Austria

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&#39;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.

ACL 2025

Alignment Drift in CEFR-prompted LLMs for Interactive Spanish Tutoring

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

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.

workshop paper

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.

Can LLMs Effectively Simulate Human Learners? Teachers' Insights from Tutoring LLM Students

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.

Adapting LLMs for Minimal-edit Grammatical Error Correction

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.

Is Lunch Free Yet? Overcoming the Cold-Start Problem in Supervised Content Scoring using Zero-Shot LLM-Generated Training Data

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 versions of the British Council's Knowledge-based Vocabulary Lists (KVL), containing 6,768 English words paired with difficulty scores and question prompts written in Spanish, German, and Mandarin Chinese. Using this 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 insights 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.

Transformer Architectures for Vocabulary Test Item Difficulty Prediction

Heterogeneity in student populations poses a challenge in formal education, with adaptive textbooks offering a potential solution by taioring content based on individual learner models. However, creating domain models for textbooks typically demands significant manual effort. Recent work by Chau et al. (2021) demonstrated automated concept extraction from digital textbooks, but relied on costly domain-specific manual annotations. This paper introduces a novel, scalable method that minimizes manual effort by combining contextualized word embeddings with weakly supervised machine learning. Our approach clusters word embeddings from textbooks and identifies domain-specific concepts using a machine learner trained on concept seeds automatically extracted from Wikipedia. We evaluate this method using 28 economics textbooks, comparing its performance against a tf-idf baseline, a supervised machine learning baseline, the RAKE keyword extraction method, and human domain experts. Results demonstrate that our weakly supervised method effectively balances accuracy with reduced annotation effort, offering a practical solution for automated concept extraction in adaptive learning environments.

Automatic concept extraction for learning domain modeling: A weakly supervised approach using contextualized word embeddings

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.

Towards a Real-time Swedish Speech Analyzer for Language Learning Games: A Hybrid AI Approach to Language Assessment

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.

LLM-based post-editing as reference-free GEC evaluation

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.

Automated Scoring of a German Written Elicited Imitation Test

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.

Do LLMs Give Psychometrically Plausible Responses in Educational Assessments?

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

Downloads

Next from ACL 2025

Can LLMs Effectively Simulate Human Learners? Teachers' Insights from Tutoring LLM Students

Similar lecture

Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

Can LLMs Effectively Simulate Human Learners? Teachers' Insights from Tutoring LLM Students

Similar lecture

Exploring LLMs for Predicting Tutor Strategy and Student Outcomes in Dialogues

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES

Downloads