arXiv subjects

Sahan Bulathwela

Publications and source records attributed to Sahan Bulathwela.

At least 19 recordsLinked to original sources

arXiv2026

The Correct Answer Trap: Pedagogically-Grounded Detection and Feedback for Hidden Misconceptions

Automated feedback systems that rely on answer correctness will reinforce, rather than address, misconceptions when students reach the correct answer through flawed reasoning. We investigate automatic detection of these hidden misconceptions using 20,964 real student responses from the Eedi mathematics platform. Fine-tuned classifiers detect only 57% of these hidden misconceptions, and standard ML interventions do not improve on this. An open-weight reasoning model detects 84%, but at realistic prevalence, false alarms outnumber genuine detections roughly 8 to 1. We present a graduated assessment rubric that separates answer correctness from method validity, and propose a detect-verify-escalate pipeline that routes uncertain cases to diagnostic follow-up questions rather than directly to teachers. Two deployment modes adapt the pipeline: a teacher dashboard where the system filters a review queue, and an autonomous tutor where flags trigger low-cost formative follow-up.

cs.CY↗

arXiv2026

Sell Me This Stock: Unsafe Recommendation Drift in LLM Agents

People increasingly use LLM agents for multi-turn financial recommendations, where the agent pulls market data through tools and tracks user preferences across turns. When tool outputs are manipulated, the recommendations stop matching the user's stated risk profile, but because standard metrics like NDCG only score general relevance, risky and safe stocks score alike, so the metric says nothing went wrong. We call this gap evaluation blindness. We replay 23-turn financial advisory conversations across eight language models, running each dialogue twice with clean and manipulated tool data. Quality scores stay nearly identical to clean sessions while the agents produce risk-mismatched recommendations in 65-99% of turns, unanimous across all eight models. The mechanism is visible turn-by-turn: 80% of risk-score citations across 1,840 turns reproduce the manipulated value verbatim, not a single turn pushes back, and safe-language framing of high-risk stocks ranges from 14% (Qwen2.5-7B) to 69% (Claude Sonnet 4.6). The property that makes frontier models good agents, faithfully grounding their reasoning in tool outputs, also makes them follow manipulated ones. The damage is not memory-driven: contaminating only the current turn still produces 95% of the violations. The model internally distinguishes the manipulation (sparse autoencoder features separate adversarial from random perturbations), but this does not translate into safer output. Activation-level interventions recover under 6% of the safety gap, prompt-level self-verification fails because the self-check reads the same manipulated data, and a parametric cross-check that flags contamination at 99-100% per turn on a frontier model still leaves aggregate suitability unchanged: the agent identifies the tampering and recommends it anyway.

cs.CL↗

arXiv2026

Tool Calling is Linearly Readable and Steerable in Language Models

When a tool-calling agent picks the wrong tool, the failure is invisible until execution: the email gets sent, the meeting gets missed. As agents take on consequential actions, one bad tool call can do real damage. We currently have no way to look inside the model and catch the mistake before it happens; this paper shows that we can. Inside the model, the choice of tool is carried by a single direction in activation space, one direction per pair of tools. Adding that direction during generation switches which tool the model picks. Across 12 instruction-tuned and 6 base models spanning Gemma 3, Qwen 3, Qwen 2.5, and Llama 3.1 (270M to 27B), this works at 83-100% accuracy on 4B+ instruction-tuned models on a 15-tool synthetic benchmark and at 77-94% on the real-API benchmark $τ$-bench airline. The JSON arguments that follow automatically adapt to the new tool's schema, so flipping the name is enough. The same per-tool directions also flag likely errors before they happen: queries where the model is unsure between two tools fail 21x more often than queries where it is not (Gemma 3 27B). This is not just topic injection: random vectors at the same magnitude give a 0% switch rate, and a probe within a single domain (14 airline tools that share one topic) still reads which tool the model will call at top-1 61-89% across five 4B-14B models. Even base models already carry the right tool internally before they can emit it: reading the chosen tool off the model's internal state (cosine readout) recovers 61-82% accuracy on BFCL while base generation lands at 2-10%, suggesting pretraining forms the representation and instruction tuning later wires it to the output. Our results cover single-turn, fixed-menu settings; on multi-turn agent loops the same intervention is less stable (matched-baseline gain or loss of up to 30 percentage points with no consistent direction).

cs.CL↗

arXiv2026

Mix and Match: Context Pairing for Scalable Topic-Controlled Educational Summarisation

Topic-controlled summarisation enables users to generate summaries focused on specific aspects of source documents. This paper investigates a data augmentation strategy for training small language models (sLMs) to perform topic-controlled summarisation. We propose a pairwise data augmentation method that combines contexts from different documents to create contrastive training examples, enabling models to learn the relationship between topics and summaries more effectively. Using the SciTLDR dataset enriched with Wikipedia-derived topics, we systematically evaluate how augmentation scale affects model performance. Results show consistent improvements in win rate and semantic alignment as the augmentation scale increases, while the amount of real training data remains fixed. Consequently, a T5-base model trained with our augmentation approach achieves competitive performance relative to larger models, despite using significantly fewer parameters and substantially fewer real training examples.

cs.CL↗

arXiv2026

Catching The Correct Answer Trap: Characterising AI Tutor Blind Spots When Analysing Student Reasoning

Intelligent tutoring systems increasingly provide automated feedback on student work, but robust feedback requires assessing reasoning, not only final answers. We study a failure mode we call the correct answer trap (CAT): models under-detect misconceptions when students reach a correct answer via flawed reasoning. Analysing real student responses from the Eedi mathematics platform, we show that 71% of these failures concentrate in just two question types, both sharing a common structure where flawed reasoning happens to produce the correct numerical answer. Comparing a fine-tuned T5 with a frontier large language model, we find that improved capabilities reduce but do not eliminate the problem (84% vs 57% detection accuracy). Even the best-performing model generates roughly four false alarms for every genuine detection, making stand-alone screening impractical at realistic class sizes. Our findings demonstrate that high overall accuracy can mask critical failures in reasoning assessment, and that careful analysis of student reasoning still benefits from human judgment.

cs.CY↗

arXiv2026

Gaze to Insight: A Scalable AI Approach for Detecting Gaze Behaviours in Face-to-Face Collaborative Learning

Previous studies have illustrated the potential of analysing gaze behaviours in collaborative learning to provide educationally meaningful information for students to reflect on their learning. Over the past decades, machine learning approaches have been developed to automatically detect gaze behaviours from video data. Yet, since these approaches often require large amounts of labelled data for training, human annotation remains necessary. Additionally, researchers have questioned the cross-configuration robustness of machine learning models developed, as training datasets often fail to encompass the full range of situations encountered in educational contexts. To address these challenges, this study proposes a scalable artificial intelligence approach that leverages pretrained and foundation models to automatically detect gaze behaviours in face-to-face collaborative learning contexts without requiring human-annotated data. The approach utilises pretrained YOLO11 for person tracking, YOLOE-26 with text-prompt capability for education-related object detection, and the Gaze-LLE model for gaze target prediction. The results indicate that the proposed approach achieves an F1-score of 0.829 in detecting students' gaze behaviours from video data, with strong performance for laptop-directed gaze and peer-directed gaze, yet weaker performance for other gaze targets. Furthermore, when compared to other supervised machine learning approaches, the proposed method demonstrates superior and more stable performance in complex contexts, highlighting its better cross-configuration robustness. The implications of this approach for supporting students' collaborative learning in real-world environments are also discussed.

cs.CV↗

arXiv2026

Next Token Knowledge Tracing: Exploiting Pretrained LLM Representations to Decode Student Behaviour

Modelling student knowledge is a key challenge when leveraging AI in education, with major implications for personalised learning. The Knowledge Tracing (KT) task aims to predict how students will respond to educational questions in learning environments, based on their prior interactions. Existing KT models typically use response correctness along with metadata like skill tags and timestamps, often overlooking the question text, which is an important source of pedagogical insight. This omission poses a lost opportunity while limiting predictive performance. We propose Next Token Knowledge Tracing (NTKT), a novel approach that reframes KT as a next-token prediction task using pretrained Large Language Models (LLMs). NTKT represents both student histories and question content as sequences of text, allowing LLMs to learn patterns in both behaviour and language. Our series of experiments significantly improves performance over state-of-the-art neural KT models and generalises much better to cold-start questions and users. These findings highlight the importance of question content in KT and demonstrate the benefits of leveraging pretrained representations of LLMs to model student learning more effectively.

cs.CL↗

arXiv2025

Examining Student Interactions with a Pedagogical AI-Assistant for Essay Writing and their Impact on Students Writing Quality

The dynamic nature of interactions between students and GenAI, as well as their relationship to writing quality, remains underexplored. While most research has examined how general-purpose GenAI can support writing, fewer studies have investigated how students interact with pedagogically designed systems across different phases of the writing process. To address this gap, we evaluated a GenAI-driven essay-writing assistant (EWA) designed to support higher education students in argumentative writing. Drawing on 1,282 interaction logs from 32 undergraduates during a two-hour writing session, Sequential Pattern Mining and K-Means clustering were used to identify behavioral patterns. Two clusters emerged: Cluster 1 emphasized outline planning and essay structure, while Cluster 2 focused on content development. A Mann-Whitney U test revealed a moderate effect size (r = 0.36) in the essay Organization dimension, with Cluster 1 showing higher scores. Qualitative analysis indicated that students with better performance actively wrote and shared essay sections with EWA for feedback, rather than interacted passively by asking questions. These findings suggest implications for teaching and system design. Teachers can encourage active engagement, while future EWAs may integrate automatic labeling and monitoring to prompt students to move from questioning to writing, enabling fuller benefits from GenAI-supported learning.

cs.CY↗

arXiv2025

Scaffolding Reshapes Dialogic Engagement in Collaborative Problem Solving: Comparative Analysis of Two Approaches

Supporting learners during Collaborative Problem Solving (CPS) is a necessity. Existing studies have compared scaffolds with maximal and minimal instructional support by studying their effects on learning and behaviour. However, our understanding of how such scaffolds could differently shape the distribution of individual dialogic engagement and behaviours across different CPS phases remains limited. This study applied Heterogeneous Interaction Network Analysis (HINA) and Sequential Pattern Mining (SPM) to uncover the structural effects of scaffolding on different phases of the CPS process among K-12 students in authentic educational settings. Students with a maximal scaffold demonstrated higher dialogic engagement across more phases than those with a minimal scaffold. However, they were extensively demonstrating scripting behaviours across the phases, evidencing the presence of overscripting. Although students with the minimal scaffold demonstrated more problem solving behaviours and fewer scripting behaviours across the phases, they repeated particular behaviours in multiple phases and progressed more to socialising behaviours. In both scaffold conditions, problem solving behaviours rarely progressed to other problem solving behaviours. The paper discusses the implications of these findings for scaffold design and teaching practice of CPS, and highlights the distinct yet complementary value of HINA and SPM approaches to investigate students' learning processes during CPS.

cs.SI↗

arXiv2025

Towards Synergistic Teacher-AI Interactions with Generative Artificial Intelligence

Generative artificial intelligence (GenAI) is increasingly used in education, posing significant challenges for teachers adapting to these changes. GenAI offers unprecedented opportunities for accessibility, scalability and productivity in educational tasks. However, the automation of teaching tasks through GenAI raises concerns about reduced teacher agency, potential cognitive atrophy, and the broader deprofessionalisation of teaching. Drawing findings from prior literature on AI in Education, and refining through a recent systematic literature review, this chapter presents a conceptualisation of five levels of teacher-AI teaming: transactional, situational, operational, praxical and synergistic teaming. The framework aims to capture the nuanced dynamics of teacher-AI interactions, particularly with GenAI, that may lead to the replacement, complementarity, or augmentation of teachers' competences and professional practice. GenAI technological affordances required in supporting teaming, along with empirical studies, are discussed. Drawing on empirical observations, we outline a future vision that moves beyond individual teacher agency toward collaborative decision-making between teachers and AI, in which both agents engage in negotiation, constructive challenge, and co-reasoning that enhance each other's capabilities and enable outcomes neither could realise independently. Further discussion of socio-technical factors beyond teacher-AI teaming is also included to streamline the synergy of teachers and AI in education ethically and practically.

cs.CY↗

arXiv2025

Exploring Human-AI Complementarity in CPS Diagnosis Using Unimodal and Multimodal BERT Models

Detecting collaborative problem solving (CPS) indicators from dialogue using machine learning techniques is a significant challenge for the field of AI in Education. Recent studies have explored the use of Bidirectional Encoder Representations from Transformers (BERT) models on transcription data to reliably detect meaningful CPS indicators. A notable advancement involved the multimodal BERT variant, AudiBERT, which integrates speech and acoustic-prosodic audio features to enhance CPS diagnosis. Although initial results demonstrated multimodal improvements, the statistical significance of these enhancements remained unclear, and there was insufficient guidance on leveraging human-AI complementarity for CPS diagnosis tasks. This workshop paper extends the previous research by highlighting that the AudiBERT model not only improved the classification of classes that were sparse in the dataset, but it also had statistically significant class-wise improvements over the BERT model for classifications in the social-cognitive dimension. However, similar significant class-wise improvements over the BERT model were not observed for classifications in the affective dimension. A correlation analysis highlighted that larger training data was significantly associated with higher recall performance for both the AudiBERT and BERT models. Additionally, the precision of the BERT model was significantly associated with high inter-rater agreement among human coders. When employing the BERT model to diagnose indicators within these subskills that were well-detected by the AudiBERT model, the performance across all indicators was inconsistent. We conclude the paper by outlining a structured approach towards achieving human-AI complementarity for CPS diagnosis, highlighting the crucial inclusion of model explainability to support human agency and engagement in the reflective coding process.

cs.CL↗

arXiv2025

Explainable Collaborative Problem Solving Diagnosis with BERT using SHAP and its Implications for Teacher Adoption

The use of Bidirectional Encoder Representations from Transformers (BERT) model and its variants for classifying collaborative problem solving (CPS) has been extensively explored within the AI in Education community. However, limited attention has been given to understanding how individual tokenised words in the dataset contribute to the model's classification decisions. Enhancing the explainability of BERT-based CPS diagnostics is essential to better inform end users such as teachers, thereby fostering greater trust and facilitating wider adoption in education. This study undertook a preliminary step towards model transparency and explainability by using SHapley Additive exPlanations (SHAP) to examine how different tokenised words in transcription data contributed to a BERT model's classification of CPS processes. The findings suggested that well-performing classifications did not necessarily equate to a reasonable explanation for the classification decisions. Particular tokenised words were used frequently to affect classifications. The analysis also identified a spurious word, which contributed positively to the classification but was not semantically meaningful to the class. While such model transparency is unlikely to be useful to an end user to improve their practice, it can help them not to overrely on LLM diagnostics and ignore their human expertise. We conclude the workshop paper by noting that the extent to which the model appropriately uses the tokens for its classification is associated with the number of classes involved. It calls for an investigation into the exploration of ensemble model architectures and the involvement of human-AI complementarity for CPS diagnosis, since considerable human reasoning is still required for fine-grained discrimination of CPS subskills.

cs.CL↗

arXiv2025

TrueReason: An Exemplar Personalised Learning System Integrating Reasoning with Foundational Models

Personalised education is one of the domains that can greatly benefit from the most recent advances in Artificial Intelligence (AI) and Large Language Models (LLM). However, it is also one of the most challenging applications due to the cognitive complexity of teaching effectively while personalising the learning experience to suit independent learners. We hypothesise that one promising approach to excelling in such demanding use cases is using a \emph{society of minds}. In this chapter, we present TrueReason, an exemplar personalised learning system that integrates a multitude of specialised AI models that can mimic micro skills that are composed together by a LLM to operationalise planning and reasoning. The architecture of the initial prototype is presented while describing two micro skills that have been incorporated in the prototype. The proposed system demonstrates the first step in building sophisticated AI systems that can take up very complex cognitive tasks that are demanded by domains such as education.

cs.CY↗

arXiv2025

A Novel Approach to Scalable and Automatic Topic-Controlled Question Generation in Education

The development of Automatic Question Generation (QG) models has the potential to significantly improve educational practices by reducing the teacher workload associated with creating educational content. This paper introduces a novel approach to educational question generation that controls the topical focus of questions. The proposed Topic-Controlled Question Generation (T-CQG) method enhances the relevance and effectiveness of the generated content for educational purposes. Our approach uses fine-tuning on a pre-trained T5-small model, employing specially created datasets tailored to educational needs. The research further explores the impacts of pre-training strategies, quantisation, and data augmentation on the model's performance. We specifically address the challenge of generating semantically aligned questions with paragraph-level contexts, thereby improving the topic specificity of the generated questions. In addition, we introduce and explore novel evaluation methods to assess the topical relatedness of the generated questions. Our results, validated through rigorous offline and human-backed evaluations, demonstrate that the proposed models effectively generate high-quality, topic-focused questions. These models have the potential to reduce teacher workload and support personalised tutoring systems by serving as bespoke question generators. With its relatively small number of parameters, the proposals not only advance the capabilities of question generation models for handling specific educational topics but also offer a scalable solution that reduces infrastructure costs. This scalability makes them feasible for widespread use in education without reliance on proprietary large language models like ChatGPT.

cs.CY↗

arXiv2024

Stereotype Detection in LLMs: A Multiclass, Explainable, and Benchmark-Driven Approach

Stereotype detection is a challenging and subjective task, as certain statements, such as "Black people like to play basketball," may not appear overtly toxic but still reinforce racial stereotypes. With the increasing prevalence of large language models (LLMs) in human-facing artificial intelligence (AI) applications, detecting these types of biases is essential. However, LLMs risk perpetuating and amplifying stereotypical outputs derived from their training data. A reliable stereotype detector is crucial for benchmarking bias, monitoring model input and output, filtering training data, and ensuring fairer model behavior in downstream applications. This paper introduces the Multi-Grain Stereotype (MGS) dataset, consisting of 51,867 instances across gender, race, profession, religion, and other stereotypes, curated from multiple existing datasets. We evaluate various machine learning approaches to establish baselines and fine-tune language models of different architectures and sizes, presenting a suite of stereotype multiclass classifiers trained on the MGS dataset. Given the subjectivity of stereotypes, explainability is essential to align model learning with human understanding of stereotypes. We employ explainable AI (XAI) tools, including SHAP, LIME, and BertViz, to assess whether the model's learned patterns align with human intuitions about stereotypes.Additionally, we develop stereotype elicitation prompts and benchmark the presence of stereotypes in text generation tasks using popular LLMs, employing the best-performing stereotype classifiers.

cs.CL↗

arXiv2023

A Toolbox for Modelling Engagement with Educational Videos

With the advancement and utility of Artificial Intelligence (AI), personalising education to a global population could be a cornerstone of new educational systems in the future. This work presents the PEEKC dataset and the TrueLearn Python library, which contains a dataset and a series of online learner state models that are essential to facilitate research on learner engagement modelling.TrueLearn family of models was designed following the "open learner" concept, using humanly-intuitive user representations. This family of scalable, online models also help end-users visualise the learner models, which may in the future facilitate user interaction with their models/recommenders. The extensive documentation and coding examples make the library highly accessible to both machine learning developers and educational data mining and learning analytics practitioners. The experiments show the utility of both the dataset and the library with predictive performance significantly exceeding comparative baseline models. The dataset contains a large amount of AI-related educational videos, which are of interest for building and validating AI-specific educational recommenders.

cs.CY↗

arXiv2023

Towards Auditing Large Language Models: Improving Text-based Stereotype Detection

Large Language Models (LLM) have made significant advances in the recent past becoming more mainstream in Artificial Intelligence (AI) enabled human-facing applications. However, LLMs often generate stereotypical output inherited from historical data, amplifying societal biases and raising ethical concerns. This work introduces i) the Multi-Grain Stereotype Dataset, which includes 52,751 instances of gender, race, profession and religion stereotypic text and ii) a novel stereotype classifier for English text. We design several experiments to rigorously test the proposed model trained on the novel dataset. Our experiments show that training the model in a multi-class setting can outperform the one-vs-all binary counterpart. Consistent feature importance signals from different eXplainable AI tools demonstrate that the new model exploits relevant text features. We utilise the newly created model to assess the stereotypic behaviour of the popular GPT family of models and observe the reduction of bias over time. In summary, our work establishes a robust and practical framework for auditing and evaluating the stereotypic bias in LLM.

cs.CL↗

arXiv2023

TrueLearn: A Python Library for Personalised Informational Recommendations with (Implicit) Feedback

This work describes the TrueLearn Python library, which contains a family of online learning Bayesian models for building educational (or more generally, informational) recommendation systems. This family of models was designed following the "open learner" concept, using humanly-intuitive user representations. For the sake of interpretability and putting the user in control, the TrueLearn library also contains different representations to help end-users visualise the learner models, which may in the future facilitate user interaction with their own models. Together with the library, we include a previously publicly released implicit feedback educational dataset with evaluation metrics to measure the performance of the models. The extensive documentation and coding examples make the library highly accessible to both machine learning developers and educational data mining and learning analytic practitioners. The library and the support documentation with examples are available at https://truelearn.readthedocs.io/en/latest.

cs.IR↗