveraged to perform QDE from text, reducing both costs and elapsed time.

We present a visualisation tool which aims to illuminate the inner workings of an LSTM model for question answering. It plots heatmaps of neurons’ firings and allows a user to check the dependency between neurons and manual features. The system possesses an interactive web-interface and can be adapted to other models and domains.

Code-Mixing (CM) is the phenomenon of alternating between two or more languages which is prevalent in bi- and multi-lingual communities. Most NLP applications today are still designed with the assumption of a single interaction language and are most likely to break given a CM utterance with multiple languages mixed at a morphological, phrase or sentence level. For example, popular commercial search engines do not yet fully understand the intents expressed in CM queries. As a first step towards fostering research which supports CM in NLP applications, we systematically crowd-sourced and curated an evaluation dataset for factoid question answering in three CM languages - Hinglish (Hindi+English), Tenglish (Telugu+English) and Tamlish (Tamil+English) which belong to two language families (Indo-Aryan and Dravidian). We share the details of our data collection process, techniques which were used to avoid inducing lexical bias amongst the crowd workers and other CM specific linguistic properties of the dataset. Our final dataset, which is available freely for research purposes, has 1,694 Hinglish, 2,848 Tamlish and 1,391 Tenglish factoid questions and their answers. We discuss the techniques used by the participants for the first edition of this ongoing challenge.