Kobi Gal

Hagai Astrin, Ayal Swaid, Avi Segal et al.

Mental health challenges are increasing worldwide, straining emotional support services and leading to counselor overload. This can result in delayed responses during critical situations, such as suicidal ideation, where timely intervention is essential. While large language models (LLMs) have shown strong generative capabilities, their application in low-resource languages, especially in sensitive domains like mental health, remains underexplored. Furthermore, existing LLM-based agents often struggle to replicate the supportive language and intervention strategies used by professionals due to a lack of training on large-scale, real-world datasets. To address this, we propose CARE (Counselor-Aligned Response Engine), a GenAI framework that assists counselors by generating real-time, psychologically aligned response recommendations. CARE fine-tunes open-source LLMs separately for Hebrew and Arabic using curated subsets of real-world crisis conversations. The training data consists of sessions rated as highly effective by professional counselors, enabling the models to capture interaction patterns associated with successful de-escalation. By training on complete conversation histories, CARE maintains the evolving emotional context and dynamic structure of counselor-help-seeker dialogue. In experimental settings, CARE demonstrates stronger semantic and strategic alignment with gold-standard counselor responses compared to non-specialized LLMs. These findings suggest that domain-specific fine-tuning on expert-validated data can significantly support counselor workflows and improve care quality in low-resource language contexts.

Anastasia Kovalkov, Benjamin Paaßen, Avi Segal et al.

Promoting creativity is considered an important goal of education, but creativity is notoriously hard to measure.In this paper, we make the journey fromdefining a formal measure of creativity that is efficientlycomputable to applying the measure in a practical domain. The measure is general and relies on coretheoretical concepts in creativity theory, namely fluency, flexibility, and originality, integratingwith prior cognitive science literature. We adapted the general measure for projects in the popular visual programming language Scratch.We designed a machine learning model for predicting the creativity of Scratch projects, trained and evaluated on human expert creativity assessments in an extensive user study. Our results show that opinions about creativity in Scratch varied widely across experts. The automatic creativity assessment aligned with the assessment of the human experts more than the experts agreed with each other. This is a first step in providing computational models for measuring creativity that can be applied to educational technologies, and to scale up the benefit of creativity education in schools.

Amir Bialer, Daniel Izmaylov, Avi Segal et al.

With the increased awareness of situations of mental crisis and their societal impact, online services providing emergency support are becoming commonplace in many countries. Computational models, trained on discussions between help-seekers and providers, can support suicide prevention by identifying at-risk individuals. However, the lack of domain-specific models, especially in low-resource languages, poses a significant challenge for the automatic detection of suicide risk. We propose a model that combines pre-trained language models (PLM) with a fixed set of manually crafted (and clinically approved) set of suicidal cues, followed by a two-stage fine-tuning process. Our model achieves 0.91 ROC-AUC and an F2-score of 0.55, significantly outperforming an array of strong baselines even early on in the conversation, which is critical for real-time detection in the field. Moreover, the model performs well across genders and age groups.

Roy Fairstein, Harper Lyon, Oshri Damty et al.

Peer-evaluation and selection systems are used when sets of agents evaluate each other in order to select the best $k$ among them. These are commonly used in real-world settings, including academic conferences where those reviewing papers are often the set of submitters. Conferences have attempted to better allocate their reviewing resources by moving to a two-stage mechanism, in which some papers are eliminated after a first stage of review and remaining papers receive additional reviewers. We investigate how two major strategyproof peer selection mechanisms, Partition and ExactDollarPartition, perform when adapted to a two-stage system, in order to try and understand the effect of the two-stage mechanism on which agents get selected. We also examine how the various parameters of the two-stage mechanism influence the outcome. We provide a theoretical basis by showing how a particular setting is influenced by the two stages. However, solving for the general case seems implausible at the moment, and we use extensive simulations of different scenarios and settings to observe which agents benefit and which are harmed by adopting two-stage mechanisms (and we vary this mechanisms parameters as well). We show that the two-stage mechanism's advantage depends the noisiness of reviewer beliefs. Borderline agents benefit most in a low noise environment, while high rank agents benefit more in noisy environments. We show that the effectiveness of these mechanisms is highly dependent on the number of chosen agents, the number of reviews requested from agents, and reviewers' correlation, indicating that organizers need to exercise caution when selecting these parameters for a reviewing process.

Ziv Fenigstein, Kobi Gal, Avi Segal et al.

Assessing teachers' geometric content knowledge is essential for geometry instructional quality and student learning, but difficult to scale. The Van Hiele model characterizes geometric reasoning through five hierarchical levels. Traditional Van Hiele assessment relies on manual expert analysis of open-ended responses. This process is time-consuming, costly, and prevents large-scale evaluation. This study develops an automated approach for diagnosing teachers' Van Hiele reasoning levels using large language models grounded in educational theory. Our central hypothesis is that integrating explicit skills information significantly improves Van Hiele classification. In collaboration with mathematics education researchers, we built a structured skills dictionary decomposing the Van Hiele levels into 33 fine-grained reasoning skills. Through a custom web platform, 31 pre-service teachers solved geometry problems, yielding 226 responses. Expert researchers then annotated each response with its Van Hiele level and demonstrated skills from the dictionary. Using this annotated dataset, we implemented two classification approaches: (1) retrieval-augmented generation (RAG) and (2) multi-task learning (MTL). Each approach compared a skills-aware variant incorporating the skills dictionary against a baseline without skills information. Results showed that for both methods, skills-aware variants significantly outperformed baselines across multiple evaluation metrics. This work provides the first automated approach for Van Hiele level classification from open-ended responses. It offers a scalable, theory-grounded method for assessing teachers' geometric reasoning that can enable large-scale evaluation and support adaptive, personalized teacher learning systems.

Max Norris, Kobi Gal, Sahan Bulathwela

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.

Eliya Naomi Aharon, Meytal Grimland, Avi Segal et al.

Effective mental health counseling is a complex, theory-driven process requiring the simultaneous integration of psychological frameworks, real-time distress signals, and strategic intervention planning. This level of clinical reasoning is critical for safety and therapeutic effectiveness but is often missing in general-purpose Large Language Models (LLMs). We introduce SAGE (Strategy-Aware Graph-Enhanced), a novel framework designed to bridge the gap between structured clinical knowledge and generative AI. SAGE constructs a heterogeneous graph that unifies conversational dynamics with a psychologically grounded layer, explicitly anchoring interactions in a theory-driven lexicon. Our architecture first employs a Next Strategy Classifier to identify the optimal therapeutic intervention. Subsequently, a Graph-Aware Attention mechanism projects graph-derived structural signals into soft prompts, conditioning the LLM to generate responses that maintain clinical depth. Validated through both automated metrics and expert human evaluation, SAGE outperforms baselines in strategy prediction and recommended response quality. By providing actionable intervention recommendations, SAGE serves as a cutting-edge decision-support tool designed to augment human expertise in high-stakes crisis counseling.

Roy Fairstein, Dan Vilenchik, Kobi Gal

Participatory Budgeting (PB) offers a democratic process for communities to allocate public funds across various projects through voting. In practice, PB organizers face challenges in selecting aggregation rules either because they are not familiar with the literature and the exact details of every existing rule or because no existing rule echoes their expectations. This paper presents a novel data-driven approach utilizing machine learning to address this challenge. By training neural networks on PB instances, our approach learns aggregation rules that balance social welfare, representation, and other societal beneficial goals. It is able to generalize from small-scale synthetic PB examples to large, real-world PB instances. It is able to learn existing aggregation rules but also generate new rules that adapt to diverse objectives, providing a more nuanced, compromise-driven solution for PB processes. The effectiveness of our approach is demonstrated through extensive experiments with synthetic and real-world PB data, and can expand the use and deployment of PB solutions.

Noga Schwartz, Roy Fairstein, Avi Segal et al.

Early detection of struggling student programmers is crucial for providing them with personalized support. While multiple AI-based approaches have been proposed for this problem, they do not explicitly reason about students' programming skills in the model. This study addresses this gap by developing in collaboration with educators a taxonomy of proficiencies that categorizes how students solve coding tasks and is embedded in the detection model. Our model, termed the Proficiency Taxonomy Model (PTM), simultaneously learns the student's coding skills based on their coding history and predicts whether they will struggle on a new task. We extensively evaluated the effectiveness of the PTM model on two separate datasets from introductory Java and Python courses for beginner programmers. Experimental results demonstrate that PTM outperforms state-of-the-art models in predicting struggling students. The paper showcases the potential of combining structured insights from teachers for early identification of those needing assistance in learning to code.

Nicholas Hoernle, Rafael Michael Karampatsis, Vaishak Belle et al.

We propose a novel way to incorporate expert knowledge into the training of deep neural networks. Many approaches encode domain constraints directly into the network architecture, requiring non-trivial or domain-specific engineering. In contrast, our approach, called MultiplexNet, represents domain knowledge as a logical formula in disjunctive normal form (DNF) which is easy to encode and to elicit from human experts. It introduces a Categorical latent variable that learns to choose which constraint term optimizes the error function of the network and it compiles the constraints directly into the output of existing learning algorithms. We demonstrate the efficacy of this approach empirically on several classical deep learning tasks, such as density estimation and classification in both supervised and unsupervised settings where prior knowledge about the domains was expressed as logical constraints. Our results show that the MultiplexNet approach learned to approximate unknown distributions well, often requiring fewer data samples than the alternative approaches. In some cases, MultiplexNet finds better solutions than the baselines; or solutions that could not be achieved with the alternative approaches. Our contribution is in encoding domain knowledge in a way that facilitates inference that is shown to be both efficient and general; and critically, our approach guarantees 100% constraint satisfaction in a network's output.

Janet Rafner, Miroslav Gajdacz, Gitte Kragh et al.

Artificial Intelligence (AI) can augment and sometimes even replace human cognition. Inspired by efforts to value human agency alongside productivity, we discuss the benefits of solving Citizen Science (CS) tasks with Hybrid Intelligence (HI), a synergetic mixture of human and artificial intelligence. Currently there is no clear framework or methodology on how to create such an effective mixture. Due to the unique participant-centered set of values and the abundance of tasks drawing upon both human common sense and complex 21st century skills, we believe that the field of CS offers an invaluable testbed for the development of HI and human-centered AI of the 21st century, while benefiting CS as well. In order to investigate this potential, we first relate CS to adjacent computational disciplines. Then, we demonstrate that CS projects can be grouped according to their potential for HI-enhancement by examining two key dimensions: the level of digitization and the amount of knowledge or experience required for participation. Finally, we propose a framework for types of human-AI interaction in CS based on established criteria of HI. This "HI lens" provides the CS community with an overview of several ways to utilize the combination of AI and human intelligence in their projects. It also allows the AI community to gain ideas on how developing AI in CS projects can further their own field.

Anastasia Kovalkov, Avi Segal, Kobi Gal

Visual programming environments are increasingly part of the curriculum in schools. Their potential for promoting creative thinking of students is an important factor in their adoption. However, there does not exist a standard approach for detecting creativity in students' programming behavior, and analyzing programs manually requires human expertise and is time consuming. This work provides a computational tool for measuring creativity in visual programming that combines theory from the literature with data mining approaches. It adapts the classical dimensions of creative processes to our setting, as well as considering new aspects such as visual elements of the projects. We apply this approach to the Scratch programming environment, measuring the creativity score of hundreds of projects. We show that current metrics of computational thinking in Scratch fail to capture important aspects of creativity, such as the visual artifacts of projects. Interviews conducted with Scratch teachers validate our approach.

Jonathan Martinez, Kobi Gal, Ece Kamar et al.

AI systems that model and interact with users can update their models over time to reflect new information and changes in the environment. Although these updates may improve the overall performance of the AI system, they may actually hurt the performance with respect to individual users. Prior work has studied the trade-off between improving the system's accuracy following an update and the compatibility of the updated system with prior user experience. The more the model is forced to be compatible with a prior version, the higher loss in accuracy it will incur. In this paper, we show that by personalizing the loss function to specific users, in some cases it is possible to improve the compatibility-accuracy trade-off with respect to these users (increase the compatibility of the model while sacrificing less accuracy). We present experimental results indicating that this approach provides moderate improvements on average (around 20%) but large improvements for certain users (up to 300%).

Laura Schelenz, Avi Segal, Kobi Gal

Artificial Intelligence based systems increasingly use personalization to provide users with relevant content, products, and solutions. Personalization is intended to support users and address their respective needs and preferences. However, users are becoming increasingly vulnerable to online manipulation due to algorithmic advancements and lack of transparency. Such manipulation decreases users' levels of trust, autonomy, and satisfaction concerning the systems with which they interact. Increasing transparency is an important goal for personalization based systems. Unfortunately, system designers lack guidance in assessing and implementing transparency in their developed systems. In this work we combine insights from technology ethics and computer science to generate a list of transparency best practices for machine generated personalization. Based on these best practices, we develop a checklist to be used by designers wishing to evaluate and increase the transparency of their algorithmic systems. Adopting a designer perspective, we apply the checklist to prominent online services and discuss its advantages and shortcomings. We encourage researchers to adopt the checklist in various environments and to work towards a consensus-based tool for measuring transparency in the personalization community.

Nicholas Hoernle, Gregory Kehne, Ariel D. Procaccia et al.

Badges are commonly used in online platforms as incentives for promoting contributions. It is widely accepted that badges "steer" people's behavior toward increasing their rate of contributions before obtaining the badge. This paper provides a new probabilistic model of user behavior in the presence of badges. By applying the model to data from thousands of users on the Q&A site Stack Overflow, we find that steering is not as widely applicable as was previously understood. Rather, the majority of users remain apathetic toward badges, while still providing a substantial number of contributions to the site. An interesting statistical phenomenon, termed "Phantom Steering," accounts for the interaction data of these users and this may have contributed to some previous conclusions about steering. Our results suggest that a small population, approximately 20%, of users respond to the badge incentives. Moreover, we conduct a qualitative survey of the users on Stack Overflow which provides further evidence that the insights from the model reflect the true behavior of the community. We argue that while badges might contribute toward a suite of effective rewards in an online system, research into other aspects of reward systems such as Stack Overflow reputation points should become a focus of the community.

Benjamin Clavié, Kobi Gal

The use of large pretrained neural networks to create contextualized word embeddings has drastically improved performance on several natural language processing (NLP) tasks. These computationally expensive models have begun to be applied to domain-specific NLP tasks such as re-hospitalization prediction from clinical notes. This paper demonstrates that using large pretrained models produces excellent results on common learning analytics tasks. Pre-training deep language models using student forum data from a wide array of online courses improves performance beyond the state of the art on three text classification tasks. We also show that a smaller, distilled version of our model produces the best results on two of the three tasks while limiting computational cost. We make both models available to the research community at large.

Nicholas Hoernle, Kobi Gal, Barbara Grosz et al.

This paper describes methods for comparative evaluation of the interpretability of models of high dimensional time series data inferred by unsupervised machine learning algorithms. The time series data used in this investigation were logs from an immersive simulation like those commonly used in education and healthcare training. The structures learnt by the models provide representations of participants' activities in the simulation which are intended to be meaningful to people's interpretation. To choose the model that induces the best representation, we designed two interpretability tests, each of which evaluates the extent to which a model's output aligns with people's expectations or intuitions of what has occurred in the simulation. We compared the performance of the models on these interpretability tests to their performance on statistical information criteria. We show that the models that optimize interpretability quality differ from those that optimize (statistical) information theoretic criteria. Furthermore, we found that a model using a fully Bayesian approach performed well on both the statistical and human-interpretability measures. The Bayesian approach is a good candidate for fully automated model selection, i.e., when direct empirical investigations of interpretability are costly or infeasible.

Avi Segal, Kobi Gal, Guy Shani et al.

The prevalence of e-learning systems and on-line courses has made educational material widely accessible to students of varying abilities and backgrounds. There is thus a growing need to accommodate for individual differences in e-learning systems. This paper presents an algorithm called EduRank for personalizing educational content to students that combines a collaborative filtering algorithm with voting methods. EduRank constructs a difficulty ranking for each student by aggregating the rankings of similar students using different aspects of their performance on common questions. These aspects include grades, number of retries, and time spent solving questions. It infers a difficulty ranking directly over the questions for each student, rather than ordering them according to the student's predicted score. The EduRank algorithm was tested on two data sets containing thousands of students and a million records. It was able to outperform the state-of-the-art ranking approaches as well as a domain expert. EduRank was used by students in a classroom activity, where a prior model was incorporated to predict the difficulty rankings of students with no prior history in the system. It was shown to lead students to solve more difficult questions than an ordering by a domain expert, without reducing their performance.

Avi Segal, Yossi Ben David, Joseph Jay Williams et al.

As e-learning systems become more prevalent, there is a growing need for them to accommodate individual differences between students. This paper addresses the problem of how to personalize educational content to students in order to maximize their learning gains over time. We present a new computational approach to this problem called MAPLE (Multi-Armed Bandits based Personalization for Learning Environments) that combines difficulty ranking with multi-armed bandits. Given a set of target questions MAPLE estimates the expected learning gains for each question and uses an exploration-exploitation strategy to choose the next question to pose to the student. It maintains a personalized ranking over the difficulties of question in the target set which is used in two ways: First, to obtain initial estimates over the learning gains for the set of questions. Second, to update the estimates over time based on the students responses. We show in simulations that MAPLE was able to improve students' learning gains compared to approaches that sequence questions in increasing level of difficulty, or rely on content experts. When implemented in a live e-learning system in the wild, MAPLE showed promising results. This work demonstrates the efficacy of using stochastic approaches to the sequencing problem when augmented with information about question difficulty.