This tutorial will aim to bridge the gap between NLP researchers and Artificial Intelligence in Education (AIED) practitioners to help participants understand the requirements and challenges of education, enabling them to develop LLMs that align with educational needs, and to enable educators to gain a deeper understanding of the capabilities and limitations of current NLP technologies, fostering effective integration of LLMs in educational contexts.

This study examines the lexical and syntactic interventions of human and LLM proofreading aimed at improving overall intelligibility in identical second language writings, and evaluates the consistency of outcomes across three LLMs (ChatGPT-4o, Llama3.1-8b, Deepseek-r1-8b). Findings show that both human and LLM proofreading enhance bigram lexical features, which may contribute to better coherence and contextual connectedness between adjacent words. However, LLM proofreading exhibits a more generative approach, extensively reworking vocabulary and sentence structures, such as employing more diverse and sophisticated vocabulary and incorporating a greater number of adjective modifiers in noun phrases. The proofreading outcomes are highly consistent in major lexical and syntactic features across the three models.

The rapid advancement of Large Language Models (LLMs) has transformed various domains, particularly computer science (CS) education. These models exhibit remarkable capabilities in code-related tasks and problem-solving, raising questions about their potential and limitations in advanced CS contexts. This study presents a novel bilingual (English–Romanian) multimodal (text and image) dataset of multiple-choice questions derived from a high-level computer science competition. A particularity of our dataset is that the problems are conceived such that some of them are easier solved using reasoning on paper, while for others writing code is more efficient. We systematically evaluate State of The Art LLMs on this dataset, analyzing their performance on theoretical programming tasks. Our findings reveal the strengths and limitations of current LLMs, including the influence of language choice (English vs. Romanian), providing insights into their applicability in CS education and competition settings. We also address critical ethical considerations surrounding educational integrity and the fairness of assessments in the context of LLM usage. These discussions aim to inform future educational practices and policies. To support further research, our dataset will be made publicly available in both English and Romanian. Additionally, we release an educational application tailored for Romanian students, enabling them to self-assess using the dataset in an interactive and practice-oriented environment.

Automatic Short Answer Grading (ASAG) refers to automated scoring of open-ended textual responses to specific questions, both in natural language form. In this paper, we propose a method to tackle this task in a setting where annotated data is unavailable. Crucially, our method is competitive with the state-of-the-art while being lighter and interpretable. We crafted a unique dataset containing a highly diverse set of questions and a small amount of answers to these questions; making it more challenging compared to previous tasks. Our method uses weak labels generated from other methods proven to be effective in this task, which are then used to train a white-box (linear) regression based on a few interpretable features. The latter are extracted expert features and learned representations that are interpretable per se and aligned with manual labeling. We show the potential of our method by evaluating it on a small annotated portion of the dataset, and demonstrate that its ability compares with that of strong baselines and state-of-the-art methods, comprising an LLM that in contrast to our method comes with a high computational price and an opaque reasoning process. We further validate our model on a public Automatic Essay Scoring dataset in English, and obtained competitive results compared to other unsupervised baselines, outperforming the LLM. To gain further insights of our method, we conducted an interpretability analysis revealing sparse weights in our linear regression model, and alignment between our features and human ratings.

Multiple-Choice Tasks are one of the most common types of assessment item, due to their feature of being easy to automatically and objectively grade. A key component of Multiple-Choice Tasks are distractors – i.e., the wrong answer options – since poor distractors affect the overall quality of the item: e.g., if they are obviously wrong, they are never selected. Thus, previous research has focused extensively on techniques for automatically generating distractors, which can be especially helpful in settings where large pools of questions are desirable or needed. However, there is no agreement within the community about the techniques that are most suited to evaluate generated distractors, and the ones used in the literature are sometimes not aligned with how distractors perform in real exams. In this review paper, we perform a comprehensive study of the approaches which are used in the literature for evaluating generated distractors, propose a taxonomy to categorise them, discuss if and how they are aligned with distractors performance in exam settings, and what are the differences for different question types and educational domains.

This paper investigates the potentials of Large Language Models (LLMs) as adaptive tutors in the context of second-language learning. In particular, we evaluate whether system prompting can reliably constrain LLMs to generate only text appropriate to the student’s competence level. We simulate full teacher-student dialogues in Spanish using instruction-tuned, open-source LLMs ranging in size from 7B to 12B parameters. Dialogues are generated by having an LLM alternate between tutor and student roles with separate chat histories. The output from the tutor model is then used to evaluate the effectiveness of CEFR-based prompting to control text difficulty across three proficiency levels (A1, B1, C1). Our findings suggest that while system prompting can be used to constrain model outputs, prompting alone is too brittle for sustained, long-term interactional contexts - a phenomenon we term alignment drift. Our results provide insights into the feasibility of LLMs for personalized, proficiency aligned adaptive tutors and provide a scalable method for low-cost evaluation of model performance without human participants.

Competitive programming contests play a crucial role in cultivating computational thinking and algorithmic skills among learners. However, generating comprehensive test cases to effectively assess programming solutions remains resource-intensive and challenging for educators. This paper introduces an innovative NLP-driven method leveraging generative AI (large language models) to automate the creation of high-quality test cases for competitive programming assessments. We extensively evaluated our approach on diverse datasets, including 25 years of Romanian Informatics Olympiad (OJI) data for 5th graders, recent competitions hosted on the Kilonova.ro platform, and the International Informatics Olympiad in Teams (IIOT). Our results demonstrate that AI-generated test cases substantially enhanced assessments, notably identifying previously undetected errors in 67% of the OJI 5th grade programming problems. These improvements underscore the complementary educational value of our technique in formative assessment contexts. By openly sharing our prompts, translated datasets, and methodologies, we offer practical NLP-based tools that educators and contest organizers can readily integrate to enhance assessment quality, reduce workload, and deepen insights into learner performance.We have uploaded a demo which showcases the process of using the prompt in order to generate the test cases for one of the problems from the Kilonova.ro platform, which is accessible through the file we uploaded in the supplementary material section.

Large Language Models (LLMs) offer many opportunities for scalably improving the teaching and learning process, for example, by simulating students for teacher training or lesson preparation. However, design requirements for building high-fidelity LLM-based simulations are poorly understood. This study aims to address this gap from the perspective of key stakeholders—teachers who have tutored LLM-simulated students. We use a mixed-method approach and conduct semi-structured interviews with these teachers, grounding our interview design and analysis in the Community of Inquiry and Scaffolding frameworks. Our findings indicate several challenges in LLM-simulated students, including authenticity, high language complexity, lack of emotions, unnatural attentiveness, and logical inconsistency. We end by categorizing four types of real-world student behaviors and provide guidelines for the design and development of LLM-based student simulations. These include introducing diverse personalities, modeling knowledge building, and promoting questions.

Decoder-only large language models have shown superior performance in the fluency-edit English Grammatical Error Correction, but their adaptation for minimal-edit English GEC is still underexplored. To improve their effectiveness in the minimal-edit approach, we explore the error rate adaptation topic and propose a novel training schedule method. Our experiments set a new state-of-the-art result for a single-model system on the BEA-test set. We also detokenize the most common English GEC datasets to match the natural way of writing text. During the process, we find that there are errors in them. Our experiments analyze whether training on detokenized datasets impacts the results and measure the impact of the usage of the datasets with corrected erroneous examples. To facilitate reproducibility, we have released the source code used to train our models.

While Generative AI has demonstrated strong potential and versatility in content generation, its application to educational contexts presents several challenges. Models often fail to align with curriculum standards and maintain grade-appropriate reading levels consistently. Furthermore, STEM education poses additional challenges in balancing scientific explanations with everyday language when introducing complex and abstract ideas and phenomena to younger students.In this work, we propose COGENT, a curriculum-oriented framework for generating grade-appropriate educational content. We incorporate three curriculum components (science concepts, core ideas, and learning objectives), control readability through length, vocabulary, and sentence complexity, and adopt a “wonder-based” approach to increase student engagement and interest. We conduct a multi-dimensional evaluation via both LLM-as-a-judge and human expert analysis. Experimental results show that COGENT consistently produces grade-appropriate passages that are comparable or superior to human references. Our work establishes a viable approach for scaling adaptive and high-quality learning resources.

In this work, we assess the potential of using synthetic data to train models for content scoring. We generate a parallel corpus of LLM-generated data for the SRA dataset. In our experiments, we train three different kinds of models (Logistic Regression, BERT, SBERT) with this data, examining their respective ability to bridge between generated training data and student-authored test data. We also explore the effects of generating larger volumes of training data than what is available in the original dataset. Overall, we find that training models from LLM-generated data outperforms zero-shot scoring of the test data with an LLM. Still, the fine-tuned models perform much worse than models trained on the original data, largely because the LLM-generated answers often do not to conform to the desired labels. However, once the data is manually relabeled, competitive models can be trained from it. With a similarity-based scoring approach, the relabeled (larger) amount of synthetic answers consistently yields a model that surpasses performance of training on the (limited) amount of answers available in the original dataset.

Establishing the difficulty of test items is an essential part of the language assessment development process. However, traditional item calibration methods are often time-consuming and difficult to scale. To address this, recent research has explored natural language processing (NLP) approaches for automatically predicting item difficulty from text. This paper investigates the use of transformer models to predict the difficulty of second language (L2) English vocabulary test items that have multilingual prompts. We introduce an extended version of the British Council’s Knowledge-based Vocabulary Lists (KVL) dataset, containing 6,768 English words paired with difficulty scores and question prompts written in Spanish, German, and Mandarin Chinese. Using this new dataset for fine-tuning, we explore various transformer-based architectures. Our findings show that a multilingual model jointly trained on all L1 subsets of the KVL achieve the best results, with analysis suggesting that the model is able to learn global patterns of cross-linguistic influence on target word difficulty. This study establishes a foundation for NLP-based item difficulty estimation using the KVL dataset, providing actionable insights for developing multilingual test items.

Heterogeneity in student populations poses achallenge in formal education, with adaptivetextbooks offering a potential solution by tai-loring content based on individual learner mod-els. However, creating domain models for text-books typically demands significant manual ef-fort. Recent work by Chau et al. (2021) demon-strated automated concept extraction from dig-ital textbooks, but relied on costly domain-specific manual annotations. This paper in-troduces a novel, scalable method that mini-mizes manual effort by combining contextu-alized word embeddings with weakly super-vised machine learning. Our approach clustersword embeddings from textbooks and identi-fies domain-specific concepts using a machinelearner trained on concept seeds automaticallyextracted from Wikipedia. We evaluate thismethod using 28 economics textbooks, com-paring its performance against a tf-idf baseline,a supervised machine learning baseline, theRAKE keyword extraction method, and humandomain experts. Results demonstrate that ourweakly supervised method effectively balancesaccuracy with reduced annotation effort, offer-ing a practical solution for automated conceptextraction in adaptive learning environments.

This paper presents an automatic speech assessment system designed for Swedish language learners. We introduce a novel hybrid approach that integrates Microsoft Azure speech services with open-source Large Language Models (LLMs). Our system is implemented as a web-based application that provides real-time quick assessment with a game-like experience. Through testing against COREFL English corpus data and Swedish L2 speech data, our system demonstrates effectiveness in distinguishing different language proficiencies, closely aligning with CEFR levels. This ongoing work addresses the gap in current low-resource language assessment technologies with a pilot system developed for automated speech analysis.

Grammatical Error Correction (GEC) relies on accurate error annotation and evaluation, yet existing frameworks, such as errant, face limitations when extended to typologically diverse languages. In this paper, we introduce a standardized, modular framework for multilingual grammatical error annotation. Our approach combines a language-agnostic foundation with structured language-specific extensions, enabling both consistency and flexibility across languages. We reimplement errant using stanza to support broader multilingual coverage, and demonstrate the framework’s adaptability through applications to English, German, Czech, Korean, and Chinese, ranging from general-purpose annotation to more customized linguistic refinements. This work supports scalable and interpretable GEC annotation across languages and promotes more consistent evaluation in multilingual settings. The complete codebase and annotation tools can be accessed at https://github.com/open-writing-evaluation/jp_errant_bea.

Evaluation of Grammatical Error Correction (GEC) systems is becoming increasingly challenging as the quality of such systems increases and traditional automatic metrics fail to adequately capture such nuances as fluency versus minimal edits, alternative valid corrections compared to the ‘ground truth’, and the difference between corrections that are useful in a language learning scenario versus those preferred by native readers. Previous work has suggested using human post-editing of GEC system outputs, but this is very labor-intensive. We investigate the use of Large Language Models (LLMs) as post-editors of English and Swedish texts, and perform a meta-analysis of a range of different evaluation setups using a set of recent GEC systems. We find that for the two languages studied in our work, automatic evaluation based on post-editing agrees well with both human post-editing and direct human rating of GEC systems. Furthermore, we find that a simple n-gram overlap metric is sufficient to measure post-editing distance, and that including human references when prompting the LLMs generally does not improve agreement with human ratings. The resulting evaluation metric is reference-free and requires no language-specific training or additional resources beyond an LLM capable of handling the given language.Evaluation of Grammatical Error Correction (GEC) systems is becoming increasingly challenging as the quality of such systems increases and traditional automatic metrics fail to adequately capture such nuances as fluency versus minimal edits, alternative valid corrections compared to the ‘ground truth’, and the difference between corrections that are useful in a language learning scenario versus those preferred by native readers. Previous work has suggested using human post-editing of GEC system outputs, but this is very labor-intensive. We investigate the use of Large Language Models (LLMs) as post-editors of English and Swedish texts, and perform a meta-analysis of a range of different evaluation setups using a set of recent GEC systems. We find that for the two languages studied in our work, automatic evaluation based on post-editing agrees well with both human post-editing and direct human rating of GEC systems. Furthermore, we find that a simple n-gram overlap metric is sufficient to measure post-editing distance, and that including human references when prompting the LLMs generally does not improve agreement with human ratings. The resulting evaluation metric is reference-free and requires no language-specific training or additional resources beyond an LLM capable of handling the given language.

In this paper, we address generic essay scoring, i.e., the use of training data from one writing task to score data from a different task. We approach this by generalizing a similarity-based essay scoring method (Xie et al., 2022) to learning from texts that are written in response to a mixture of different prompts. In our experiments, we compare within-prompt and cross-prompt performance on two large datasets (ASAP and PERSUADE). We combine different amounts of prompts in the training data and show that our generalized method substantially improves cross-prompt performance, especially when an increasing number of prompts is used to form the training data. In the most extreme case, this leads to more than double the performance, increasing QWK from .26 to .55.

We present an approach to the automated scoring of a German Written Elicited Imitation Test, designed to assess literacy-dependent procedural knowledge in German as a foreign language. In this test, sentences are briefly displayed on a screen and, after a short pause, test-takers are asked to reproduce the sentence in writing as accurately as possible. Responses are rated on a 5-point ordinal scale, with grammatical errors typically penalized more heavily than lexical deviations. We compare a rule-based model that implements the categories of the scoring rubric through hand-crafted rules, and a deep learning model trained on pairs of stimulus sentences and written responses. Both models achieve promising performance with quadratically weighted kappa (QWK) values around .87. However, their strengths differ – the rule-based model performs better on previously unseen stimulus sentences and at the extremes of the rating scale, while the deep learning model shows advantages in scoring mid-range responses, for which explicit rules are harder to define.

Since the release of ChatGPT, Large Langauge Models (LLMs) have been proposed as potential tutors to students in the education outcomes. Such an LLM-as-tutors metaphor is problematic, notably due to the counterfactual generation, perception of learned skills as mastered by an automated system and hence non-valuable, and learning LLM over-reliance.We propose instead the LLM-as-mentee tutoring schema, leveraging the Learning-by-Teaching protégé effect in peer tutoring - LLM Protégés. In this configuration, counterfactual generation is desirable, allowing students to operationalize the learning material and better understand the limitations of LLM-based systems, both a skill in itself and an additional learning motivation. Our preliminary results suggest that LLM Protégés are effective. Students in an introductory algorithms class who successfully diagnosed an LLM teachable agent system prompted to err on a course material gained an average of 0.72 points on a 1-6 scale. Remarkably, if fully adopted, this approach would reduce the failure rate in the second midterm from 28% to 8%, mitigating 72% of midterm failure.We publish code for on-premises deployment of LLM Protégés on https://github.com/Reliable-Information-Lab-HEVS/LLM_Proteges.

Translating technical terms into lexically similar, low-resource Indian languages remains a challenge due to limited parallel data and the complexity of linguistic structures. We propose a novel use-case of Sanskrit-based segments for linguistically informed translation of such terms, leveraging subword-level similarity and morphological alignment across related languages. Our approach uses character-level segmentation to identify meaningful subword units, facilitating more accurate and context-aware translation. To enable this, we utilize a Character-level Transformer model for Sanskrit Word Segmentation (CharSS), which addresses the complexities of sandhi and morpho-phonemic changes during segmentation. We observe consistent improvements in two experimental settings for technical term translation using Sanskrit-derived segments, averaging 8.46 and 6.79 chrF++ scores, respectively. Further, we conduct a post hoc human evaluation to verify the quality assessment of the translated technical terms using automated metrics. This work has important implications for the education field, especially in creating accessible, high-quality learning materials in Indian languages. By supporting the accurate and linguistically rooted translation of technical content, our approach facilitates inclusivity and aids in bridging the resource gap for learners in low-resource language communities.

Knowing how test takers answer items in educational assessments is essential for test development, to evaluate item quality, and to improve test validity. However, this process usually requires extensive pilot studies with human participants. If large language models (LLMs) exhibit human-like response behavior to test items, this could open up the possibility of using them as pilot participants to accelerate test development. In this paper, we evaluate the human-likeness or psychometric plausibility of responses from 18 instruction-tuned LLMs with two publicly available datasets of multiple-choice test items across three subjects: reading, U.S. history, and economics. Our methodology builds on two theoretical frameworks from psychometrics which are commonly used in educational assessment, classical test theory and item response theory. The results show that while larger models are excessively confident, their response distributions can be more human-like when calibrated with temperature scaling. In addition, we find that LLMs tend to correlate better with humans in reading comprehension items compared to other subjects. However, the correlations are not very strong overall, indicating that LLMs should not be used for piloting educational assessments in a zero-shot setting.

Generative AI systems have rapidly advanced, with multimodal input capabilities enabling reasoning beyond text-based tasks. In education, these advancements could influence assessment design and question answering, presenting both opportunities and challenges. To investigate these effects, we introduce a high-quality dataset of 201 university-level STEM questions, manually annotated with features such as image type, role, problem complexity, and question format. Our study analyzes how these features affect generative AI performance compared to students. We evaluate four model families with five prompting strategies, comparing results to the average of 546 student responses per question. Although the best model correctly answers on average 58.5% of the questions using majority vote aggregation, human participants consistently outperform AI on questions involving visual components. Interestingly, human performance remains stable across question features but varies by subject, whereas AI performance is susceptible to both subject matter and question features. Finally, we provide actionable insights for educators, demonstrating how question design can enhance academic integrity by leveraging features that challenge current AI systems without increasing the cognitive burden for students

Large language models (LLMs) are increasingly used to generate distractors for multiple-choice questions (MCQs), especially in domains like math education. However, existing approaches are limited in ensuring that the generated distractors are consistent with common student errors. We propose LookAlike, a method that improves error–distractor consistency via preference optimization. Our two main innovations are: (a) mining synthetic preference pairs from model inconsistencies, and (b) alternating supervised fine-tuning (SFT) with Direct Preference Optimization (DPO) to stabilize training. Unlike prior work that relies on heuristics or manually annotated preference data, LookAlike uses its own generation inconsistencies as dispreferred samples, thus enabling scalable and stable training. Evaluated on a real-world dataset of 1,400+ math MCQs, LookAlike achieves 51.6% accuracy in distractor generation and 57.2% in error generation under LLM-as-a-judge evaluation, outperforming an existing state-of-the-art method (45.6% / 47.7%). These improvements highlight the effectiveness of preference-based regularization and inconsistency mining for generating consistent math MCQ distractors at scale.

Designing vocabulary learning activities for foreign/second language (L2) learners highly depends on the successful identification of difficult words. In this paper, we present a novel personalised word difficulty classifier for L2 Spanish, using the LexComSpaL2 corpus as training data and a BiLSTM model as the architecture. We train a base version (using the original LexComSpaL2 data) and a word family version of the classifier (adding word family knowledge as an extra feature). The base version obtains reasonably good performance (F1 = 0.53) and shows weak positive predictive power (φ = 0.32), underlining the potential of automated methods in determining vocabulary difficulty for individual L2 learners. The “word family classifier” is able to further push performance (F1 = 0.62 and φ = 0.45), highlighting the value of well-chosen linguistic features in developing word difficulty classifiers.

Recent trends in NLP have shifted towards modeling lexical complexity as a continuous value, but practical implementations often remain binary. This opinion piece argues for the importance of truly graded lexical complexity prediction, particularly in language learning. We examine the evolution of lexical complexity modeling, highlighting the “data bottleneck” as a key obstacle. Overcoming this challenge can lead to significant benefits, such as enhanced personalization in language learning and improved text simplification. We call for a concerted effort from the research community to create high-quality, graded complexity datasets and to develop methods that fully leverage continuous complexity modeling, while addressing ethical considerations. By fully embracing the continuous nature of lexical complexity, we can develop more effective, inclusive, and personalized language technologies.

This paper introduces IPPOLIS Write, an open source, web-based tool designed to provide automated feedback on the formal aspects of scientific documents. Aimed at addressing the variability in writing and language skills among scientists and the challenges faced by supervisors in providing consistent feedback on student theses, IPPOLIS Write integrates several open source tools and custom implementations to analyze documents for a range of formal issues, including grammatical errors, consistent introduction of acronyms, comparison of literature entries with several databases, referential integrity of figures and tables, and consistent link access dates.IPPOLIS Write generates reports with statistical summaries and annotated documents that highlight specific issues and suggest improvements while also providing additional background information where appropriate. To evaluate its effectiveness, a qualitative assessment is conducted using a small but diverse dataset of bachelor’s and master’s theses sourced from arXiv. Our findings demonstrate the tool’s potential to enhance the quality of scientific documents by providing targeted and consistent feedback, thereby aiding both students and professionals in refining their document preparation skills.

Students’ argumentative writing benefits from receiving automated feedback, particularly throughout the writing process. While Argument Mining (AM) technology shows promise for delivering automated feedback on argumentative structures, existing systems are frequently trained on completed essays, providing rich context information and raising concerns about their usefulness for offering writing support on incomplete texts during the writing process. This study evaluates the robustness of AM algorithms on artificially fragmented learner texts from two large-scale corpora of secondary school essays: the German DARIUS corpus and the English PERSUADE corpus. Our analysis reveals that token-level sequence-tagging methods, while highly effective on complete essays, suffer significantly when context is limited or misleading. Conversely, sentence-level classifiers maintain relative stability under such conditions. We show that deliberately training AM models on fragmented input substantially mitigates these context-related weaknesses, enabling AM systems to support dynamic educational writing scenarios better.

The rapid development of Large Language Models (LLMs) opens up the possibility of using them aspersonal tutors. This has led to the development of several intelligent tutoring systems and learning assistants that use LLMs as back-ends with various degrees of engineering. In this study, we seek to compare human tutors with LLM tutorsin terms of engagement, empathy, scaffolding, and conciseness. We ask human tutors to compare the performance of an LLM tutor with that of a human tutor in teaching grade-school math word problems on these qualities. We find that annotators with teaching experience perceive LLMs as showing higher performance than human tutors in all 4 metrics. The biggest advantage is in empathy, where 80% of our annotators prefer the LLM tutor more often than the human tutors. Our study paints a positive picture of LLMs as tutors and indicates that these models can be used to reduce the load on human teachers in the future.

Real-word spelling errors (RWSEs) pose special challenges for detection methods, as they ‘hide’ in the form of another existing word and in many cases even fit in syntactically. We present a modern Transformer-based implementation of earlier probabilistic methods based on confusion sets and show that RWSEs can be detected with a good balance between missing errors and raising too many falsealarms. The confusion sets are dynamically configurable, allowing teachers to easily adjust which errors trigger feedback.

Weakly supervised learning (WSL) is a machine learning approach used when labeled data is scarce or expensive to obtain. In such scenarios, models are trained using weaker supervision sources instead of human-annotated data. However, these sources are often noisy and may introduce unquantified biases during training. This issue is particularly pronounced in automated scoring (AS) of second language (L2) learner output, where high variability and limited generalizability pose significant challenges.In this paper, we investigate analytical scoring of L2 learner responses under weak and semi-supervised learning conditions, leveraging Prediction-Powered Inference (PPI) to provide statistical guarantees on score validity. We compare two approaches: (1) synthetic scoring using large language models (LLMs), and (2) a semi-supervised setting in which a machine learning model, trained on a small gold-standard set, generates predictions for a larger unlabeled corpus. In both cases, PPI is applied to construct valid confidence intervals for assessing the reliability of the predicted scores.Our analysis, based on a dataset of L2 learner conversations with an AI agent, shows that PPI is highly informative for evaluating the quality of weakly annotated data. Moreover, we demonstrate that PPI can increase the effective sample size by over 150% relative to the original human-scored subset, enabling more robust inference in educational assessment settings where labeled data is scarce.

Inference making is an essential but complex skill in reading comprehension (RC). Some inferences require resolving references across sentences, and some rely on using prior knowledge to fill in the detail that is not explicitly written in the text. Diagnostic RC questions can help educators provide more effective and targeted reading instruction and interventions for school-age students. We introduce a taxonomy of inference types for RC and use it to analyze the distribution of items within a diagnostic RC item bank. Next, we present experiments using GPT-4o to generate bridging-inference RC items for given reading passages via few-shot prompting, comparing conditions with and without chain-of-thought prompts. Generated items were evaluated on three aspects: overall item quality, appropriate inference type, and LLM reasoning, achieving high inter-rater agreements above 0.90. Our results show that GPT-4o produced 93.8% good-quality questions suitable for operational use in grade 3-12 contexts; however, only 42.6% of the generated questions accurately matched the targeted inference type. We conclude that combining automatic item generation with human judgment offers a promising path toward scalable, high-quality diagnostic RC assessments.

Aligning Learning Objectives (LOs) in course descriptions with educational frameworks such as Bloom’s revised taxonomy is an important step in maintaining educational quality, yet it remains a challenging and often manual task. With the growing availability of large language models (LLMs), a natural question arises: can these models meaningfully automate LO classification, or are non-LLM methods still sufficient? In this work, we systematically compare LLM- and non-LLM-based methods for mapping LOs to Bloom’s taxonomy levels, using expert annotations as the gold standard. LLM-based methods consistently outperform non-LLM methods and offer more balanced distributions across taxonomy levels. Moreover, contrary to common concerns, we do not observe significant biases (e.g. verbosity or positional) or notable sensitivity to prompt structure in LLM outputs. Our results suggest that a more consistent and precise formulation of LOs, along with improved methods, could support both automated and expert-driven efforts to better align LOs with taxonomy levels.

We present LangEye, a mobile application for contextual vocabulary learning that combines learner-curated content with generative NLP. Learners use their smartphone camera to capture real-world objects and create personalized “memories” enriched with definitions, example sentences, and pronunciations generated via object recognition, large language models, and machine translation.LangEye features a three-phase review system — progressing from picture recognition to sentence completion and free recall. In a one-week exploratory study with 20 French (L2) learners, the learner-curated group reported higher engagement and motivation than those using pre-curated materials. Participants valued the app’s personalization and contextual relevance. This study highlights the potential of integrating generative NLP with situated, learner-driven interaction. We identify design opportunities for adaptive review difficulty, improved content generation, and better support for language-specific features. LangEye points toward scalable, personalized vocabulary learning grounded in real-world contexts.

As generative AI tools become increasingly integrated into educational research workflows, large language models (LLMs) have shown substantial promise in automating complex tasks such as topic modeling. This paper presents a user study that evaluates AI-enabled topic modeling (AITM) within the domain of P-20 education research. We investigate the benefits and trade-offs of integrating LLMs into expert document analysis through a case study of school improvement plans, comparing four analytical conditions. Our analysis focuses on three dimensions: (1) the marginal financial and environmental costs of AITM, (2) the impact of LLM assistance on annotation time, and (3) the influence of AI suggestions on topic identification. The results show that LLM increases efficiency and decreases financial cost, but potentially introduce anchoring bias that awareness prompts alone fail to mitigate.

With the rise and widespread adoption of Large Language Models (LLMs) in recent years, extensive research has been conducted on their applications across various domains. One such domain is education, where a key area of interest for researchers is investigating the implementation and reliability of LLMs in grading student responses. This review paper examines studies on the use of LLMs in grading across six academic sub-fields: educational assessment, essay grading, natural sciences and technology, social sciences and humanities, computer science and engineering, and mathematics. It explores how different LLMs are applied in automated grading, the prompting techniques employed, the effectiveness of LLM-based grading for both structured and open-ended responses, and the patterns observed in grading performance. Additionally, this paper discusses the challenges associated with LLM-based grading systems, such as inconsistencies and the need for human oversight. By synthesizing existing research, this paper provides insights into the current capabilities of LLMs in academic assessment and serves as a foundation for future exploration in this area.

We train a relative sentence readability estimator from a corpus without absolute sentence readability.Since sentence readability depends on the reader’s knowledge, objective and absolute readability assessments require costly annotation by experts.Therefore, few corpora have absolute sentence readability, while parallel corpora for text simplification with relative sentence readability between two sentences are available for many languages.With multilingual applications in mind, we propose a method to estimate relative sentence readability based on parallel corpora for text simplification.Experimental results on ranking a set of English sentences by readability show that our method outperforms existing unsupervised methods and is comparable to supervised methods based on absolute sentence readability.

This study presents a teacher-centered evaluation of an AI-powered reading comprehension tool, developed to support learners with language-based difficulties. Drawing on the Social Acceptance of Technology (SAT) framework, we investigate not only technical usability but also the pedagogical, ethical, and contextual dimensions of AI integration in classrooms. We explore how teachers perceive the platform’s alignment with inclusive pedagogies, instructional workflows, and professional values through a mixed-methods approach, including questionnaires and focus groups with educators. Findings a shift from initial curiosity to critical, practice-informed reflection, with trust, transparency, and adaptability emerging as central concerns. The study contributes a replicable evaluation framework and highlights the importance of engaging teachers as co-designers in the development of educational technologies.

We release LORuGEC – the first rule-annotated corpus for Russian Grammatical Error Correction. The corpus is designed for diagnostic purposes and contains 348 validation and 612 test sentences specially selected to represent complex rules of Russian writing. This makes our corpus significantly different from other Russian GEC corpora. We apply several large language models and approaches to our corpus, the best F0.5 score of 83% is achieved by 5-shot learning using YandexGPT-5 Pro model.To move further the boundaries of few-shot learning, we are the first to apply a GECTOR-like encoder model for similar examples retrieval. GECTOR-based example selection significantly boosts few-shot performance. This result is true not only for LORuGEC but for other Russian GEC corpora as well. On LORuGEC, the GECTOR-based retriever might be further improved using contrastive tuning on the task of rule label prediction. All these results hold for a broad class of large language models.

Comparative judgment (CJ) is an assessment method in which multiple assessors determine the holistic quality of essays through pairwise comparisons.While CJ is recognized for generating reliable and valid scores, it falls short in providing transparency about the specific quality aspects these holistic scores represent.Our study addresses this limitation by predicting scores on a set of rubrics that measure text quality, thereby explaining the holistic scores derived from CJ.We developed feature-based machine learning models that leveraged complexity and genre features extracted from a collection of Dutch essays.We evaluated the predictability of rubric scores for text quality based on linguistic features.Subsequently, we evaluated the validity of the predicted rubric scores by examining their ability to explain the holistic scores derived from CJ.Our findings indicate that feature-based prediction models can predict relevant rubric scores moderately well. Furthermore, the predictions can be used to explain holistic scores from CJ, despite certain biases. This automated approach to explain holistic quality scores from CJ can enhance the transparency of CJ assessments and simplify the evaluation of their validity.

Automated Essay Scoring (AES) plays a crucial role in assessing language learners’ writingquality, reducing grading workload, and providing real-time feedback. The lack of annotatedessay datasets inhibits the development of Arabic AES systems. This paper leverages LargeLanguage Models (LLMs) and Transformermodels to generate synthetic Arabic essays forAES. We prompt an LLM to generate essaysacross the Common European Framework ofReference (CEFR) proficiency levels and introduce and compare two approaches to errorinjection. We create a dataset of 3,040 annotated essays with errors injected using our twomethods. Additionally, we develop a BERTbased Arabic AES system calibrated to CEFRlevels. Our experimental results demonstratethe effectiveness of our synthetic dataset in improving Arabic AES performance. We makeour code and data publicly available

Locally deployed Small Language Models (SLMs) offer a promising solution for providing timely and effective programming feedback to students learning to code. However, SLMs often produce misleading or hallucinated feedback, limiting their reliability in educational settings. Current approaches for improving SLM feedback rely on existing human annotations or LLM-generated feedback. This paper addresses a fundamental challenge: Can we improve SLMs’ feedback capabilities without relying on human or LLM-generated annotations? We demonstrate that training SLMs on the proxy task of program repair is sufficient to enhance their ability to generate high-quality feedback. To this end, we introduce Direct Repair Optimization (DRO), a self-supervised online reinforcement learning strategy that trains language models to reason about how to efficiently fix students’ programs.Our experiments, using DRO to fine-tune LLaMA-3.1–3B and Qwen-2.5–3B on a large-scale dataset of Python submissions from real students, show substantial improvements on downstream feedback tasks. We release our code to support further research in educational feedback and highlight promising directions for future work.

Interviews are central to the Empathy phase of Design Thinking, helping designers uncover user needs and experience. Although interviews are widely used to support human centered innovation, evaluating their quality, especially from a cognitive perspective, remains underexplored. This study introduces a structured framework for evaluating interview quality in the context of Design Thinking, using Bloom’s Taxonomy as a foundation. We propose the Cognitive Interview Quality Score, a composite metric that integrates three dimensions: Effectiveness Score, Bloom Coverage Score, and Distribution Balance Score. Using human-annotations, we assessed 15 interviews across three domains to measure cognitive diversity and structure. We compared CIQS-based rankings with human experts and found that the Bloom Coverage Score aligned more closely with expert judgments. We evaluated the performance of LMA-3-8B-Instruct and GPT-4o-mini, using zero-shot, few-shot, and chain-of-thought prompting, finding GPT-4o-mini, especially in zero-shot mode, showed the highest correlation with human annotations in all domains. Error analysis revealed that models struggled more with mid-level cognitive tasks (e.g., Apply, Analyze) and performed better on Create, likely due to clearer linguistic cues. These findings highlight both the promise and limitations of using NLP models for automated cognitive classification and underscore the importance of combining cognitive metrics with qualitative insights to comprehensively assess interview quality.

Easy language and text simplification are currently topical research questions, with important applications in many contexts, and with various approaches under active investigation, including prompt-based methods. The estimation of the level of difficulty of a text becomes a crucial challenge when the estimator is employed in a simplification workflow as a quality-control mechanism. It can act as a critic in frameworks where it can guide other models, which are responsible for generating text at a specified level of difficulty, as determined by the user’s needs.We present our work in the context of simplified Finnish. We discuss problems in collecting corpora for training models for estimation of text difficulty, and our experiments with estimation models.The results of the experiments are promising: the models appear usable both for assessment and for deployment as a component in a larger simplification framework.

Large language models (LLMs) are increasingly being adopted in educational settings. These applications expand beyond English, though current LLMs remain primarily English-centric. In this work, we ascertain if their use in education settings in non-English languages is warranted. We evaluated the performance of popular LLMs on four educational tasks: identifying student misconceptions, providing targeted feedback, interactive tutoring, and grading translations in eight languages (Mandarin, Hindi, Arabic, German, Farsi, Telugu, Ukrainian, Czech) in addition to English. We find that the performance on these tasks somewhat corresponds to the amount of language represented in training data, with lower-resource languages having poorer task performance. However, at least some models are able to more or less maintain their levels of performance across all languages. Thus, we recommend that practitioners first verify that the LLM works well in the target language for their educational task before deployment.

Vocabulary use is a fundamental aspect of second language (L2) proficiency. To date, its assessment by automated systems has typically examined the context-independent, or part-of-speech (PoS) related use of words. This paper introduces a novel approach to enable fine-grained vocabulary evaluation exploiting the precise use of words within a sentence. The scheme combines large language models (LLMs) with the English Vocabulary Profile (EVP). The EVP is a standard lexical resource that enables in-context vocabulary use to be linked with proficiency level. We evaluate the ability of LLMs to assign proficiency levels to individual words as they appear in L2 learner writing, addressing key challenges such as polysemy, contextual variation, and multi-word expressions. We compare LLMs to a PoS-based baseline. LLMs appear to exploit additional semantic information that yields improved performance.We also explore correlations between word-level proficiency and essay-level proficiency. Finally, the approach is applied to examine the consistency of the EVP proficiency levels. Results show that LLMs are well-suited for the task of vocabulary assessment.

Question Generation (QG), the task of automatically generating questions from a source input, has seen significant progress in recent years. Difficulty-controllable QG (DCQG) enables control over the difficulty level of generated questions while considering the learner’s ability. Additionally, narrative-controllable QG (NCQG) allows control over the narrative aspects embedded in the questions. However, research in QG lacks a focus on combining these two types of control, which is important for generating questions tailored to educational purposes. To address this gap, we propose a strategy for Joint Narrative and Difficulty Control, enabling simultaneous control over these two attributes in the generation of reading comprehension questions. Our evaluation provides preliminary evidence that this approach is feasible, though it is not effective across all instances. Our findings highlight the conditions under which the strategy performs well and discuss the trade-offs associated with its application.

We present the framework Omethi, which is aimed at scoring short text responses in a semi-automatic fashion, particularly fit to international large-scale assessments. We evaluate its effectiveness for the massively multilingual PISA tests. Responses are passed through a conditional flow of hierarchically combined scoring components to assign a score. Once a score is assigned, hierarchically lower components are discarded. Models implemented in this study ranged from lexical matching of normalized texts—with excellent accuracy but weak generalizability—to fine-tuned large language models—with lower accuracy but high generalizability. If not scored by any automatic component, responses are passed on to manual scoring. The paper is the first to provide an evaluation of automatic scoring on multilingual PISA data in eleven languages (including Arabic, Finnish, Hebrew, and Kazakh) from three domains (_n_ = 3.8 million responses). On average, results show a manual effort reduction of 71 percent alongside an agreement of _κ_ = .957, when including manual scoring, and _κ_ = .804 for only the automatically scored responses. The evaluation underscores the framework’s effective adaptivity and operational feasibility with its shares of used components varying substantially across domains and languages while maintaining homogeneously high accuracy.

Advances in Large Language Models (LLMs) have sparked growing interest in their potential as explainable text evaluators. While LLMs have shown promise in assessing machine-generated texts in tasks such as summarization and machine translation, their effectiveness in evaluating human-written content—such as student writing in classroom settings—remains underexplored. In this paper, we investigate LLM-based specificity assessment of student reflections written in response to prompts, using three instruction-tuned models. Our findings indicate that although LLMs may underperform compared to simpler supervised baselines in terms of scoring accuracy, they offer a valuable interpretability advantage. Specifically, LLMs can generate user-friendly explanations that enhance the transparency and usability of automated specificity scoring systems.

Identifying instances of first language (L1) transfer – the application of the linguistics structures of a speaker’s first language to their second language(s) – can facilitate second language (L2) learning as it can inform learning and teaching resources, especially when instances of negative transfer (that is, interference) can be identified. While studies of transfer between two languages A and B require a priori linguistic structures to be analyzed with three datasets (data from L1 speakers of language A, L1 speakers of language B, and L2 speakers of A or B), native language identification (NLI) – a machine learning task to predict one’s L1 based on one’s L2 production – has the advantage to detect instances of subtle and unpredicted transfer, casting a “wide net” to capture patterns of transfer that were missed before (Jarvis and Crossley, 2018). This study aims to apply NLI tasks to find potential instances of transfer of collocations. Our results, compared to previous transfer studies, indicate that NLI can be used to reveal collocation transfer, also in understudied L2 languages.

Across cultures, names tell a lot about their bearers as they carry deep personal, historical, and cultural significance. Names have also been found to serve as powerful signals of gender, race, and status in the social hierarchy–a pecking order in which individual positions shape others’ expectations on their perceived competence and worth (Podolny, 2005). With the widespread adoption of Large Language Models (LLMs) in education and given that names are often an input for LLMs, it is crucial to evaluate whether LLMs may sort students into status positions based on first and last names and, if so, whether it is in an unfair, biased fashion. While prior work has primarily investigated biases in first names, little attention has been paid to last names and even less to the combined effects of first and last names. In this study, we conduct a large-scale analysis with bootstrap standard errors of 45,000 name variations across 5 ethnicities to examine how AI-generated responses exhibit systemic name biases. Our study investigates three key characteristics of inequality and finds that LLMs reflect, construct, and reinforce status hierarchies based on names that signal gender and ethnicity as they encode differential expectations of competence, leadership, and economic potential. Contrary to the common assumption that AI tends to favor Whites, we show that East and, in some contexts, South Asian names receive higher rankings. We also disaggregate Asians, a population projected to be the largest immigrant group in the U.S. by 2055. Our results challenge the monolithic Asian model minority assumption, illustrating a more complex and stratified model of bias. Additionally, spanning cultural categories by adopting Western first names improves AI-perceived status for East and Southeast Asian students, particularly for girls. Our findings underscore the importance of intersectional and more nuanced understandings of race, gender, and mixed identities in the evaluation of LLMs, rather than relying on broad, monolithic, and mutually exclusive categories. By examining LLM bias and discrimination in our multicultural contexts, our study illustrates potential harms of using LLMs in education as they do not merely reflect implicit biases but also actively construct new social hierarchies that can unfairly shape long-term life trajectories. An LLM that systematically assigns lower grades or subtly less favorable evaluations to students with certain name signals reinforces a tiered system of privilege and opportunity. Some groups may face structural disadvantages, while others encounter undue pressure from inflated expectations.

This study investigates the use of ChatGPT for Automated Essay Scoring (AES) in assessing Italian middle school students’ written texts. Using rubrics targeting grammar, coherence and argumentation, we compare AI-generated feedback with that of a human teacher on a newly collected corpus of students’ essays. Despite some differences, ChatGPT provided detailed and timely feedback that complements the teacher’s role. These findings underscore the potential of generative AI to improve the assessment of writing, providing useful insights for educators and supporting students in developing their writing skills.

Engaging students in a coherent classroom discussion is one aspect of high-quality instruction and is an important skill that requires practice to acquire. With the goal of providing teachers with formative feedback on their classroom discussions, we investigate automated means for evaluating teachers’ ability to lead coherent discussions in simulated classrooms. While prior work has shown the effectiveness of large language models (LLMs) in assessing the coherence of relatively short texts, it has also found that LLMs struggle when assessing instructional quality. We evaluate the generalizability of task formulation strategies for assessing the coherence of classroom discussions across different subject domains using GPT-4o and discuss how these formulations address the previously reported challenges—the overestimation of instructional quality and the inability to extract relevant parts of discussions. Finally, we report lack of generalizability across domains and the misalignment with humans in the use of evidence from discussions as remaining challenges.

Understanding how linguistic topics are related to each another is essential for designing effective and adaptive second-language (L2) instruction. We present a data-driven framework to model topic dependencies and their difficulty within a L2 learning curriculum. First, we estimate topic difficulty and student ability using a three-parameter Item Response Theory (IRT) model. Second, we construct topic-level knowledge graphs—as directed acyclic graphs (DAGs)—to capture the prerequisite relations among the topics, comparing a threshold-based method with the statistical Grow-Shrink Markov Blanket algorithm. Third, we evaluate the alignment between IRT-inferred topic difficulty and the structure of the graphs using edge-level and global ordering metrics. Finally, we compare the IRT-based estimates of learner ability with assessments of the learners provided by teachers to validate the model’s effectiveness in capturing learner proficiency. Our results show a promising agreement between the inferred graphs, IRT estimates, and human teachers’ assessments, highlighting the framework’s potential to support personalized learning and adaptive curriculum design in intelligent tutoring systems.

Recent advancements in natural language processing, particularly large language models (LLMs), are making the automated evaluation of classroom discussions more achievable. In this work, we propose a method to improve the performance of LLMs on classroom discussion quality assessment by utilizing in-context learning (ICL) example retrieval. Specifically, we leverage example re-ranking and label ratio regulation, which forces a specific ratio of different types of examples on the ICL examples.While a standard ICL example retrieval approach shows inferior performance compared to using a predetermined set of examples, our approach improves performance in all tested dimensions. We also conducted experiments to examine the ineffectiveness of the generic ICL example retrieval approach and found that the lack of positive and hard negative examples can be a potential cause. Our analyses emphasize the importance of maintaining a balanced distribution of classes (positive, non-hard negative, and hard negative examples) in creating a good set of ICL examples, especially when we can utilize educational knowledge to identify instances of hard negative examples.

Tutoring dialogues have gained significant attention in recent years, given the prominence of online learning and the emerging tutoring abilities of artificial intelligence (AI) agents powered by large language models (LLMs). Recent studies have shown that the strategies used by tutors can have significant effects on student outcomes, necessitating methods to predict how tutors will behave and how their actions impact students. However, few works have studied predicting tutor strategy in dialogues. Therefore, in this work we investigate the ability of modern LLMs, particularly Llama 3 and GPT-4o, to predict both future tutor moves and student outcomes in dialogues, using two math tutoring dialogue datasets. We find that even state-of-the-art LLMs struggle to predict future tutor strategy while tutor strategy is highly indicative of student outcomes, outlining a need for more powerful methods to approach this task.

This paper presents a data-driven analysis of Romanian secondary school textbooks through the lens of Bloom’s Taxonomy, focusing on the promotion of critical thinking in instructional design. Using the ROTEX corpus, we extract and annotate almost 2 million words of Romanian Language and Literature textbooks (grades 5-8) with Bloom-aligned labels for verbs associated with pedagogical tasks. Our annotation pipeline combines automatic verb extraction, human filtering based on syntactic form and task relevance, and manual assignment of Bloom labels supported by in-text concordance checks. The resulting dataset enables fine-grained analysis of task complexity both across and within textbooks and grade levels. Our findings reveal a general lack of structured cognitive progression across most textbook series. We also propose a multi-dimensional framework combining cognitive-level and linguistic evaluation to assess instructional design quality. This work contributes annotated resources and reproducible methods for NLP-based educational content analysis in low-resource languages.

This paper presents a strategy for improving AI assistants embedded in short e-learning courses. The proposed method is implemented within a Retrieval-Augmented Generation (RAG) architecture and evaluated using several retrieval variants. The results show that query quality improves when the knowledge base is enriched with definitions of key concepts discussed in the course. Our main contribution is a lightweight enhancement approach that increases response quality without overloading the course with additional instructional content.

This paper presents an LLM-powered approach for generating concept maps to enhance digital reading comprehension in higher education. While particularly focused on supporting neurodivergent students with their distinct information processing patterns, this approach benefits all learners facing the cognitive challenges of digital text. We use GPT-4o-mini to extract concepts and relationships from educational texts across ten diverse disciplines using open-domain prompts without predefined categories or relation types, enabling discipline-agnostic extraction. Section-level processing achieved higher precision (83.62%) in concept extraction, while paragraph-level processing demonstrated superior recall (74.51%) in identifying educationally relevant concepts. We implemented an interactive web-based visualization tool https://simplified-cognitext.streamlit.app that transforms extracted concepts into navigable concept maps. User evaluation (n=14) showed that participants experienced a 31.5% reduction in perceived cognitive load when using concept maps, despite spending more time with the visualization (22.6% increase). They also completed comprehension assessments more efficiently (14.1% faster) with comparable accuracy. This work demonstrates that LLM-based concept mapping can significantly reduce cognitive demands while supporting non-linear exploration.

Correct verb placement is difficult to acquire for second-language learners of Germanic languages. However, word order errors and, consequently, verb placement errors, are heavily underrepresented in benchmark datasets of NLP tasks such as grammatical error detection/correction and linguistic acceptability assessment. If they are present, they are most often naively introduced, or classification occurs at the sentence level, preventing the precise identification of individual errors and the provision of appropriate feedback to learners. To remedy this, we present GermDetect: Universal Dependencies-based, linguistically informed verb placement error detection datasets for learners of Germanic languages, designed as a token classification task. As our datasets are UD-based, we are able to provide them in most major Germanic languages: Afrikaans, German, Dutch, Faroese, Icelandic, Danish, Norwegian (Bokmål and Nynorsk), and Swedish.We train multilingual BERT models on GermDetect and show that linguistically informed, UD-based error induction results in more effective models for verb placement error detection than models trained on naively introduced errors. Finally, we conduct ablation studies on multilingual training and find that lower-resource languages benefit from the inclusion of structurally related languages in training.

This study examines vulnerabilities in transformer-based automated short-answer grading systems used in medical education, with a focus on how these systems can be manipulated through adversarial gaming strategies. Our research identifies three main types of gaming strategies that exploit the system’s weaknesses, potentially leading to false positives. To counteract these vulnerabilities, we implement several adversarial training methods designed to enhance the system’s robustness. Our results indicate that these methods significantly reduce the susceptibility of grading systems to such manipulations, especially when combined with ensemble techniques like majority voting and Ridge regression, which further improve the system’s defense against sophisticated adversarial inputs. Additionally, employing large language models suchasGPT-4with varied prompting techniques has shown promise in recognizing and scoring gaming strategies effectively. The findings underscore the importance of continuous improvements in AI-driven educational tools to ensure their reliability and fairness in high-stakes settings.

Recent advances in LLMs offer new opportunities for supporting student writing, particularly through real-time, composition-level feedback. However, for such support to be effective, LLMs need to generate text completions that align with the writer’s internal representation of their developing message, a representation that is often implicit and difficult to observe. This paper investigates the use of eye-tracking data, specifically lookback fixations during pauses in text production, as a cue to this internal representation. Using eye movement data from students composing texts, we compare human-generated completions with LLM-generated completions based on prompts that either include or exclude words and sentences fixated during pauses. We find that incorporating lookback fixations enhances human-LLM alignment in generating text completions. These results provide empirical support for generating fixation-aware LLM feedback and lay the foundation for future educational tools that deliver real-time, composition-level feedback grounded in writers’ attention and cognitive processes.

We present a method for labeling spans of text with large language models (LLMs) and apply it to the task of identifying shell language, language which plays a structural or connective role without constituting the main content of a text. We compare several recent LLMs by evaluating their “annotations” against a small human-curated test set, and train a smaller supervised model on thousands of LLM-annotated examples. The described method enables workflows that can learn complex or nuanced linguistic phenomena without tedious, large-scale hand-annotations of training data or specialized feature engineering.

Large language models (LLMs) hold great promise for educational applications, particularly in intelligent tutoring systems. However, effective tutoring requires alignment with pedagogical strategies – something current LLMs lack without task-specific adaptation. In this work, we explore whether fine-grained annotation of teacher intents can improve the quality of LLM-generated tutoring responses. We focus on MathDial, a dialog dataset for math instruction, and apply an automated annotation framework to re-annotate a portion of the dataset using a detailed taxonomy of eleven pedagogical intents. We then fine-tune an LLM using these new annotations and compare its performance to models trained on the original four-category taxonomy. Both automatic and qualitative evaluations show that the fine-grained model produces more pedagogically aligned and effective responses. Our findings highlight the value of intent specificity for controlled text generation in educational settings, and we release our annotated data and code to facilitate further research.

Large Language Models (LLMs) offer exciting potential as educational tutors, and much research explores this potential. Unfortunately, there’s little research in understanding the baseline behavioral pattern differences that LLM tutors exhibit, in contrast to human tutors. We conduct a preliminary study of these differences with the CIMA dataset and three state-of-the-art LLMs (GPT-4o, Gemini Pro 1.5, and LLaMA 3.1 450B). Our results reveal systematic deviations in these baseline patterns, particulary in the tutoring actions selected, complexity of responses, and even within different LLMs. This research brings forward some early results in understanding how LLMs when deployed as tutors exhibit systematic differences, which has implications for educational technology design and deployment. We note that while LLMs enable more powerful and fluid interaction than previous systems, they simultaneously develop characteristic patterns distinct from human teaching. Understanding these differences can inform better integration of AI in educational settings.

Large Language Models (LLMs) have shown impressive performance in mathematical reasoning tasks when guided by Chain-of-Thought (CoT) prompting. However, they tend to produce highly confident yet incorrect outputs, which poses significant risks in domains like education, where users may lack the expertise to assess reasoning steps. To address this, we propose a structured framework that models stepwise confidence as a temporal signal and evaluates it using Signal Temporal Logic (STL). In particular, we define formal STL-based constraints to capture desirable temporal properties and compute robustness scores that serve as structured, interpretable confidence estimates. Our approach also introduces a set of uncertainty reshaping strategies to enforce smoothness, monotonicity, and causal consistency across the reasoning trajectory. Experiments show that our approach consistently improves calibration metrics and provides more reliable uncertainty estimates than conventional confidence aggregation and post-hoc calibration.

This paper presents an automated scoring approach for a formative assessment tool aimed at helping learner physicians enhance their communication skills through simulated patient interactions. The system evaluates transcribed learner responses by detecting key communicative behaviors, such as acknowledgment, empathy, and clarity. Built on an adapted version of the ACTA scoring framework, the model achieves a mean binary F1 score of 0.94 across 8 clinical scenarios. A central contribution of this work is the investigation of how to balance scoring accuracy with scalability. We demonstrate that synthetic training data offers a promising path toward reducing reliance on large, annotated datasets—making automated scoring more accurate and scalable.

This study presents a computational analysis to classify actionability in teacher feedback. We fine-tuned a RoBERTa model on 662 manually annotated feedback examples from West African classrooms, achieving strong classification performance (accuracy = 0.94, precision = 0.90, recall = 0.96, f1 = 0.93). This enabled classification of over 12,000 feedback instances. A comparison of linguistic features indicated that actionable feedback was associated with lower word count but higher readability, greater lexical diversity, and more modifier usage. These findings suggest that concise, accessible language with precise descriptive terms may be more actionable for teachers. Our results support focusing on clarity in teacher observation protocols while demonstrating the potential of computational approaches in analyzing educational feedback at scale.

This paper presents EduCSW, a novel pipeline for generating Mandarin-English code-switched text to support AI-powered educational tools that adapt computer science instruction to learners’ language proficiency through mixed-language delivery. To address the scarcity of code-mixed datasets, we propose an encoder-decoder architecture that generates natural code-switched text using only minimal existing code-mixed examples and parallel corpora. Evaluated on a corpus curated for computer science education, human annotators rated 60–64% of our model’s outputs as natural, significantly outperforming both a baseline fine-tuned neural machine translation (NMT) model (22–24%) and the DeepSeek-R1 model (34–44%). The generated text achieves a Code-Mixing Index (CMI) of 25.28%, aligning with patterns observed in spontaneous Mandarin-English code-switching. Designed to be generalizable across language pairs and domains, this pipeline lays the groundwork for generating training data to support the development of educational tools with dynamic code-switching capabilities.

The advent of artificial intelligence (AI) has marked a transformative era in educational measurement and evaluation, particularly in the development of assessment items. Large language models (LLMs) have emerged as promising tools for scalable automatic item generation (AIG), yet concerns remain about the validity of AI-generated items in various domains. To address this issue, we propose STAIR-AIG (Systematic Tool for Assessment Item Review in Automatic Item Generation), a human-in-the-loop framework that integrates expert judgment to optimize the quality of AIG items. To explore the functionality of the tool, AIG items were generated in the domain of critical thinking. Subsequently, the human expert and four OpenAI LLMs conducted a review of the AIG items. The results show that while the LLMs demonstrated high consistency in their rating of the AIG items, they exhibited a tendency towards leniency. In contrast, the human expert provided more variable and strict evaluations, identifying issues such as the irrelevance of the construct and cultural insensitivity. These findings highlight the viability of STAIR-AIG as a structured human-AI collaboration approach that facilitates rigorous item review, thus optimizing the quality of AIG items. Furthermore, STAIR-AIG enables iterative review processes and accumulates human feedback, facilitating the refinement of models and prompts. This, in turn, would establish a more reliable and comprehensive pipeline to improve AIG practices.

We present UPSC2M, a large-scale dataset comprising two million multiple-choice question attempts from over 46,000 students, spanning nearly 9,000 questions across seven subject areas. The questions are drawn from the Union Public Service Commission (UPSC) examination, one of India’s most competitive and high-stakes assessments. Each attempt includes both response correctness and time taken, enabling fine-grained analysis of learner behavior and question characteristics. Over this dataset, we define two core benchmark tasks: question difficulty estimation and student performance prediction. The first task involves predicting empirical correctness rates using only question text. The second task focuses on predicting the likelihood of a correct response based on prior interactions. We evaluate simple baseline models on both tasks to demonstrate feasibility and establish reference points. Together, the dataset and benchmarks offer a strong foundation for building scalable, personalized educational systems. We release the dataset and code to support further research at the intersection of content understanding, learner modeling, and adaptive assessment.

Despite recent advances in AI detection methods, their practical application, especially in education, remains limited. Educators need functional tools pointing to AI indicators within texts, rather than merely estimating whether AI was used. GPTZero’s new AI Vocabulary feature, which highlights parts of a text likely to be AI-generated based on frequent words and phrases from LLM-generated texts, offers a potential solution. However, its effectiveness has not yet been empirically validated.In this study, we examine whether GPTZero’s AI Vocabulary can effectively distinguish between LLM-generated and student-written essays. We analyze the AI Vocabulary lists published from October 2024 to March 2025 and evaluate them on a subset of the Ghostbuster dataset, which includes student and LLM essays. We train multiple Bag-of-Words classifiers using GPTZero’s AI Vocabulary terms as features and examine their individual contributions to classification.Our findings show that simply checking for the presence, not the frequency, of specific AI terms yields the best results, particularly with ChatGPT-generated essays. However, performance drops to near-random when applied to Claude-generated essays, indicating that GPTZero’s AI Vocabulary may not generalize well to texts generated by LLMs other than ChatGPT. Additionally, all classifiers based on GPTZero’s AI Vocabulary significantly underperform compared to Bag-of-Words classifiers trained directly on the full dataset vocabulary. These findings suggest that fixed vocabularies based solely on lexical features, despite their interpretability, have limited effectiveness across different LLMs and educational writing contexts.

We present a case study on building task-specific models for grammatical error correction and explanation generation tailored to learners of Estonian. Our approach handles whole paragraphs instead of sentences and leverages prompting proprietary large language models for generating synthetic training data, addressing the limited availability of error correction data and the complete absence of correction justification/explanation data in Estonian. We describe the chosen approach and pipeline and provide technical details for the experimental part. The final outcome is a set of open-weight models, which are released with a permissive license along with the generated synthetic error correction and explanation data.

Automated item generation (AIG) is a key enabler for scaling language proficiency assessments. We present an end-to-end methodology for automated generation, annotation, and integration of adaptive writing items for the EF Standard English Test (EFSET), leveraging recent advances in large language models (LLMs). Our pipeline uses few-shot prompting with state-of-the-art LLMs to generate diverse, proficiency-aligned prompts, rigorously validated by expert reviewers. For robust scoring, we construct a synthetic response dataset via majority-vote LLM annotation and fine-tune a LLaMA 3.1 (8B) model. For each writing item, a range of proficiency-aligned synthetic responses, designed to emulate authentic student work, are produced for model training and evaluation. These results demonstrate substantial gains in scalability and validity, offering a replicable framework for next-generation adaptive language testing.

Communicative practice is critical for second language development, yet learners often lack targeted, engaging opportunities to use new grammar structures. While large language models (LLMs) can offer coherent interactions, they are not inherently aligned with pedagogical goals or proficiency levels. In this paper, we explore how LLMs can be integrated into a structured framework for contextually-constrained, grammar-focused interaction, building on an existing goal-oriented dialogue system. Through controlled simulations, we evaluate five LLMs across 75 A2-level tasks under two conditions: (i) grammar-targeted, task-anchored prompting and (ii) the addition of a lightweight post-generation validation pipeline using a grammar annotator.Our findings show that template-based prompting alone substantially increases target-form coverage up to 91.4% for LLaMA 3.1-70B-Instruct, while reducing overly advanced grammar usage. The validation pipeline provides an additional boost in form-focused tasks, raising coverage to 96.3% without significantly degrading appropriateness.

Large Language Models (LLMs) are increasingly used as proxy students in the development of Intelligent Tutoring Systems (ITSs) and in piloting test questions. However, to what extent these proxy students accurately emulate the behavior and characteristics of real students remains an open question. To investigate this, we collected a dataset of 489 items from the National Assessment of Educational Progress (NAEP), covering mathematics and reading comprehension in grades 4, 8, and 12. We then apply an Item Response Theory (IRT) model to position 11 diverse and state-of-the-art LLMs on the same ability scale as real student populations. Our findings reveal that, without guidance, strong general-purpose models consistently outperform the average student at every grade, while weaker or domain-mismatched models may align incidentally. Using grade-enforcement prompts changes models’ performance, but whether they align with the average grade-level student remains highly model- and prompt-specific: no evaluated model–prompt pair fits the bill across subjects and grades, underscoring the need for new training and evaluation strategies. We conclude by providing guidelines for the selection of viable proxies based on our findings. All related code and data have been made available (https://github.com/kvadityasrivatsa/IRT-for-LLMs-as-Students).

This paper details an LLM-assisted system designed to support curriculum writing within a Finnish higher education institution. Developed over 18 months through iterative prototyping, workshops, and user testing with faculty, the tool functions as a collaborative partner. It provides structured suggestions and analyzes course content for alignment with institutional goals and standards like UN SDGs, aiming to reduce educator cognitive load while keeping humans central to the process. The paper presents the system’s technical architecture, findings from user feedback (including quotes and evaluation metrics), and discusses its potential to aid complex educational planning compared to generic AI tools.

This shared task has aimed to assess pedagogical abilities of AI tutors powered by large language models (LLMs), focusing on evaluating the quality of tutor responses aimed at student’s mistake remediation within educational dialogues. The task consisted of five tracks designed to automatically evaluate the AI tutor’s performance across key dimensions of mistake identification, precise location of the mistake, providing guidance, and feedback actionability, grounded in learning science principles that define good and effective tutor responses, as well as the track focusing on detection of the tutor identity. The task attracted over 50 international teams across all tracks. The submitted models were evaluated against gold-standard human annotations, and the results, while promising, show that there is still significant room for improvement in this domain: the best results for the four pedagogical ability assessment tracks range between macro F1 scores of 58.34 (for providing guidance) and 71.81 (for mistake identification) on three-class problems, with the best F1 score in the tutor identification track reaching 96.98 on a 9-class task. In this paper, we overview the main findings of the shared task, discuss the approaches taken by the teams, and analyze their performance. All resources associated with this task are made publicly available to support futureresearch in this critical domain (https://github.com/kaushal0494/UnifyingAITutorEvaluation/tree/main/BEA_Shared_Task_2025_Datasets).

With the rapid development of smart education, educational conversation systems have become an important means to support personalized learning. Identifying tutors and understanding their unique teaching style are crucial to optimizing teaching quality. However, accurately identifying tutors from multi-round educational conversation faces great challenges due to complex contextual semantics, long-term dependencies, and implicit pragmatic relationships. This paper proposes a dual-tower encoding architecture to model the conversation history and tutor responses separately, and enhances semantic fusion through four feature interaction mechanisms. To further improve the robustness, this paper adopts a model ensemble voting strategy based on five-fold cross-validation. Experiments on the BEA 2025 shared task dataset show that our method achieves 89.65% Marco-F1 in tutor identification, ranks fourth among all teams(4/20), demonstrating its effectiveness and potential in educational AI applications.We have made the corresponding code publicly accessible at https://github.com/leibnizchen/Dual-Encoder.

The potential of large language models (LLMs) as AI tutors to facilitate student learning has garnered significant interest, with numerous studies exploring their efficacy in educational contexts. Notably, Wang and Chen (2025) suggests that the relationship between AI model performance and educational outcomes may not always be positively correlated; less accurate AI models can sometimes achieve similar educational impacts to their more accurate counterparts if designed into learning activities appropriately. This underscores the need to evaluate the pedagogical capabilities of LLMs across various dimensions, empowering educators to select appropriate dimensions and LLMs for specific analyses and instructional activities. Addressing this imperative, the BEA 2025 workshop initiated a shared task aimed at comprehensively assessing the pedagogical potential of AI-powered tutors. In this task, our team employed parameter-efficient fine-tuning (PEFT) on Llama-3.2-3B to automatically assess the quality of feedback generated by LLMs in student-teacher dialogues, concentrating on mistake identification, mistake location, guidance provision, and guidance actionability. The results revealed that the fine-tuned Llama-3.2-3B demonstrated notable performance, especially in mistake identification, mistake location, and guidance actionability, securing a top-ten ranking across all tracks. These outcomes highlight the robustness and significant promise of the PEFT method in enhancing educational dialogue analysis.

The growing use of large language models (LLMs) for AI-powered tutors in education highlights the need for reliable evaluation of their pedagogical abilities. In this work, we propose a reasoning-based evaluation methodology that leverages pedagogical domain knowledge to assess LLM-generated feedback in mathematical dialogues while providing insights into why a particular evaluation is given. We design structured prompts to invoke pedagogically-informed reasoning from LLMs and compare base model candidates selected for their strengths in reasoning, mathematics, and overall instruction-following. We employ Group Relative Policy Optimization (GRPO), a reinforcement learning method known to improve reasoning performance, to train models to perform evaluation in four pedagogically motivated dimensions, Mistake Identification, Mistake Location, Providing Guidance, and Actionability. Experimental results show that our GRPO-based models consistently outperform the base model and GPT-4.1, and surpass models trained using supervised fine-tuning in three out of four dimensions. Notably, our method achieved top-ranked performance in Actionability and competitive performance in two other dimensions in the BEA 2025 Shared Task under the team name bea-jh, underscoring the value of generating pedagogically grounded rationales for improving the quality of educational feedback evaluation.

Automatic evaluation of AI tutor responses in educational dialogues is a challenging task, requiring accurate identification of mistakes and the provision of pedagogically effective guidance. In this paper, we propose a classification model based on BERT, enhanced with a cross-attention mechanism that explicitly models the interaction between the tutor’s response and preceding dialogue turns. This design enables better alignment between context and response, supporting more accurate assessment along the educational dimensions defined in the BEA 2025 Shared Task. To address the substantial class imbalance in the dataset, we employ data augmentation techniques for minority classes. Our system consistently outperforms baseline models across all tracks. However, performance on underrepresented labels remains limited, particularly when distinguishing between semantically similar cases. This suggests room for improvement in both model expressiveness and data coverage, motivating future work with stronger decoder-only models and auxiliary information from systems like GPT-4.1. Overall, our findings offer insights into the potential and limitations of LLM-based approaches for pedagogical feedback evaluation.

We present a prompt-based evaluation framework for assessing AI-generated math tutoring responses across four pedagogical dimensions: mistake identification, mistake location, guidance quality, and actionability. Our approach leverages task-aware prompt tuning on a large language model, supplemented by data augmentation techniques including dialogue shuffling and class-balanced downsampling. In experiments on the BEA 2025 Shared Task benchmark, our system achieved first place in mistake identification and strong top-five rankings in the other tracks. These results demonstrate the effectiveness of structured prompting and targeted augmentation for enhancing LLMs’ ability to provide pedagogically meaningful feedback.

In this paper, we propose a novel framework for the tutor identification track of the BEA 2025 shared task (Track 5). Our framework integrates data-algorithm co-design, dynamic class compensation, and structured prediction optimization. Specifically, our approach employs noise augmentation, a fine-tuned DeBERTa-v3-small model with inverse-frequency weighted loss, and Hungarian algorithm-based label assignment to address key challenges, such as severe class imbalance and variable-length dialogue complexity. Our method achieved 0.969 Macro-F1 score on the official test set, securing second place in this competition. Ablation studies revealed significant improvements: a 9.4% gain in robustness from data augmentation, a 5.3% boost in minority-class recall thanks to the weighted loss, and a 2.1% increase in Macro-F1 score through Hungarian optimization. This work advances the field of educational AI by providing a solution for tutor identification, with implications for quality control in LLM-assisted learning environments.

This paper describes our approaches for the BEA-2025 Shared Task on assessing pedagogical ability and attributing tutor identities in AI-powered tutoring systems. We explored three methodological paradigms: in-context learning (ICL), supervised fine-tuning (SFT), and reinforcement learning from human feedback (RLHF). Results indicate clear methodological strengths: SFT is highly effective for structured classification tasks such as mistake identification and feedback actionability, while ICL with advanced prompting excels at open-ended tasks involving mistake localization and instructional guidance. Additionally, fine-tuned models demonstrated strong performance in identifying tutor authorship. Our findings highlight the importance of aligning methodological strategy and task structure, providing insights toward more effective evaluations of educational AI systems.

As Large Language Models (LLMs) are increasingly deployed in educational environments, two critical challenges emerge: identifying the source of tutoring responses and evaluating their pedagogical effectiveness. This paper presents our comprehensive approach to the BEA 2025 Shared Task, addressing both tutor identity classification (Track 5) and actionability assessment (Track 4) in mathematical tutoring dialogues. For tutor identity classification, we distinguish between human tutors (expert/novice) and seven distinct LLMs using cross-response context augmentation and ensemble techniques. For actionability assessment, we evaluate whether responses provide clear guidance on student next steps using selective attention masking and instruction-guided training. Our multi-task approach combines transformer-based models with innovative contextual feature engineering, achieving state-of-the-art performance with a CV macro F1 score of 0.9596 (test set 0.9698) for identity classification and 0.655 (test set Strict F1 0.6906) for actionability assessment. We were able to score rank 5th in Track 4 and rank 1st in Track 5. Our analysis reveals that despite advances in human-like responses, LLMs maintain detectable fingerprints while showing varying levels of pedagogical actionability, with important implications for educational technology development and deployment.

For the BEA 2025 shared task on pedagogi- cal ability assessment, we introduce LUCERA (Lexical Understanding for Cue Density–Based Escalation and Reflective Assessment), a rubric-grounded evaluation framework for sys- tematically analyzing tutor responses across configurable pedagogical dimensions. The ar- chitecture comprises three core components: (1) a rubric-guided large language model (LLM) agent that performs lexical and dialogic cue extraction in a self-reflective, goal-driven manner; (2) a cue-complexity assessment and routing mechanism that sends high-confidence cases to a fine-tuned T5 classifier and esca- lates low-confidence or ambiguous cases to a reasoning-intensive LLM judge; and (3) an LLM-as-a-judge module that performs struc- tured, multi-step reasoning: (i) generating a domain-grounded reference solution, (ii) iden- tifying conceptual, procedural and cognitive gaps in student output, (iii) inferring the tutor’s instructional intent, and (iv) applying the rubric to produce justification-backed classifications. Results show that this unique combination of LLM powered feature engineering, strategic routing and rubrics for grading, enables com- petitive performance without sacrificing inter- pretability and cost effectiveness.

This paper presents the approach and findings of Averroes Team in the BEA 2025 Shared Task Track 1: Mistake Identification. Our system uses the multilingual understanding capabilities of general text embedding models. Our approach involves full-model fine-tuning, where both the pre-trained language model and the classification head are optimized to detect tutor recognition of student mistakes in educational dialogues. This end-to-end training enables the model to better capture subtle pedagogical cues, leading to improved contextual understanding. Evaluated on the official test set, our system achieved an exact macro-F_1 score of 0.7155 and an accuracy of 0.8675, securing third place among the participating teams. These results underline the effectiveness of task-specific optimization in enhancing model sensitivity to error recognition within interactive learning contexts.

The rapid adoption of AI in educational technology is changing learning settings, making the thorough evaluation of AI tutor pedagogical performance is quite important for promoting student success. This paper describes our solution for the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered tutors, which assesses tutor replies over several pedagogical dimensions. We developed transformer-based approaches for five diverse tracks: mistake identification, mistake location, providing guidance, actionability, and tutor identity prediction using the MRBench dataset of mathematical dialogues. We evaluated several pre-trained models including DeBERTa-V3, RoBERTa-Large, SciBERT, and EduBERT. Our approach addressed class imbalance problems by incorporating strategic fine-tuning with weighted loss functions. The findings show that, for all tracks, DeBERTa architectures have higher performances than the others, and our models have obtained in the competitive positions, including 9th of Tutor Identity (Exact F1 of 0.8621), 16th of Actionability (Exact F1 of 0.6284), 19th of Providing Guidance (Exact F1 of 0.4933), 20th of Mistake Identification (Exact F1 of 0.6617) and 22nd of Mistake Location (Exact F1 of 0.4935). The difference in performance over tracks highlights the difficulty of automatic pedagogical evaluation, especially for tasks whose solutions require a deep understanding of educational contexts. This work contributes to ongoing efforts to develop robust automated tools for assessing.

In this paper, we present the RETUYT-INCO participation at the BEA 2025 shared task. Our participation was characterized by the decision of using relatively small models, with fewer than 1B parameters. This self-imposed restriction tries to represent the conditions in which many research groups or institutions are in the Global South, where computational power is not easily accessible due to its prohibitive cost. Even under this restrictive self-imposed setting, our models managed to stay competitive with the rest of teams that participated in the shared task. According to the exact F1 scores published by the organizers, our models had the following distances with respect to the winners: 6.46 in Track 1; 10.24 in Track 2; 7.85 in Track 3; 9.56 in Track 4; and 13.13 in Track 5. Considering that the minimum difference with a winner team is 6.46 points — and the maximum difference is 13.13 — according to the exact F1 score, we find that models with a size smaller than 1B parameters are competitive for these tasks, all of which can be run on computers with a low-budget GPU or even without a GPU.

This paper presents automatic evaluation systems for assessing the pedagogical capabilities of LLM-based AI tutors. Drawing from a shared task, our systems specifically target four key dimensions of tutor responses: Mistake Identification, Mistake Location, Providing Guidance, and Actionability. These dimensions capture the educational quality of responses from multiple perspectives, including the ability to detect student mistakes, accurately identify error locations, provide effective instructional guidance, and offer actionable feedback. We propose GPT-4.1-based automatic evaluation systems, leveraging their strong capabilities in comprehending diverse linguistic expressions and complex conversational contexts to address the detailed evaluation criteria across these dimensions. Our systems were quantitatively evaluated based on the official criteria of each track. In the Mistake Location track, our evaluation systems achieved an Exact macro F1 score of 58.80% (ranked in the top 3), and in the Providing Guidance track, they achieved 56.06% (ranked in the top 5). While the systems showed mid-range performance in the remaining tracks, the overall results demonstrate that our proposed automatic evaluation systems can effectively assess the quality of tutor responses, highlighting their potential for evaluating AI tutor effectiveness.

Effective AI tutoring hinges on guiding learners with the right balance of support. In this work, we introduce CODE (COntextually-aware Distilled Evaluator), a framework that harnesses advanced large language models (i.e., GPT-4o and Claude-2.7) to generate synthetic, context-aware justifications for human-annotated tutor responses in the BEA 2025 Shared Task. By distilling these justifications into a smaller open-source model (i.e, Phi-3.5-mini-instruct) via initial supervised fine-tuning and then Group Relative Policy Optimization, we achieve substantial gains in label prediction over direct prompting of proprietary LLMs. Our experiments show that CODE reliably identifies strong positive and negative guidance, but like prior work, struggles to distinguish nuanced “middle-ground” cases where partial hints blur with vagueness. We argue that overcoming this limitation will require the development of explicit, feature-based evaluation metrics that systematically map latent pedagogical qualities to model outputs, enabling more transparent and robust assessment of AI-driven tutoring.

This paper presents our contribution to the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-Powered Tutors. The objective of this shared task was to assess the quality of conversational feedback provided by LLM-based math tutors to students regarding four facets: whether the tutors 1) identified mistakes, 2) identified the mistake’s location, 3) provided guidance, and whether they 4) provided actionable feedback. To leverage information across all four labels, we approached the problem with FLAN-T5 models, which we fit for this task using a multi-step pipeline involving regular fine-tuning as well as model merging using the DARE-TIES algorithm. We can demonstrate that our pipeline is beneficial to overall model performance compared to regular fine-tuning. With results on the test set ranging from 52.1 to 68.6 in F1 scores and 62.2% to 87.4% in accuracy, our best models placed 11th of 44 teams in Track 1, 8th of 31 teams in Track 2, 11th of 35 teams in Track 3, and 9th of 30 teams in Track 4. Notably, the classifiers’ recall was relatively poor for underrepresented classes, indicating even greater potential for the employed methodology.

While large language models show promise as AI tutors, evaluating their pedagogical capabilities remains challenging. In this paper, we, team LexiLogic presents our participation in the BEA 2025 shared task on evaluating AI tutors across five dimensions: Mistake Identification, Mistake Location, Providing Guidance, Actionability, and Tutor Identification. We approach all tracks as classification tasks using fine-tuned transformer models on a dataset of 300 educational dialogues between a student and a tutor in the mathematical domain. Our results show varying performance across tracks, with macro average F1 scores ranging from 0.47 to 0.82, achieving rankings between 4th and 31st place. Such models have the potential to be used in developing automated scoring metrics for assessing the pedagogical skills of AI math tutors.

As AI’s presence in educational environments grows, it becomes critical to evaluate how its feedback may impact students’ learning processes. Pedagogical theory, with decades of effort into understanding how human instructors give good-quality feedback to students, may provide a rich source of insight into feedback automation. In this paper, we propose a novel architecture based on pedagogical-theory feature extraction from the conversation history and tutor response to predict pedagogical guidance on MRBench. Such features are based on Brookhart’s canonical work in pedagogical theory, and extracted by prompting the language model LearnLM. The features are then used to train a random-forest classifier to predict the ‘providing guidance’ dimension of the MRBench dataset. Our approach ranked 8th in the dimension’s leaderboard with a test Macro F1-score of ~0.54. Our work provides some evidence in support of using pedagogical theory qualitative factors treated separately to provide clearer guidelines on how to improve low-scoring intelligent tutoring systems. Finally, we observed several inconsistencies between pedagogical theory and MRBench’s inherent relaxation of the tutoring problem implied in evaluating on a single-conversation basis, calling for the development of more elaborate measures which consider student profiles to serve as true heuristics of AI tutors’ usefulness.

We present MSA-MathEval, our submission to the BEA 2025 Shared Task on evaluating AI tutor responses across four instructional dimensions: Mistake Identification, Mistake Location, Providing Guidance, and Actionability. Our approach uses a unified training pipeline to fine-tune a single instruction-tuned language model across all tracks, without any task-specific architectural changes. To improve prediction reliability, we introduce a disagreement-aware ensemble inference strategy that enhances coverage of minority labels. Our system achieves strong performance across all tracks, ranking 1st in Providing Guidance, 3rd in Actionability, and 4th in both Mistake Identification and Mistake Location. These results demonstrate the effectiveness of scalable instruction tuning and disagreement-driven modeling for robust, multi-dimensional evaluation of LLMs as educational tutors.

In light of the growing adoption of large language models (LLMs) as educational tutors, it is crucial to effectively evaluate their pedagogical capabilities across multiple dimensions. Toward this goal, we address the Mistake Identification sub-task of the BEA 2025 Shared task, aiming to assess the accuracy of tutors in detecting and identifying student errors. We experiment with several LLMs, including GPT-4o-mini, Mistral-7B, and Llama-3.1-8B, evaluating them in both zero-shot and fine-tuned settings. To address class imbalance, we augment the training data with synthetic examples, targeting underrepresented labels, generated by Command R+. Our GPT-4o model finetuned on the full development set achieves a strict macro-averaged F1 score of 71.63%, ranking second in the shared task. Our work highlights the effectiveness of fine-tuning on task-specific data and suggests that targeted data augmentation can further support LLM performance on nuanced pedagogical evaluation tasks.

This paper presents DiReC (Disentangled Contrastive Representation), a novel two-stage framework designed to address the BEA 2025 Shared Task 5: Tutor Identity Classification. The task involves distinguishing between responses generated by nine different tutors, including both human educators and large language models (LLMs). DiReC leverages a disentangled representation learning approach, separating semantic content and stylistic features to improve tutor identification accuracy. In Stage 1, the model learns discriminative content representations using cross-entropy loss. In Stage 2, it applies supervised contrastive learning on style embeddings and introduces a disentanglement loss to enforce orthogonality between style and content spaces. Evaluated on the validation set, DiReC achieves strong performance, with a macro-F1 score of 0.9101 when combined with a CatBoost classifier and refined using the Hungarian algorithm. The system ranks third overall in the shared task with a macro-F1 score of 0.9172, demonstrating the effectiveness of disentangled representation learning for tutor identity classification.

This paper describes our approach for 5 classification tasks from Building Educational Applications (BEA) 2025 Shared Task.Our methods range from classical machine learning models to large-scale transformers with fine-tuning and prompting strategies. Despite the diversity of approaches, performance differences were often minor, suggesting a strong surface-level signal and the limiting effect of annotation noise—particularly around the “To some extent” label. Under lenient evaluation, simple models perform competitively, showing their effectiveness in low-resource settings. Our submissions ranked in the top 10 in four of five tracks.

This paper describes the system created for the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered Tutors. The task aims to assess how well AI tutors identify and locate errors made by students, provide guidance and ensure actionability, among other features of their responses in educational dialogues. Transformer-based models, especially DeBERTa and RoBERTa, are improved by multitask learning, threshold tweaking, ordinal regression, and oversampling. The efficiency of pedagogically driven training methods and bespoke transformer models for evaluating AI tutor quality is demonstrated by the high performance of their best systems across all evaluation tracks.

This paper presents our system for Track 1: Mistake Identification in the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered Tutors. The task involves evaluating whether a tutor’s response correctly identifies a mistake in a student’s mathematical reasoning. We explore four approaches: (1) an ensemble of machine learning models over pooled token embeddings from multiple pretrained langauge models (LMs); (2) a frozen sentence-transformer using [CLS] embeddings with an MLP classifier; (3) a history-aware model with multi-head attention between token-level history and response embeddings; and (4) a retrieval-augmented few-shot prompting system with a large language model (LLM) i.e. GPT 4o. Our final system retrieves semantically similar examples, constructs structured prompts, and uses schema-guided output parsing to produce interpretable predictions. It outperforms all baselines, demonstrating the effectiveness of combining example-driven prompting with LLM reasoning for pedagogical feedback assessment.

We present our submission for Tracks 3 (Providing Guidance), 4 (Actionability), and 5 (Tutor Identification) of the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-Powered Tutors. Our approach sought to investigate the performance of directly using sentence embeddings of tutor responses as input to downstream classifiers (that is, without employing any fine-tuning). To this end, we benchmarked two general-purpose sentence embedding models: gte-modernbert-base (GTE) and all-MiniLM-L12-v2, in combination with two downstream classifiers: XGBoost and multilayer perceptron. Feeding GTE embeddings to a multilayer perceptron achieved macro-F1 scores of 0.4776, 0.5294, and 0.6420 on the official test sets for Tracks 3, 4, and 5, respectively. While overall performance was modest, these results offer insights into the challenges of pedagogical response evaluation and establish a baseline for future improvements.

We present Team BD’s submission to the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered Tutors, under Track 1 (Mistake Identification) and Track 2 (Mistake Location). Both tracks involve three-class classification of tutor responses in educational dialogues – determining if a tutor correctly recognizes a student’s mistake (Track 1) and whether the tutor pinpoints the mistake’s location (Track 2). Our system is built on MPNet, a Transformer-based language modelthat combines BERT and XLNet’s pre-training advantages. We fine-tuned MPNet on the task data using a class-weighted cross-entropy loss to handle class imbalance, and leveraged grouped cross-validation (10 folds) to maximize the use of limited data while avoiding dialogue overlap between training and validation. We then performed a hard-voting ensemble of the best models from each fold, which improves robustness and generalization by combining multiple classifiers. Ourapproach achieved strong results on both tracks, with exact-match macro-F1 scores of approximately 0.7110 for Mistake Identification and 0.5543 for Mistake Location on the official test set. We include comprehensive analysis of our system’s performance, including confusion matrices and t-SNE visualizations to interpret classifier behavior, as well as a taxonomy of common errors with examples. We hope our ensemble-based approach and findings provide useful insights for designing reliable tutor response evaluation systems in educational dialogue settings.

This paper presents Thapar Titan/s’ submission to the BEA 2025 Shared Task on Pedagogical Ability Assessment of AI-powered Tutors. The shared task consists of five subtasks; our team ranked 18th in Mistake Identification, 15th in Mistake Location, and 18th in Actionability. However, in this paper, we focus exclusively on presenting results for Task 1: Mistake Identification, which evaluates a system’s ability to detect student mistakes.Our approach employs contextual data augmentation using a RoBERTa based masked language model to mitigate class imbalance, supplemented by oversampling and weighted loss training. Subsequently, we fine-tune three separate classifiers: RoBERTa, BERT, and DeBERTa for three-way classification aligned with task-specific annotation schemas. This modular and scalable pipeline enables a comprehensive evaluation of tutor feedback quality in educational dialogues.