Proceedings of the Impactful and Responsible AI Systems for Education Workshop Held in COEX Convention and Exhibition Center, Seoul, South Korea on 28 June 2026 Published as Volume 339 by the Proceedings of Machine Learning Research on 10 August 2026. Volume Edited by: Debshila Basu Mallick Simon Woodhead Zichao Wang Muktha Ananda Jill Burstein April Murphy Series Editors: Neil D. Lawrence Hoel Kervadec Tegan Emerson https://proceedings.mlr.press/v339/ Mon, 10 Aug 2026 07:31:03 +0000 Mon, 10 Aug 2026 07:31:03 +0000 Jekyll v3.10.0 Monitoring classroom-wide students’ visible behavioral engagement is important for teacher reflection and instructional adjustment, yet video-based classroom analytics raise substantial privacy concerns when students may be identifiable from visual data. To address this tension, this paper presents a teacher-facing, region-level classroom engagement analytics framework that converts classroom video into temporal and spatial summaries while reducing reliance on identity-sensitive visual information. The framework combines stochastic head obfuscation, adversarial identity decoupling through a gradient reversal layer, and differentially private release of aggregated engagement summaries through Laplace noise injection. Using 21 short-duration classroom video segments from the Dataset for Classroom Group Engagement Recognition and a 13-minute extended instructional sequence, we examine what interpretable temporal and spatial signals the system can generate for classroom reflection across short-duration and extended analyses. We also conduct a professional evaluation with four middle school teachers to examine whether these outputs are understandable and aligned with teachers’ professional observations. The generated outputs and teacher evaluation suggest that region-level timelines and heatmaps can help teachers review moments and classroom areas that warrant contextual interpretation. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/xiang26a.html https://proceedings.mlr.press/v339/xiang26a.html Can AI support the research reasoning required to use unfamiliar educational datasets? We evaluated an AI-assisted discovery tool across two educational archives, ASSISTments and SEDA, using 50-item expert-designed assessments. We varied the sources available to the tool: dataset documentation, published papers, synthetic data samples, and their combination. Full triangulation across all three sources outperformed documentation alone in overall accuracy: 89% versus 67% on ASSISTments and 90% versus 73% on SEDA. However, more context was not always better. With codebook-oriented documentation, adding synthetic data without papers reduced appropriate-use accuracy from 80% to 40%, while adding papers without data reduced code-generation accuracy from 90% to 75%. Effective AI-assisted research support requires matching the information sources to how well a dataset’s design fits the research questions. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/wei26a.html https://proceedings.mlr.press/v339/wei26a.html Educational AI agents are usually introduced as tutors, homework helpers, or study companions. As children use them regularly, learning help can become a disclosure site: a student may begin with a math problem and then describe bullying, fear of adult reaction, or school avoidance. This position paper asks how child-facing educational agents should preserve student trust while deciding what stays private, what can be shared under student control, and what must be routed to accountable adults. Full parental visibility would turn help-seeking into surveillance; full secrecy would leave serious harm unsupported. We argue for bounded confidentiality as a communication-governance framework: ordinary tutoring remains private unless student choice or credible danger creates a reason to involve adults. The paper specifies a response ladder for minimal records, student-reviewed communication, care-oriented adult prompts, and escalation to locally accountable humans. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/shi26a.html https://proceedings.mlr.press/v339/shi26a.html Benchmarks for evaluating the pedagogical ability of LLM tutors are increasingly central to educational AI research, yet their psychometric properties are rarely examined formally. We apply a comprehensive measurement validation pipeline to MRBench V2, a benchmark of 200 mathematical tutoring dialogues annotated across eight pedagogical dimensions. Using exploratory and confirmatory factor analyses, graded response modelling, item-level validity diagnostics, measurement invariance testing, and generalizability theory, we find that six of the eight dimensions form a coherent unidimensional scale with excellent structural fit (CFI $= 0.998$, RMSEA $= 0.058$) and strong generalizability ($G_\text{rel} = 0.974$). Two dimensions show psychometric properties inconsistent with their intended role. We further detect measurement non-equivalence across model sizes and highlight open questions regarding gaps in construct validity and predictive validity. Our findings demonstrate the value of psychometric analysis as a standard practice in educational AI evaluation. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/sharma26a.html https://proceedings.mlr.press/v339/sharma26a.html Large language models are increasingly used to generate educational content, but many existing approaches focus on a single question format and rely on human review or same model validation for quality assurance. This limits scalability and can leave systematic evaluator blind spots unaddressed. We present \textit{HeteroForge}, a three pass pipeline for curriculum-aligned STEM exercise generation. Pass 1 uses separate prompts and JSON schemas for different question formats, followed by validation and self-refinement, to produce structurally valid exercises from curriculum metadata. Pass 2 performs heterogeneous multi-agent debate, where three models critique the exercises across accuracy, grade appropriateness, and scaffolded rigor before a judge model renders a structured verdict. Pass 3 recalibrates rejected exercises using judge feedback and a persistent curriculum-specific evaluation summary. The pipeline is modular across question formats, curriculum taxonomies, and model backends, allowing new formats, curricula, or evaluators to be added without changing the overall architecture. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/scaria26b.html https://proceedings.mlr.press/v339/scaria26b.html Generative AI is increasingly being positioned as a tutoring and learning-support technology, often through chatbot-based interfaces. While flexible, this design can make it difficult to constrain system behavior, align interactions with curriculum goals, and connect learner activity to demonstrated skill performance. This position paper proposes \textit{Learning in Blocks}, a skill-driven framework for AI-supported learning in which learners engage with structured assessments and practice tasks rather than open-ended conversation as the primary interface. The framework organizes learning into blocks of target and prerequisite skills. Bounded pedagogical agents support assessment generation, assessment evaluation, diagnostic recommendation, spaced review, and mastery-based progression. We argue that responsible AI in education should center structured evidence of learner skill. By integrating adaptive learning, formative assessment, multifaceted evaluation, spaced repetition, and mastery learning into a single loop, Learning in Blocks offers a design pattern for more transparent, auditable, and pedagogically aligned AI-supported learning. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/scaria26a.html https://proceedings.mlr.press/v339/scaria26a.html We present EditoriaLLM, a web application that uses large language models to generate competitive programming editorials from problem statements, reference solutions, algorithmic tags, audience settings, and style templates. This approach streamlines using a general-purpose chat interface into a simple, all-in-one tool, tailored to our particular use-case. EditoriaLLM is designed as a human-in-the-loop assistant for competitive programming education, with early adoption already observed in national competition contexts. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/moroianu26a.html https://proceedings.mlr.press/v339/moroianu26a.html Conversational learner simulations are valuable tools for testing learning theories, evaluating instructional materials and automated tutors, or powering teachable agents. Recently, large language models (LLM) have enabled richer, more naturalistic interactions with simulated learners; however, no open framework exists for evaluating whether such simulations faithfully reproduce real learner behavior. We introduce \textbf{EvalConvoLearn}, an open-source framework that assesses learner simulations along two axes: \emph{learning behavior} (skill-conditioned mastery outcomes) and \emph{conversational quality} (talk moves, error type distributions, question rate, turn length). EvalConvoLearn measures how closely a simulated learner approximates answer distributions observed in data by grounding metrics in authentic tutoring conversation datasets, and anchoring generated tutor responses in existing tutor utterances. The framework is demonstrated on a dataset of tutoring dialogues, including results for two LLM-based learner simulations, and the published GitHub code. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/moreau-pernet26a.html https://proceedings.mlr.press/v339/moreau-pernet26a.html Research on Knowledge Tracing (KT) models traditionally focuses on improving predictive accuracy. However, responsible real-world deployment requires models to know when to defer uncertain predictions to a human teacher. We introduce an intrinsic selective prediction layer for existing KT models using Monte Carlo Dropout (MC-Dropout) to quantify uncertainty. We evaluate this approach across three architectures (DKT, SAKT, and AKT) using the Eedi mathematics dataset. Abstaining on the 20% most uncertain predictions lifts accuracy by 2.3 to 3.0 percentage points, AUC by 1.9 to 2.4 percentage points and F1 by 1.4 to 4.3 percentage points without any retraining. This abstention strategy is highly targeted: the deferred set exhibits 1.45 to 1.60 times the error rate of the kept set. Furthermore, this targeting holds within every question-difficulty quartile and remains fair across student-ability levels. Importantly, MC-Dropout variance gives roughly five times the AUC lift of a calibrated two-parameter logistic (2PL) Item Response Theory (IRT) baseline as a selective-prediction signal. A variance decomposition of the model’s epistemic uncertainty (BALD) reveals that the entire classical psychometric stack, comprising question difficulty, student ability, IRT-style outcome ambiguity, and historical curriculum coverage, explains less than 4% of the signal under linear modeling and at most 23% even with a non-linear regressor. This leaves 77% to 90% as architecture-specific epistemic content that MC-Dropout surfaces and simpler proxies cannot recover. Selective prediction with model-native epistemic uncertainty is therefore a necessary component of responsible KT deployment, complementary to subgroup-fairness audits and downstream classroom evaluation rather than a substitute for them. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/mitton26a.html https://proceedings.mlr.press/v339/mitton26a.html Though LLM-simulated practice may support psychotherapy education, open questions remain about which system features provide the most educational value. We developed an AI-simulated practice tool for use in psychotherapy classrooms that provides speaking-practice with LLM-based patients, followed by a post-practice activity that includes feedback from a fine-tuned LLM and written reflection exercises. We deployed the tool in a graduate psychotherapy course (n=25, 5 weeks) using a micro-randomized trial (MRT)—a method which can estimate the causal excursion effect of an intervention by leveraging longitudinal repeated randomization with participants. Testing four conditions crossing AI feedback (present/absent) with reflection granularity (utterance/session-level), we found surprising interaction effects on student engagement and educational value: utterance-level reflection paired with AI feedback significantly outperformed all conditions. Students valued comparing the AI suggestions to their own rather than passively accepting them. As the first MRT in a health education setting testing an LLM-simulation system, this work demonstrates that MRTs, a method mostly used in mobile intervention research, offers a viable evaluation paradigm for AI systems deployed in health education settings. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/louie26a.html https://proceedings.mlr.press/v339/louie26a.html As Generative Artificial Intelligence (GenAI) capabilities rapidly expand in educational settings, assessing the validity, reliability, and fairness of these tools is critical before widespread classroom deployment. While there has been a large amount of research, it is not yet clear what methods are being used to evaluate these foundational concepts and what the findings are. We present a work-in-progress systematic literature review focusing exclusively on automated item scoring. Based on a final coded corpus of 107 empirical studies published since January 2023, our preliminary analysis reveals that while prompt-based Large Language Models (LLMs) are achieving parity with traditional transformer-based methods, significant gaps remain regarding fairness evaluation and the representation of diverse student populations. Quadratic weighted kappa (QWK) is the dominant agreement metric, used in 41.8% of studies, and fairness is acknowledged by 55% of studies but empirically evaluated in only 11.2%. This paper offers evidence-based guidance for responsible classroom integration of AI. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/li26a.html https://proceedings.mlr.press/v339/li26a.html Learning at scale increasingly incorporates generative Al assets, with developers of products and services finding it useful to incorporate multiple Al components in their toolchain and to power particular features. Yet, the practices of choosing, tuning, and measuring Al components for inclusion in a broader product remains somewhat ad hoc, particularly with regard to evaluating Al according to learning sciences concepts. This has led researchers, industry, and funders to call for public goods that would be useful as infrastructure: datasets, benchmarks, and models that can measure learning sciences concepts, strengthen their operationalization, and provide guardrails to protect student privacy and safety. At the launch of a funding program related to this call, a public Request for Information (RFI) was launched and received over 100 responses from researchers, industry, and educators all over the world. Here we report on a work-in-progress to analyze the RFI responses. Using the responses, we answer questions that include: What kinds of infrastructural datasets, models and benchmarks are most needed to better include the learning sciences? What dimensions of learner variability should be emphasized? What privacy, safety and data governance issues are most in need of guardrails? This work presents community-informed insights on AI infrastructure in K-12 education as an initial roadmap to transition towards evidence-based public goods that serve all learners at scale. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/lee26b.html https://proceedings.mlr.press/v339/lee26b.html This study examines how a transdisciplinary team experienced the co-design of an AI-enhanced Multimodal Learning Analytics platform for collaborative secondary Science, Technology, Engineering, and Mathematics classrooms. While artificial intelligence and learning analytics systems are often framed in terms of efficiency and personalization, far less is understood about how such systems are collaboratively designed across stakeholders. Drawing on sociocultural theory and participatory design, this study analyzes in-depth semi-structured interviews with four members of a co-design team engaged in a year-long, online collaboration. Using interpretative phenomenological analysis and theory-informed social network analysis, we examine both participants’ lived experiences and patterns of influence shaping the co-design process. Findings show that co-design unfolded as a negotiated process across personal, interpersonal, and community planes. Educator expertise anchored problem definition and guided development, while roles and authority emerged through interaction. Care and trust sustained collaboration, and participation became a site of adult learning and identity work. Network analysis reveals a dense and reciprocal ecology of influence extending beyond the team. This study reframes co-design not as a procedural step for feedback incorporation, but as a relational process through which educational futures with AI are collectively imagined. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/lee26a.html https://proceedings.mlr.press/v339/lee26a.html Large Language Models (LLMs) offer great potential to reduce Teaching Assistant workloads through automated grading. However, their rapid integration shifts the primary challenge from technical scalability to responsible governance, as LLMs frequently diverge from expert human graders on complex rubrics. Existing Human-in-the-Loop (HITL) systems focus on micro-level calibration for individual submissions, leaving instructors blind to how rubrics alter class-wide grade distributions. To address this gap, we introduce \textbf{Co-TA}, a deployable web application that streamlines the grading workflow while shifting the paradigm to rubric co-design and fairness auditing. Featuring batch uploads and uncertainty flagging, Co-TA allows instructors to generate rubric variations and immediately simulate their impact against synthetic student personas. By visualizing resulting grade distributions prior to live grading, Co-TA serves as a critical fairness auditing mechanism, ensuring AI assessment remains transparent, verifiable, and firmly under human pedagogical control. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/jin26a.html https://proceedings.mlr.press/v339/jin26a.html The rapid integration of Artificial Intelligence (AI) into K-12 education presents both significant instructional opportunities and emerging governance challenges. While existing research has largely focused on the technical capabilities or theoretical implications of AI in education, limited empirical work captures how educators themselves perceive its use in real classroom settings, particularly at the district and state levels. This study addresses that gap by examining educator perspectives on AI adoption across nine school districts in Arkansas, in collaboration with the Arkansas Department of Higher Education. Using an IRB-approved survey instrument, data were collected from 46 educators across five domains: instructional and administrative benefits, classroom risks and downsides, trust and ethical considerations, governance and oversight structures, and professional support and training needs. Findings indicate strong recognition of AI’s efficiency and instructional value, particularly in lesson planning, grading, and content generation. However, these benefits are accompanied by substantial concerns, including the loss of critical thinking, risks to academic integrity, misinformation, and data privacy risks. Notably, results reveal a pronounced gap between individual-level AI usage and institutional-level trust, with educators expressing strong demand for structured oversight, standardized tools, and formal training. Building on these findings, this study proposes a structured governance framework for AI integration in K-12 education that links technological capability, risk emergence, institutional trust, and policy design to the preservation of student cognitive autonomy. The framework emphasizes that the central challenge of AI adoption is not technological readiness, but the absence of coordinated governance mechanisms that align AI use with educational objectives. By providing localized, policy-relevant evidence and a conceptual model for responsible integration, this study contributes to ongoing discussions on AI governance, educational standardization, and the future of teaching and learning in the AI era. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/jeon26a.html https://proceedings.mlr.press/v339/jeon26a.html Spaced repetition is a technique widely used in digital learning applications (notably for language practice) that determines the frequency with which concepts need to be reviewed by learners in order to boost long-term retention, scheduling reviews at increasing intervals. Research in the area of AI in education investigates how data from digital learning applications can be leveraged to develop predictive models that enable personalization of spaced repetition, and as such maximize learning efficiency. Previous research on Duolingo’s digital platform for language practice developed a personalized model (the Half-Life Regression model) that predicts when lexemes (abstract units of mearning) are likely to be forgotten by individual learners. A limitation of previous research is its reliance on a small set of features, which restricts the performance of the original model. Moreover, attempts to improve this performance often rely on deep learning methods, which in turn reduce the model’s explainability. In the current study, we propose an alternative approach (combining psychometrics with machine learning) that incorporates relevant learner, lexeme and contextual features (e.g., time interval since the lexeme was last shown) aimed at improving both the performance and explainability of earlier Half-Life Regression models. More specifically, we introduce a mixed-effects version of Half-Life Regression, and use the estimated random effects as input to a Random Forest. In comparison with the original model, our results show improvements in predictive performance across a range of measures. SHAP analysis of the best-performing method (the Random Forest) indicates the importance of contextual features and random-effects features that are likely to reflect aspects like lexeme difficulty and learner ability. By achieving both explainability and higher predictive performance than the earlier models, our work contributes to more impactful and responsible use of AI for personalized spaced repetition in language learning. \begin{keywords} spaced repetition, half-life regression, language learning, AI in education\end{keywords} Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/ilidio26a.html https://proceedings.mlr.press/v339/ilidio26a.html Large language models (LLMs) are increasingly used to provide automated feedback in introductory programming courses, yet empirical evidence from authentic classroom deployments comparing different feedback modalities remains limited. In this work, we present a large-scale classroom study in which AI-generated feedback was deployed through a randomized protocol in an introductory Python programming course. Students received one of three feedback conditions on incorrect submissions: natural language hints, AI-generated failing test cases, or no AI feedback. We release the resulting dataset, \textsc{ProgFeed}, which captures 6{,}693 submissions from 215 consenting students across 17 labs, including feedback conditions, execution-based performance measures, and fine-grained temporal information. Using this data, we analyze learning trajectories, feedback quality, and submission behavior over repeated attempts. We find that natural language feedback is significantly associated with higher completion rates and faster convergence to correct solutions. Test case feedback, by contrast, exhibits heterogeneous effects that depend critically on feedback validity. Our results suggest that the form of AI-generated feedback matters, and that evaluating feedback quality—not just its presence—is essential for understanding its pedagogical impact. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/heickal26a.html https://proceedings.mlr.press/v339/heickal26a.html AI-powered educational tools proliferate globally, reaching hundreds of millions of learners through national education platforms, EdTech products, and AI tutoring systems. A critical gap persists between the pace of AI deployment and the depth of cognitive grounding that underlies these systems. This paper argues that responsible AI for education requires a \emph{cognitive grounding phase} before scaling: a systematic, empirically validated assessment of what existing educational systems actually demand cognitively, and whether AI tools are designed to enhance or merely replicate those demands. Drawing on a decade of empirical research classifying over 3,000 assessment items from national school board examinations using the revised Bloom’s Taxonomy, we demonstrate that the dominant examination systems in a major Global South context overwhelmingly test the lowest cognitive levels, precisely the levels at which large language models perform most capably. We introduce the concept of digital rote to describe the phenomenon in which AI systems technologically reproduce pedagogically shallow learning at scale, and propose a three-stage Cognitive Grounding Framework — Cognitive Audit, Cognitive Specification, and Cognitive Validation — developed in the context of LLM-based assessment generation and extensible in principle to other AI-in-education applications. We demonstrate the framework’s practical viability through a patented AI-based assessment system whose design was directly informed by cognitive-level analysis. Finally, we discuss policy implications for national-scale AI deployment in education and outline open questions for the responsible AI-in-education community. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/das26a.html https://proceedings.mlr.press/v339/das26a.html Rapid model progress can make educational benchmark results stale, but responsible evaluation requires more than refreshing aggregate scores. We revisit a benchmark of 376 curriculum-authentic, image-required middle-school mathematics items using GPT-5.5 and compare its behavior with representative prior models. In the \textit{with-images} condition, accuracy increases from about 88% for GPT-5.4 to about 93% for GPT-5.5, while the strict never-solved set decreases from 31 items to 17. In contrast, \textit{without-images} performance remains near zero, confirming that the benchmark remains genuinely image-required. We then audit this residual set and find that its composition clarifies what the remaining failures mean: they include inaccessible-image cases, insufficient-context items, answer-format mismatches, and benchmark-quality issues rather than only generic visual-reasoning failures. We treat representation-sensitive probes as future work, using the audit to identify which residual items are appropriate probe candidates. We argue that continuous evaluation of educational AI systems should pair aggregate performance updates with residual item audits that distinguish model limitations from dataset, scoring, accessibility, and task-specification issues. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/croteau26a.html https://proceedings.mlr.press/v339/croteau26a.html Large language models (LLMs) have shown considerable promise for automatic feedback generation (AFG) in educational settings, yet their inaccuracies and unpredictable failures underscore the need for robust quality assurance (QA) before deployment at scale. Traditional QA approaches relying on human raters are inherently time-consuming, costly, and difficult to scale. This study extends the LLM-as-judge paradigm beyond evaluation to introduce a multi-agent QA system that not only assesses but autonomously revises LLM-generated feedback until predefined quality standards are met. The system comprises seven specialized agents operating in a closed evaluate–revise–re-evaluate loop. Skill-level feedback from a formative assessment measuring professional competencies ($n = 841$) is evaluated against six quality criteria using chain-of-thought prompting, and revised using role-based prompting where quality thresholds are not met. The system achieved a mean QA total of 5.99 out of 6 across all records, approved 99.0% of feedback within three iterations, and escalated only 1.0% of cases for human review. A single-agent baseline is compared against the proposed system to assess the contribution of the multi-agent architecture. The baseline performed comparably on five of the six criteria but failed markedly on second-person language (32.1% vs. 100%) and escalated 68.0% of records, compared to 1.0% in the proposed system, demonstrating that distributing QA responsibilities across specialized agents substantially improves both evaluation precision and revision effectiveness. A criterion-level stress test on 180 distorted records demonstrated a 100% correction rate for all detected violations and a 0% false positive rate, with detection rates ranging from 80.0% for actionability to 100% for second-person language. Human expert review of flagged and revised cases largely confirmed the system’s judgments, though reviewers noted that certain linguistic corrections, particularly those involving punctuation, were unlikely to meaningfully affect student comprehension. These findings establish the viability of multi-agent LLM systems as scalable, automated QA and remediation pipelines for AFG. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/bulut26a.html https://proceedings.mlr.press/v339/bulut26a.html Peer-argumentation is an effective and widely adopted learning activity, yet it remains difficult to enact at scale. While pedagogical conversational agents (PCAs) have been extensively studied, their role as arguing peers that engage students in dialogic argumentation is underexplored. To address this gap, we present ArguBot, a didactic AI arguing partner designed to support scalable case-based peer-argumentation activities. ArguBot deliberately adopts an opposing stance, challenging students’ correct claims with common student misconceptions and countering incorrect claims with arguments grounded in course lecture materials, supported by Retrieval-Augmented Generation (RAG). We conducted a semester-long field study in a graduate robotics course of 172 students, in which both in-class peer-peer argumentation and after-class peer-AI argumentation with ArguBot were offered under regular instructional conditions. Analyses draw on pre-post multiple-choice question responses from students who voluntarily participated in in-class activities ($n = 85$) and/or engaged with ArguBot after class ($n = 68$), as well as post-semester perception surveys ($n = 57$). Results show no statistically significant differences between conditions in error correction or correctness preservation. In contrast, students rated the AI arguing partner significantly lower than a human peer, including reporting lower levels of interest/enjoyment. These findings provide early, ecologically valid evidence on the opportunities and challenges of deploying AI arguing agents in higher education to scale argumentation-based learning. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/brender26a.html https://proceedings.mlr.press/v339/brender26a.html Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/basu-mallick26a.html https://proceedings.mlr.press/v339/basu-mallick26a.html Annotating geometric figures from scanned documents has long been addressed by adapting generic annotation tools, tools not originally designed for such tasks, to use cases where they are suboptimal. An interactive web interface is described that is purpose-built for validating automatically generated geometry figure descriptions, allowing annotators to review and correct conditional declaration language (CDL) descriptions while simultaneously adjusting figure crops and editing source problem text. Submissions pass through two independent annotators in sequence, with each round fully logged. The interface is currently deployed and has been used by 12 annotators to validate 483 problem entries. Mon, 10 Aug 2026 00:00:00 +0000 https://proceedings.mlr.press/v339/badea26a.html https://proceedings.mlr.press/v339/badea26a.html