Vinitra Swamy

Zeming Chen, Alejandro Hernández Cano, Angelika Romanou et al. · allen-ai

Large language models (LLMs) can potentially democratize access to medical knowledge. While many efforts have been made to harness and improve LLMs' medical knowledge and reasoning capacities, the resulting models are either closed-source (e.g., PaLM, GPT-4) or limited in scale (<= 13B parameters), which restricts their abilities. In this work, we improve access to large-scale medical LLMs by releasing MEDITRON: a suite of open-source LLMs with 7B and 70B parameters adapted to the medical domain. MEDITRON builds on Llama-2 (through our adaptation of Nvidia's Megatron-LM distributed trainer), and extends pretraining on a comprehensively curated medical corpus, including selected PubMed articles, abstracts, and internationally-recognized medical guidelines. Evaluations using four major medical benchmarks show significant performance gains over several state-of-the-art baselines before and after task-specific finetuning. Overall, MEDITRON achieves a 6% absolute performance gain over the best public baseline in its parameter class and 3% over the strongest baseline we finetuned from Llama-2. Compared to closed-source LLMs, MEDITRON-70B outperforms GPT-3.5 and Med-PaLM and is within 5% of GPT-4 and 10% of Med-PaLM-2. We release our code for curating the medical pretraining corpus and the MEDITRON model weights to drive open-source development of more capable medical LLMs.

Vinitra Swamy, Sijia Du, Mirko Marras et al.

Deep learning models for learning analytics have become increasingly popular over the last few years; however, these approaches are still not widely adopted in real-world settings, likely due to a lack of trust and transparency. In this paper, we tackle this issue by implementing explainable AI methods for black-box neural networks. This work focuses on the context of online and blended learning and the use case of student success prediction models. We use a pairwise study design, enabling us to investigate controlled differences between pairs of courses. Our analyses cover five course pairs that differ in one educationally relevant aspect and two popular instance-based explainable AI methods (LIME and SHAP). We quantitatively compare the distances between the explanations across courses and methods. We then validate the explanations of LIME and SHAP with 26 semi-structured interviews of university-level educators regarding which features they believe contribute most to student success, which explanations they trust most, and how they could transform these insights into actionable course design decisions. Our results show that quantitatively, explainers significantly disagree with each other about what is important, and qualitatively, experts themselves do not agree on which explanations are most trustworthy. All code, extended results, and the interview protocol are provided at https://github.com/epfl-ml4ed/trusting-explainers.

Vinitra Swamy, Bahar Radmehr, Natasa Krco et al.

Neural networks are ubiquitous in applied machine learning for education. Their pervasive success in predictive performance comes alongside a severe weakness, the lack of explainability of their decisions, especially relevant in human-centric fields. We implement five state-of-the-art methodologies for explaining black-box machine learning models (LIME, PermutationSHAP, KernelSHAP, DiCE, CEM) and examine the strengths of each approach on the downstream task of student performance prediction for five massive open online courses. Our experiments demonstrate that the families of explainers do not agree with each other on feature importance for the same Bidirectional LSTM models with the same representative set of students. We use Principal Component Analysis, Jensen-Shannon distance, and Spearman's rank-order correlation to quantitatively cross-examine explanations across methods and courses. Furthermore, we validate explainer performance across curriculum-based prerequisite relationships. Our results come to the concerning conclusion that the choice of explainer is an important decision and is in fact paramount to the interpretation of the predictive results, even more so than the course the model is trained on. Source code and models are released at http://github.com/epfl-ml4ed/evaluating-explainers.

Mohammad Asadi, Vinitra Swamy, Jibril Frej et al.

Time series is the most prevalent form of input data for educational prediction tasks. The vast majority of research using time series data focuses on hand-crafted features, designed by experts for predictive performance and interpretability. However, extracting these features is labor-intensive for humans and computers. In this paper, we propose an approach that utilizes irregular multivariate time series modeling with graph neural networks to achieve comparable or better accuracy with raw time series clickstreams in comparison to hand-crafted features. Furthermore, we extend concept activation vectors for interpretability in raw time series models. We analyze these advances in the education domain, addressing the task of early student performance prediction for downstream targeted interventions and instructional support. Our experimental analysis on 23 MOOCs with millions of combined interactions over six behavioral dimensions show that models designed with our approach can (i) beat state-of-the-art educational time series baselines with no feature extraction and (ii) provide interpretable insights for personalized interventions. Source code: https://github.com/epfl-ml4ed/ripple/.

Vinitra Swamy, Malika Satayeva, Jibril Frej et al.

Predicting multiple real-world tasks in a single model often requires a particularly diverse feature space. Multimodal (MM) models aim to extract the synergistic predictive potential of multiple data types to create a shared feature space with aligned semantic meaning across inputs of drastically varying sizes (i.e. images, text, sound). Most current MM architectures fuse these representations in parallel, which not only limits their interpretability but also creates a dependency on modality availability. We present MultiModN, a multimodal, modular network that fuses latent representations in a sequence of any number, combination, or type of modality while providing granular real-time predictive feedback on any number or combination of predictive tasks. MultiModN's composable pipeline is interpretable-by-design, as well as innately multi-task and robust to the fundamental issue of biased missingness. We perform four experiments on several benchmark MM datasets across 10 real-world tasks (predicting medical diagnoses, academic performance, and weather), and show that MultiModN's sequential MM fusion does not compromise performance compared with a baseline of parallel fusion. By simulating the challenging bias of missing not-at-random (MNAR), this work shows that, contrary to MultiModN, parallel fusion baselines erroneously learn MNAR and suffer catastrophic failure when faced with different patterns of MNAR at inference. To the best of our knowledge, this is the first inherently MNAR-resistant approach to MM modeling. In conclusion, MultiModN provides granular insights, robustness, and flexibility without compromising performance.

Thiemo Wambsganss, Xiaotian Su, Vinitra Swamy et al.

Large Language Models (LLMs) are increasingly utilized in educational tasks such as providing writing suggestions to students. Despite their potential, LLMs are known to harbor inherent biases which may negatively impact learners. Previous studies have investigated bias in models and data representations separately, neglecting the potential impact of LLM bias on human writing. In this paper, we investigate how bias transfers through an AI writing support pipeline. We conduct a large-scale user study with 231 students writing business case peer reviews in German. Students are divided into five groups with different levels of writing support: one classroom group with feature-based suggestions and four groups recruited from Prolific -- a control group with no assistance, two groups with suggestions from fine-tuned GPT-2 and GPT-3 models, and one group with suggestions from pre-trained GPT-3.5. Using GenBit gender bias analysis, Word Embedding Association Tests (WEAT), and Sentence Embedding Association Test (SEAT) we evaluate the gender bias at various stages of the pipeline: in model embeddings, in suggestions generated by the models, and in reviews written by students. Our results demonstrate that there is no significant difference in gender bias between the resulting peer reviews of groups with and without LLM suggestions. Our research is therefore optimistic about the use of AI writing support in the classroom, showcasing a context where bias in LLMs does not transfer to students' responses.

Fares Fawzi, Vinitra Swamy, Dominik Glandorf et al.

Language models can be used to provide interactive, personalized student feedback in educational settings. However, real-world deployment faces three key challenges: privacy concerns, limited computational resources, and the need for pedagogically valid responses. These constraints require small, open-source models that can run locally and reliably ground their outputs in correct information. We introduce SCRIBE, a framework for multi-hop, tool-augmented reasoning designed to generate valid responses to student questions about feedback reports. SCRIBE combines domain-specific tools with a self-reflective inference pipeline that supports iterative reasoning, tool use, and error recovery. We distil these capabilities into 3B and 8B models via two-stage LoRA fine-tuning on synthetic GPT-4o-generated data. Evaluation with a human-aligned GPT-Judge and a user study with 108 students shows that 8B-SCRIBE models achieve comparable or superior quality to much larger models in key dimensions such as relevance and actionability, while being perceived on par with GPT-4o and Llama-3.3 70B by students. These findings demonstrate the viability of SCRIBE for low-resource, privacy-sensitive educational applications.

Thiemo Wambsganss, Vinitra Swamy, Roman Rietsche et al.

Natural Language Processing (NLP) has become increasingly utilized to provide adaptivity in educational applications. However, recent research has highlighted a variety of biases in pre-trained language models. While existing studies investigate bias in different domains, they are limited in addressing fine-grained analysis on educational and multilingual corpora. In this work, we analyze bias across text and through multiple architectures on a corpus of 9,165 German peer-reviews collected from university students over five years. Notably, our corpus includes labels such as helpfulness, quality, and critical aspect ratings from the peer-review recipient as well as demographic attributes. We conduct a Word Embedding Association Test (WEAT) analysis on (1) our collected corpus in connection with the clustered labels, (2) the most common pre-trained German language models (T5, BERT, and GPT-2) and GloVe embeddings, and (3) the language models after fine-tuning on our collected data-set. In contrast to our initial expectations, we found that our collected corpus does not reveal many biases in the co-occurrence analysis or in the GloVe embeddings. However, the pre-trained German language models find substantial conceptual, racial, and gender bias and have significant changes in bias across conceptual and racial axes during fine-tuning on the peer-review data. With our research, we aim to contribute to the fourth UN sustainability goal (quality education) with a novel dataset, an understanding of biases in natural language education data, and the potential harms of not counteracting biases in language models for educational tasks.

Vinitra Swamy, Mirko Marras, Tanja Käser

Despite the increasing popularity of massive open online courses (MOOCs), many suffer from high dropout and low success rates. Early prediction of student success for targeted intervention is therefore essential to ensure no student is left behind in a course. There exists a large body of research in success prediction for MOOCs, focusing mainly on training models from scratch for individual courses. This setting is impractical in early success prediction as the performance of a student is only known at the end of the course. In this paper, we aim to create early success prediction models that can be transferred between MOOCs from different domains and topics. To do so, we present three novel strategies for transfer: 1) pre-training a model on a large set of diverse courses, 2) leveraging the pre-trained model by including meta information about courses, and 3) fine-tuning the model on previous course iterations. Our experiments on 26 MOOCs with over 145,000 combined enrollments and millions of interactions show that models combining interaction data and course information have comparable or better performance than models which have access to previous iterations of the course. With these models, we aim to effectively enable educators to warm-start their predictions for new and ongoing courses.

Vinitra Swamy, Jibril Frej, Tanja Käser

Explainable Artificial Intelligence (XAI) plays a crucial role in enabling human understanding and trust in deep learning systems. As models get larger, more ubiquitous, and pervasive in aspects of daily life, explainability is necessary to minimize adverse effects of model mistakes. Unfortunately, current approaches in human-centric XAI (e.g. predictive tasks in healthcare, education, or personalized ads) tend to rely on a single post-hoc explainer, whereas recent work has identified systematic disagreement between post-hoc explainers when applied to the same instances of underlying black-box models. In this paper, we therefore present a call for action to address the limitations of current state-of-the-art explainers. We propose a shift from post-hoc explainability to designing interpretable neural network architectures. We identify five needs of human-centric XAI (real-time, accurate, actionable, human-interpretable, and consistent) and propose two schemes for interpretable-by-design neural network workflows (adaptive routing with InterpretCC and temporal diagnostics with I2MD). We postulate that the future of human-centric XAI is neither in explaining black-boxes nor in reverting to traditional, interpretable models, but in neural networks that are intrinsically interpretable.

Vinitra Swamy, Davide Romano, Bhargav Srinivasa Desikan et al.

Recent advances in eXplainable AI (XAI) for education have highlighted a critical challenge: ensuring that explanations for state-of-the-art AI models are understandable for non-technical users such as educators and students. In response, we introduce iLLuMinaTE, a zero-shot, chain-of-prompts LLM-XAI pipeline inspired by Miller's cognitive model of explanation. iLLuMinaTE is designed to deliver theory-driven, actionable feedback to students in online courses. iLLuMinaTE navigates three main stages - causal connection, explanation selection, and explanation presentation - with variations drawing from eight social science theories (e.g. Abnormal Conditions, Pearl's Model of Explanation, Necessity and Robustness Selection, Contrastive Explanation). We extensively evaluate 21,915 natural language explanations of iLLuMinaTE extracted from three LLMs (GPT-4o, Gemma2-9B, Llama3-70B), with three different underlying XAI methods (LIME, Counterfactuals, MC-LIME), across students from three diverse online courses. Our evaluation involves analyses of explanation alignment to the social science theory, understandability of the explanation, and a real-world user preference study with 114 university students containing a novel actionability simulation. We find that students prefer iLLuMinaTE explanations over traditional explainers 89.52% of the time. Our work provides a robust, ready-to-use framework for effectively communicating hybrid XAI-driven insights in education, with significant generalization potential for other human-centric fields.

Vinitra Swamy, Syrielle Montariol, Julian Blackwell et al.

In human-centric settings like education or healthcare, model accuracy and model explainability are key factors for user adoption. Towards these two goals, intrinsically interpretable deep learning models have gained popularity, focusing on accurate predictions alongside faithful explanations. However, there exists a gap in the human-centeredness of these approaches, which often produce nuanced and complex explanations that are not easily actionable for downstream users. We present InterpretCC (interpretable conditional computation), a family of intrinsically interpretable neural networks at a unique point in the design space that optimizes for ease of human understanding and explanation faithfulness, while maintaining comparable performance to state-of-the-art models. InterpretCC achieves this through adaptive sparse activation of features before prediction, allowing the model to use a different, minimal set of features for each instance. We extend this idea into an interpretable, global mixture-of-experts (MoE) model that allows users to specify topics of interest, discretely separates the feature space for each data point into topical subnetworks, and adaptively and sparsely activates these topical subnetworks for prediction. We apply InterpretCC for text, time series and tabular data across several real-world datasets, demonstrating comparable performance with non-interpretable baselines and outperforming intrinsically interpretable baselines. Through a user study involving 56 teachers, InterpretCC explanations are found to have higher actionability and usefulness over other intrinsically interpretable approaches.

Vinitra Swamy

Deep neural networks form the backbone of artificial intelligence research, with potential to transform the human experience in areas ranging from autonomous driving to personal assistants, healthcare to education. However, their integration into the daily routines of real-world classrooms remains limited. It is not yet common for a teacher to assign students individualized homework targeting their specific weaknesses, provide students with instant feedback, or simulate student responses to a new exam question. While these models excel in predictive performance, this lack of adoption can be attributed to a significant weakness: the lack of explainability of model decisions, leading to a lack of trust from students, parents, and teachers. This thesis aims to bring human needs to the forefront of eXplainable AI (XAI) research, grounded in the concrete use case of personalized learning and teaching. We frame the contributions along two verticals: technical advances in XAI and their aligned human studies. We investigate explainability in AI for education, revealing systematic disagreements between post-hoc explainers and identifying a need for inherently interpretable model architectures. We propose four novel technical contributions in interpretability with a multimodal modular architecture (MultiModN), an interpretable mixture-of-experts model (InterpretCC), adversarial training for explainer stability, and a theory-driven LLM-XAI framework to present explanations to students (iLLuMinaTE), which we evaluate in diverse settings with professors, teachers, learning scientists, and university students. By combining empirical evaluations of existing explainers with novel architectural designs and human studies, our work lays a foundation for human-centric AI systems that balance state-of-the-art performance with built-in transparency and trust.

Vinitra Swamy, Angelika Romanou, Martin Jaggi

Transformer-based language models trained on large text corpora have enjoyed immense popularity in the natural language processing community and are commonly used as a starting point for downstream tasks. While these models are undeniably useful, it is a challenge to quantify their performance beyond traditional accuracy metrics. In this paper, we compare BERT-based language models through snapshots of acquired knowledge at sequential stages of the training process. Structured relationships from training corpora may be uncovered through querying a masked language model with probing tasks. We present a methodology to unveil a knowledge acquisition timeline by generating knowledge graph extracts from cloze "fill-in-the-blank" statements at various stages of RoBERTa's early training. We extend this analysis to a comparison of pretrained variations of BERT models (DistilBERT, BERT-base, RoBERTa). This work proposes a quantitative framework to compare language models through knowledge graph extraction (GED, Graph2Vec) and showcases a part-of-speech analysis (POSOR) to identify the linguistic strengths of each model variant. Using these metrics, machine learning practitioners can compare models, diagnose their models' behavioral strengths and weaknesses, and identify new targeted datasets to improve model performance.