The researchers assigned 88 fifth-year medical students to one of two sequences according to their cardiology clerkship schedules. One group received the interactive 360 video before observing a catheterization laboratory, while the other observed the laboratory first and then completed the video module. Each student was therefore intended to experience both approaches, but allocation was not randomized. The design allowed the investigators to compare the approaches within participants while accounting for baseline understanding and the period in which each session occurred. Because sequence was determined by clerkship scheduling and the study was retrospective, the findings are associational rather than a controlled estimate of a causal teaching effect.

The module used 360-degree recordings made in an actual catheterization laboratory. Its functions included pointer-based object identification, playback synchronized with a headset, and real-time monitoring of learners’ gaze. These features differentiated it from passive 360-degree observation, although the study did not test which individual technical function, if any, was responsible for the observed difference. Self-rated understanding was measured at baseline, after the first session, and after the second session using the mean of 10 items, each rated from 1 to 5.

Baseline scores were comparable between the two sequences: 2.70 plus or minus 0.56 for the video-first group and 2.52 plus or minus 0.83 for the laboratory-first group, with p equal to 0.37. The largest separation occurred after the first session. Students assigned to interactive video first reported a mean understanding score of 4.34 plus or minus 0.58, whereas those assigned to catheterization-laboratory observation first averaged 3.41 plus or minus 0.92; the difference was reported as p less than 0.001. The interactive-video session was associated with significant score increases in either sequence. By contrast, the catheterization-laboratory session produced a significant increase only when it was delivered first.

After both sessions, the groups’ reported scores converged: 4.45 plus or minus 0.53 for the video-first sequence and 4.41 plus or minus 0.60 for the laboratory-first sequence, with p equal to 0.80. The investigators concluded that the interactive module was associated with greater early gains in self-rated understanding and proposed that similar modules might complement catheterization-laboratory observation in supporting preparation and consolidation. The evidence is limited to the study’s retrospective, non-randomized design and its use of self-reported understanding. It does not establish that the module improved objectively tested knowledge, procedural competence or patient outcomes, and it does not show that the two approaches differed after both had been completed.

Subject of Research: Effects of interactive 360 video and catheterization-laboratory observation on medical students’ self-rated cardiology understanding

Article Title: Interactive 360 video for catheterization-laboratory education in undergraduate cardiology clerkship: a retrospective two-period crossover study

Article References: Tanimoto, T., Takahata, M., Hirano, T., Kashiwagi, M., Sasaki, Y., Mori, M., Nakatsugawa, Y., Mizumoto, K., Tanaka, A., & Murata, K.-Y. (2026). Interactive 360 video for catheterization-laboratory education in undergraduate cardiology clerkship: a retrospective two-period crossover study. BMC Medical Education. https://doi.org/10.1186/s12909-026-10396-3

Image Credits: AI Generated

DOI: 10.1186/s12909-026-10396-3

Keywords: medical education, 360-degree video, cardiology clerkship, catheterization laboratory, fifth-year medical students, self-rated understanding, retrospective crossover study, immersive learning, Interactive, video, catheterization-laboratory, education

The researchers developed the multimedia through a research and development process using the ADDIE model: Analysis, Design, Development, Implementation and Evaluation. Needs analysis included preliminary observations and semi-structured interviews with students and lecturers. Reported problems included low motivation, difficulty understanding abstract grammatical structures, limited vocabulary mastery and weak engagement with conventional learning media. The needs analysis also found that 75% of students were dissatisfied with the instructional media then in use.

CTL provided the product’s pedagogical foundation. Its stated principles include constructivism, inquiry, modeling, learning community, reflection and authentic assessment, with instructional content connected to meaningful contexts. Previous studies cited by the researchers associated CTL with participation, communication skills and motivation, but most prior implementations were confined to conventional classrooms or low-technology materials. The new project sought to integrate those principles with a digital, interactive environment specifically designed for university-level Arabic instruction.

Kotobee, an interactive authoring platform developed by the Cairo-based company Vijua, was used to embed contextual reading texts, pronunciation audio, videos, vocabulary exercises, annotations, navigation tools and interactive quizzes. Exercises included multiple-choice, true-or-false, drag-and-drop and short-answer formats. Students could submit answers, view scores and incorrect responses, and clear answers to rework questions. Campus-related scenarios, including self-introduction and daily communication, were incorporated to place Arabic in situations relevant to students’ experience.

Because random assignment was not feasible in the institution’s regular Arabic program, the researchers used a non-equivalent control-group design. Both groups were taught by the same instructor, and learning objectives and content were standardized. The intervention lasted six weeks, with two instructional sessions of approximately 200 minutes each per week. The experimental group used the CTL-based Kotobee multimedia, while the control group received conventional instruction through printed materials, lectures and translation exercises. Classroom observations were used to monitor adherence to the planned procedures.

The main quantitative instrument was a 30-item learning-interest questionnaire covering attention, participation and willingness to engage in Arabic learning. Items used a five-point Likert scale and were reviewed by three specialists in Arabic education and educational technology. Retained item correlations ranged from 0.412 to 0.789, and the questionnaire had a Cronbach’s alpha of 0.87. Pretest scores did not differ significantly between the groups, with an independent-samples t-test yielding t(48) = 0.35 and p = 0.725.

Posttest scores were not normally distributed, so the researchers used a Mann–Whitney U test. The groups differed significantly after the intervention, with U = 95.0 and p < 0.001. The reported rank-biserial correlation was 0.70, described as a large effect, while the Hodges–Lehmann 95% confidence interval for the median difference was 8.5 to 14.2 points. N-Gain analysis placed the experimental group’s increase in the effective category and the control group’s in the ineffective category; the abstract reports respective increases of 76.64% and 41.26% in learning interest.

Experts rated the multimedia as “very feasible.” Two media experts evaluated display quality, interactivity and practicality, while one material expert assessed content and exercise quality using five-point rubrics converted to percentages. All four dimensions received ratings above 90%. The findings support the product’s technical usability and instructional suitability, but this was an expert assessment rather than a formal evaluation of student usability. The authors say informal classroom feedback, rather than a standardized end-user instrument, informed subsequent product refinement.

The results are preliminary because the study occurred at one institution, used a convenience sample of 50 students from two classes, and lasted six weeks. The design cannot establish whether the observed effect was caused solely by the digital format, the CTL design, their combination, or other classroom differences, and the short period did not allow assessment of long-term retention or proficiency development. Future studies would need to examine other institutions, learner populations, language skills and platforms, as well as use standardized student usability measures. Within those limits, the study provides evidence that a pedagogically integrated Kotobee product was associated with higher Arabic-learning interest among the participating Indonesian university students.

Subject of Research: A quasi-experimental evaluation of CTL-based Kotobee multimedia for Arabic-language learning among 50 Indonesian university students.

Article Title: Interactive multimedia for Arabic language learning based on contextual teaching and learning using Kotobee

Article References: Shodiq, M. J., & Siregar, A. F. (2026). Interactive multimedia for Arabic language learning based on contextual teaching and learning using Kotobee. Discover Education, 5(1), Article 928. https://doi.org/10.1007/s44217-026-02073-w

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02073-w

Keywords: Arabic language learning, interactive multimedia, Kotobee, Contextual Teaching and Learning, learning interest, higher education, Indonesia, quasi-experiment, Interactive, multimedia, Arabic, language

Relative humidity had an approximately U-shaped association with incidence. The pollutant exposure–response curves also differed from one another: sulfur dioxide and nitrogen dioxide had S-shaped relationships, while carbon monoxide showed a continuously rising association with incidence. All five environmental factors had statistically significant lagged effects within particular lag windows, indicating that their modeled relationships with reported cases were not confined to the same day.

Seven meteorology–pollutant pairs were evaluated for interaction. Only the mean-temperature–carbon-monoxide combination showed statistically significant synergy. Its relative excess risk due to interaction was 0.237, with a 95% confidence interval of 0.032 to 0.477, while its attributable proportion due to interaction was 0.232, with a 95% confidence interval of 0.039 to 0.39.

The other five tested pairs did not have statistically significant interactions. These model estimates indicate a joint association beyond the additive combination of the two factors in the analyzed data; they do not establish that temperature or carbon monoxide caused mumps infections or that either factor directly increased an individual’s risk.

Subject of Research: Environmental associations with reported mumps incidence in Chongqing, 2015–2024

Article Title: Interactive effects of meteorological factors and air pollutants on mumps incidence in Southwest China, 2015–2024

Article References: Yang, C., Yang, L., Wang, N., Yang, S., Tang, X., & Qi, L. (2026). Interactive effects of meteorological factors and air pollutants on mumps incidence in Southwest China, 2015–2024. BMC Infectious Diseases. https://doi.org/10.1186/s12879-026-14314-2

Image Credits: AI Generated

DOI: 10.1186/s12879-026-14314-2

Keywords: mumps, Chongqing, air pollution, carbon monoxide, temperature, relative humidity, infectious disease, environmental epidemiology, time-series analysis, Interactive, effects, meteorological

The pedagogical insight behind the platform is deceptively simple. In real classrooms, teachers rarely want students to produce just any mathematically equivalent expression — they want a specific canonical form at each stage. When solving an equation of the form ax + b = c, a teacher typically expects Step 1 to become ax = c − b and only then Step 2 to become x = (c − b)/a. A tool that accepts broad algebraic equivalence may happily accept a shortcut that skips the intended transformation, quietly undermining the scaffold the teacher built. The same tension appears when solving matrix-form linear systems AX = B, where students are guided through isolating one variable, computing it, substituting it into the remaining equation, and solving for the second. The Rabat platform therefore enforces teacher-specified canonical forms rather than merely checking equivalence — a deliberate narrowing that most tutoring systems have avoided.

Technically, the platform is built around a strict separation of concerns. The symbolic formula a student types is verified by a compact machine-learning classifier, while the numeric value is checked deterministically by the server with a tolerance. This split is pedagogically meaningful because it distinguishes symbolic misconceptions from arithmetic slips — a sign error is a different kind of mistake from a division slip, and the feedback messages reflect that. The classifier itself is a hybrid: the student’s formula string is represented using character n-gram TF–IDF vectors, which preserve micro-patterns like operator placement, sign structure, and denominator layout without requiring a full symbolic parser. These textual features are concatenated with targeted numeric cues — the step index, whether a division symbol is present, whether the left-hand side matches the expected form, whether numerator and denominator appear inverted. The combined matrix feeds into a compact multilayer perceptron with two hidden layers, trained with class weights to handle imbalance and early stopping to select the best checkpoint.

The performance numbers are startling for such a lightweight architecture. On the single-variable equation task, the classifier reaches 99.62 percent accuracy and 0.9962 macro-F1 on held-out test data, averaged over five random seeds. On the more structured AX = B task — which spans sixteen step-specific categories including four correct forms and twelve realistic distractors — the corresponding classifier reaches 99.87 percent accuracy and 0.9987 macro-F1. The baselines make the design choices visible: replacing character n-grams with word unigrams still yields a strong baseline around 96.5 percent accuracy, but removing the targeted numeric cues causes a dramatic drop to roughly 86.8 percent. In other words, the numeric cues are not cosmetic; they carry the signal that separates surface-similar formulas like ax = c − b from ax = b − c, where the difference is a single operator and the pedagogical difference is a sign misconception.

The error taxonomy is where the platform becomes genuinely teacher-facing. For the equation task, the labels include sign errors on b, swapping b and c, missing b on the right-hand side, denominator inversion, and missing division by a. For the matrix task, the taxonomy covers confusions between x₁ and x₂, using the wrong right-hand side, missing coefficients, numerator-denominator inversions, missing divisions, sign errors on substituted terms, and divisions by the wrong coefficient. These categories were chosen not to exhaust every mathematically possible wrong form but to cover the most frequent and pedagogically relevant near-miss patterns under the scaffold. Because the labels are stable and interpretable, the downstream analytics become directly actionable: a teacher can see not just that students are struggling but which specific symbolic operation is generating the friction.

The interface design is where the inclusive framing becomes concrete rather than rhetorical. All feedback is text-based and immediately visible on screen — no reliance on audio cues, which matters for deaf and hard-of-hearing learners. Instructions and the expected target form remain persistently visible at each step. Progression is step-locked: the next step is not revealed until the current one is completed correctly, which prevents scaffold-breaking shortcuts and reduces cognitive overload. Students can attempt each step multiple times without penalty, and the feedback messages are non-judgmental and specific rather than punitive. On the server side, every verified step is logged under a pseudonymous student code, capturing attempts, timestamps, step durations, and structured error labels without storing personally identifiable information. The participation of deaf and hard-of-hearing students took place within an established cooperation framework between the Faculty of Sciences at Mohammed V University and the Fondation Lalla Asmaa, which supports higher-education training of deaf students within the faculty.

The classroom logs reveal patterns that synthetic evaluation alone cannot. For the equation task, Step 1 consistently consumes more time than Step 2, confirming that the transformation ax = c − b is the main symbolic bottleneck rather than the final division. Real classroom mistakes concentrate overwhelmingly at Step 1, especially sign errors on b and the swap between b and c — exactly the operations that deserve targeted remediation. For the matrix task, the logs show a striking separation between cohorts: the hearing-impaired cohort had a median completion time of 808 seconds over 124 solved items, while the other cohort had a median of 278.4 seconds over 123 solved items. The authors are careful to note that this difference should not be interpreted as a causal effect of the system; rather, it demonstrates that the platform captures meaningful process variation in authentic use and can therefore support differentiated teacher monitoring. Learning curves across successive exercises suggest pacing stabilization over repeated practice while maintaining visible separation between cohorts.

The authors frame their educational claims with deliberate caution. The classroom evidence is presented as exploratory rather than causal — the study included no control groups, no pre/post testing, and no formal learning-gain estimation. This is an important distinction in a field where technology interventions are often oversold. What the results do establish is that a locally deployable, CPU-only platform can enforce canonical intermediate forms, distinguish formula errors from numeric errors, provide immediate step-level feedback, and produce teacher-facing analytics that remain meaningful in authentic classroom use. The practical value lies not in deciding whether a student is correct but in surfacing which step and which symbolic transformation deserves instructional attention — a form of closing the loop in learning analytics, where data become educationally useful only when they support timely intervention.

The positioning against prior work clarifies the niche. Compared with constraint-based tutors coupled to computer algebra systems, which can provide strong pedagogical control but carry substantial authoring and infrastructure burden, the Rabat platform reduces deployment cost by focusing on short symbolic inputs under fixed scaffolds. Compared with open-ended conversational models, which can solve equations but are not optimized for reproducible step-locked verification with stable error labels, the platform offers a different kind of value: operational reliability on ordinary hardware. Compared with exploratory misconception mining, which discovers latent error structure after the fact, the platform adds a real-time supervised component that recognizes errors during interaction and attaches explicit labels to them. The contribution is therefore not a single-task demonstration but a platform-level architecture that can host multiple algebra tasks under shared interaction philosophy, logging strategy, and lightweight deployment model.

The limitations define the scope honestly. Both tasks rely heavily on synthetic supervision for the formula classifiers, and the extent to which the current taxonomies cover all naturally occurring student variants remains an open empirical question. The classroom study is context-specific and moderate in size. The AX = B task is deliberately restricted to a controlled upper-triangular matrix setting rather than full general linear-system solving — a design choice that preserves canonicality and classroom tractability but limits generalization. The platform emphasizes keyboard-based textual interaction and does not yet address handwriting, speech, or richer multimodal accessibility pathways. Future work, the authors suggest, should include stronger evaluation designs such as pre/post measures or quasi-experimental protocols, expanded taxonomies trained on larger collections of authentic student inputs, and richer accessibility support while preserving local deployability. What the platform already demonstrates, though, is a coherent and extensible foundation: a step-locked, CPU-only framework that combines reliable symbolic diagnosis with interpretable teacher-facing analytics across more than one algebra task in inclusive, low-stakes educational settings.

Subject of Research: A lightweight interactive platform for stepwise formative assessment in algebra from single variable equations to matrix form linear systems in inclusive classrooms

Article Title: A lightweight interactive platform for stepwise formative assessment in algebra from single variable equations to matrix form linear systems in inclusive classrooms

Article References: Bendourou, M. A., Riadsolh, A., Massaoudi Souchak, S., & Ez-Zahraouy, H. (2026). A lightweight interactive platform for stepwise formative assessment in algebra from single variable equations to matrix form linear systems in inclusive classrooms. Discover Education, 5(1), Article 836. https://doi.org/10.1007/s44217-026-02085-6

Image Credits: AI Generated

DOI: 10.1007/s44217-026-02085-6

Keywords: lightweight, interactive, platform, stepwise, formative, assessment, algebra, single, variable, equations, matrix, form

The ontological argument offered by Song and Evans deserves particular attention because it provides a vocabulary for a problem that is often discussed only in practical terms. Heidegger’s distinction between objects that merely stand at our disposal and things that gather a world of relations around themselves maps onto a genuine dilemma in virtual museology. A high-resolution 3D scan of a vessel, a manuscript, or an architectural fragment can be reproduced endlessly, transmitted instantly, and rendered in any environment a developer chooses, yet none of these operations automatically preserve the web of provenance, ritual use, historical circumstance, and scholarly interpretation that gives the original its significance. The claim that artefacts acquire meaning through gathered relations rather than intrinsic physical properties implies that the digital replica is not inherently impoverished; it is impoverished only if the relations are not carried across. This reframes the curator’s task: the question is no longer whether a digital surrogate can substitute for the original, but whether the curatorial apparatus that assembles meaning around the object can be extended into the virtual environment without being dissolved by it.

The empirical findings from Furusawa and colleagues on embodiment and aesthetic response add an important corrective to a common assumption in personalisation research. Much of the literature on adaptive museum systems treats the visitor as a bundle of preferences to be inferred: prior knowledge, dwell time, interaction history, stated interests. The finding that aesthetic and emotional response varies substantially with embodied locomotion and the physical scale of artworks, independently of where visitors direct their gaze, suggests that preference-based models capture only part of what shapes the experience. Two visitors may attend to precisely the same visual features of a painting yet undergo markedly different affective experiences because one approached it slowly across a large gallery while the other encountered it as a small reproduction during rapid movement. For AI-driven personalisation this implies that the sensor suite and the underlying model must extend beyond attention tracking into the domain of the body: gait, posture, proximity, and the kinesthetics of movement through space. It also implies that the design of the virtual environment itself, including room dimensions and the scale at which artworks are presented, is not a neutral container for personalisation but an active variable in the emotional outcome the system is trying to modulate.

The Museu XYZ case study raises questions that extend well beyond blockchain technology into the political economy of virtual museums. Decentralised curatorial models, in which authority is distributed among artists and communities rather than vested in a single institution, align naturally with the accessibility promises of the Metaverse: anyone with a connection can visit, contribute, and in principle shape the collection. Yet the dependence on cryptoart and NFT market logics means that the economic infrastructure of such museums is tied to speculative markets whose volatility and environmental costs have been extensively debated. A community-curated virtual museum whose exhibits are minted as tokens inherits the incentive structures of those tokens, including the possibility that curatorial decisions become entangled with asset valuation. This does not invalidate the model, but it illustrates the editorial’s central worry: commercial and infrastructural interests can encroach on curatorial judgement not through a dramatic takeover but through the quiet adoption of platforms and market mechanisms whose logics were designed for other purposes.

It is worth situating the AI Director concept of Çakmak and Çekmiş within the longer history of adaptive systems in cultural settings. Early adaptive hypermedia research in the 1990s explored personalised tour paths through digital collections, and subsequent decades produced increasingly sophisticated visitor models, from stereotype-based user profiles to machine-learned predictions of interest. What distinguishes the current generation is the coupling of these models with real-time control of the experience itself: not merely recommending which artefact to view next, but adjusting narrative pacing, feedback timing, and interaction modality on the fly. The finding that this works best when artefact identity and narrative structure are fixed by design is a genuinely useful design principle. It suggests that the appropriate division of labour between human and machine is not a question of how much authority to delegate in the abstract, but of identifying which layers of the experience are load-bearing for meaning and must therefore remain invariant, and which layers can safely vary in response to inferred visitor states. Personalisation, on this view, is not the opposite of curatorial intent but operates within a space that curatorial intent defines.

The question of affective computing in museums also merits careful treatment. Systems that infer emotion from facial expression, voice, posture, or physiological signals promise a form of responsiveness that no human docent could scale to thousands of visitors, and there is genuine evidence that emotionally adaptive experiences can increase engagement and recall. But the inference of inner states from external signals remains scientifically contested, with documented biases across cultures, ages, and individual differences in expressiveness. A system that misreads a visitor’s contemplative stillness as boredom, or their culturally specific expression of reverence as confusion, may adapt the experience in ways that distort rather than support the encounter with heritage. The ethical stakes are heightened in cultural settings, where the emotional register of an encounter with sacred objects, traumatic histories, or ancestral artefacts is not incidental but constitutive of their meaning. Institutions deploying such systems will need not only technical validation but also transparency about what is being sensed, what is being inferred, and what the visitor can control.

The editorial’s call for comparative studies across curatorial models deserves elaboration, because the field’s evidence base remains thin in characteristic ways. Most published evaluations of XR heritage experiences are single-site, short-duration studies with small samples, often conducted with visitors who are already technologically enthusiastic. Non-adaptive baselines are frequently absent, making it difficult to attribute observed benefits to the AI component rather than to novelty effects or the general appeal of immersive media. Longitudinal evaluation across repeated visits is especially rare, yet the value of a museum experience may lie less in a single memorable session than in whether it motivates return visits, sustained learning, or changed attitudes toward heritage. Comparative designs that hold content constant while varying the curatorial model, from centralised institutional control to decentralised community authorship, would begin to answer whether the ontological and political questions raised here have measurable experiential consequences.

The regulatory and economic dimension flagged in the editorial is likely to grow in importance as AI-driven curation matures. Cultural institutions increasingly depend on commercial platforms for the delivery of immersive experiences, and these platforms accumulate data about visitor behaviour that has value far beyond the museum’s own evaluative purposes. Questions of data governance, algorithmic transparency, and the terms under which cultural institutions delegate interpretive authority to third-party systems are not peripheral technicalities; they bear directly on the editorial’s question of whether AI assumes responsibility for curatorial intent in ways that serve the public or serve the infrastructure. Professional bodies in museology and digital heritage are only beginning to develop guidance for these arrangements, and the Research Topic’s insistence on sustained critical attention to economic structures provides a timely prompt for that work.

Finally, the vision of digital artefacts as first-class objects of curatorial, educational, and social interaction represents a significant reorientation for the field. Under the older surrogate model, the digital copy was judged by its fidelity to the original and its usefulness in lieu of a visit. Under the emerging model, a born-digital artwork, a procedurally reconstructed lost building, or a community-authored virtual exhibit may have no original at all, and its significance is established through the relations it gathers: the narratives it supports, the communities that curate it, the interactions it affords. This does not diminish the value of physical collections, but it does mean that the discipline’s theories, evaluation methods, and professional practices, developed largely around the primacy of the original object, will need continued adaptation. The contributions gathered in this Research Topic, spanning ontological theory, constrained implementation, decentralised experimentation, and embodied empirical study, offer a coherent starting point for that adaptation, and the questions they converge upon, about where meaning resides and who has authority over it, will remain live as the technologies mature.

Subject of Research: Editorial: AI-driven XR applications in cultural heritage: a new era of interactive museums

Article Title: Editorial: AI-driven XR applications in cultural heritage: a new era of interactive museums

Article References: Nofal, E., Hammady, R., & De Marco, R. (2026). Editorial: AI-driven XR applications in cultural heritage: a new era of interactive museums. Frontiers in Virtual Reality, 7, Article 1990419. https://doi.org/10.3389/frvir.2026.1990419

Image Credits: AI Generated

DOI: 10.3389/frvir.2026.1990419

Keywords: Editorial, AI-driven, applications, cultural, heritage, interactive, museums, scientific research

The core difficulty is that planning quality depends on translating noisy camera data into a reliable sequence of actions. The system derives each object’s position from live depth imagery, converts those detections into structured scene text, and then asks a large language model to produce an ordered removal plan. Every plan must satisfy three conditions: it must be complete, it must respect physical constraints such as removing obstructing parts before the parts they block, and it must never reference objects that do not actually exist in the scene.

To keep the model honest, the researchers layered multiple safeguards around it. A generator model produces a candidate plan under a strict list-only output rule that forbids explanations, headings, or bullets. A separate verifier model then re-checks the candidate against the same scene representation, validating format, duplicates, and ordering. Finally, a deterministic grounding filter strips out any line whose label and coordinates cannot be matched to a detected object, guaranteeing that hallucinated parts never reach the robot.

The upgraded version replaces the earlier ensemble approach, which sampled multiple candidate plans and required six generator calls plus one verifier call, with a single generator call and, when needed, one verifier call. The model was also given a more structured reasoning prompt and a clearer set of matching rules for interpreting user requests, including how to handle ambiguous references and when to declare that no valid target exists.

On the offline benchmark, the single-call generator already achieved strong results, and adding the verifier and grounding filter pushed full-plan accuracy to 96.5 percent, up from 82.4 percent for the earlier ensemble pipeline with the same replacement model. The researchers stress that these figures come from a constrained benchmark and may overstate performance on unseen battery layouts or unfamiliar operator phrasing.

The team also built a fully integrated web interface that streams live camera views, clickable object detections, and every intermediate planning stage to the operator. Users can inspect the structured input, the generator output, the verifier output, and the final filtered plan, then accept or reject the result. In a pilot study with twelve participants, the new interface cut the average time to accept a plan from 22 seconds to 15 seconds on the first task and from 15 seconds to 11 seconds on the second, while retrospective workload scores dropped from 24.1 to 14.1 on a 100-point scale. Every participant preferred the new interface.

The researchers caution that the pilot was small and exploratory, that planning quality is tied to the accuracy of the underlying perception, and that the deterministic filter can reject valid plans whose formatting or coordinates deviate from expectations. They also note that the results are tied to one specific model and do not establish model-agnostic performance. Future work will focus on broader benchmarks, richer scene representations, and testing across multiple language models under identical prompts and scoring rules.

The benchmark underlying these findings deserves closer attention, because it explains why the accuracy numbers can be compared cleanly across pipeline versions. The evaluation set consists of 200 distinct scenes, each paired with three operator intents, yielding 600 planning episodes in total. Because every version of the system was run against the same episodes, the reported gains reflect changes to the planning architecture rather than differences in task difficulty. The metric, mean full-list correctness, is deliberately strict: a plan counts as correct only if every removal step is present, properly ordered, and grounded in the scene, so partial credit is not awarded for plans that are merely plausible.

Viewed through that lens, the progression of scores traces a clear engineering narrative. The earlier six-sample ensemble pipeline, which sampled multiple candidate plans from a single model checkpoint and reconciled them through verification, reached a mean full-list correctness of 0.824. Simply swapping in the replacement language model while keeping the ensemble machinery intact raised the figure to 0.896, indicating that a substantial share of the improvement came from the stronger instruction-following behavior of the newer model rather than from the ensemble structure itself. The upgraded configuration, which discarded parallel sampling in favor of a single generator call with revised prompts, reached 0.935 on its own, and the addition of the verifier stage and the deterministic grounding filter lifted the final figure to 0.959.

That the upgraded generator alone outperformed the full prior pipeline is notable, because it suggests that prompt design and output contracts can matter as much as model choice. The revised generator prompt emphasizes strict list formatting, explicit handling of ambiguous references, and clear rules for declaring when no valid target exists, which reduces the formatting drift that previously had to be repaired downstream. The verifier then operates on a shared formal grammar with the generator, so both stages agree on what constitutes a well-formed plan, and the verifier can correct ordering errors or duplicates before the deterministic filter performs its final token-level check against the perceived scene payload.

The reduction in language model calls from seven to two carries practical consequences beyond accuracy. Each call adds latency, compute cost, and another opportunity for the model to introduce inconsistencies, so halving the call count while improving correctness represents a genuine efficiency gain rather than a trade-off. For deployment settings such as battery recycling facilities, where operators wait on the planner between actions, fewer calls translate directly into shorter planning cycles and lower per-plan operating expense.

The human factors results, though drawn from a small within-subject pilot, align with a broader principle in human-robot interaction research: operators trust and use automated planning tools more effectively when they can see what the system is reasoning about. By streaming the exact text sent to each model stage and displaying raw, cleaned, verified, and filtered plans side by side, the interface turns the planner from a black box into an inspectable artifact. The consistent preference of all twelve participants, together with lower raw NASA-TLX workload scores and faster acceptance times on both tasks, suggests that transparency did not come at the cost of added cognitive burden, which is often a concern when interfaces expose internal system states.

The LUZD coordinate representation also plays a quiet but important role in the pipeline. Depth cameras produce point clouds in a sensor frame that can shift as the camera moves, so plans expressed in raw sensor coordinates are fragile. By rewriting all detections into a frame-consistent representation, the system ensures that the labels and coordinates the language model reasons over match what the deterministic filter later checks, and that accepted plans can be executed by the robot without an additional coordinate reconciliation step.

It is worth situating the anti-hallucination filter within the wider literature on hallucination mitigation. Sampling-based consistency checks such as SelfCheckGPT estimate reliability by comparing multiple generated outputs, which is effective but expensive. In robotic settings, where every emitted action may be physically executed, a deterministic post hoc filter that enforces exact input-output consistency is attractive precisely because it is cheap, predictable, and auditable. The trade-off, acknowledged by the authors, is that such filters can reject valid plans whose formatting or coordinates deviate from expectations, so filter strictness must be balanced against unnecessary plan rejection.

The application domain itself shapes many of these design choices. Electric-vehicle batteries arrive at end-of-life in highly variable states: pack designs differ across manufacturers, fasteners and modules may be corroded or deformed, and prior partial disassembly can leave the scene in unexpected configurations. Technoeconomic analyses of battery recycling consistently identify this uncertainty, together with non-standardized designs, as a barrier to full automation, which is why human-in-the-loop architectures remain prominent in the field. A planner that can accept natural-language corrections from an operator, while guaranteeing through deterministic grounding that no phantom component enters the action list, fits that operational reality better than either fully manual teardown or fully autonomous scheduling.

Finally, the authors are careful to bound their claims. The benchmark scenes are constrained, the pilot cohort is small, and the results are tied to one specific model checkpoint, so the findings should be read as evidence that structured scene text, clear prompts, verifier-based correction, deterministic grounding, and inspectable interfaces form a productive combination, rather than as proof that any language model would perform equivalently. The stated agenda of testing multiple models under identical prompts and scoring rules, and of extending the scene representation beyond current limits, points toward the kind of controlled comparisons that will determine whether these gains generalize across the messy variety of real disassembly floors.

Subject of Research: Robotics and AI

Article Title: Interactive and inspectable LLM planning for human-in-the-loop EV battery disassembly

Article References: Pedretti, M., Contreras, C. A., Stolkin, R., Rastegarpanah, A., & Pagani, R. (2026). Interactive and inspectable LLM planning for human-in-the-loop EV battery disassembly. Frontiers in Robotics and AI, 13, Article 1916454. https://doi.org/10.3389/frobt.2026.1916454

Image Credits: AI Generated

DOI: 10.3389/frobt.2026.1916454

Keywords: robotics, AI, EV batteries, disassembly planning, human-robot interaction, large language models, Frontiers in Robotics and AI, Interactive, inspectable, planning, human-in-the-loop, battery

The study, published in the Journal of New Approaches in Educational Research, was led by Aida Moreno-Rus and Mercedes Ventura as co-first authors, together with Noelia Ventura-Campos, Dana G. Stefanescu and Zoe Falomir. Fifty-two university students aged 18 to 26, drawn from the sciences, health sciences, social sciences and humanities, were initially screened with two computerized detection tasks. The first, a time-limited task, gave participants just three seconds to judge whether an equation correctly matched a problem statement, a constraint designed to expose automatic, unreflective mistranslations. The second allowed unlimited time for students to construct their own equations using an on-screen interface. Only students who persisted in making reversal errors, either under time pressure alone or in both conditions, were selected for the intervention.

Reversal errors are deceptively simple. Cognitive theories dating to Clement’s classic 1982 work propose two main mechanisms. The word-order matching account holds that students translate text linearly, mimicking the sequence of words rather than modeling the underlying relation: “six times as many students as professors” becomes 6·S = P. The static comparison account holds that students form a correct mental image, one professor for every six students, but then encode that image as a symbolic shortcut such as 1·P = 6·S, which simplifies to the same reversed equation. Later scholarship, including work by Jankvist and Niss, has argued that the idiom “times as many as” itself poses a combined syntactic and semantic obstacle, particularly in languages such as Spanish where the phrase “veces más que” can blur the boundary between multiplication and addition.

Crucially, the intervention targeted not just reversal errors but the full family of translation mistakes observed in the screening tasks. Operator errors arose when students confused comparative idioms, writing addition where multiplication was required or vice versa. Structural errors appeared when students placed a number as a divisor or subtrahend of the variable itself, as in 9/P = S instead of P/9 = S, reflecting a shallow understanding of what a variable represents. Hybrid errors combined reversal and operator confusions in a single incorrect equation. Before the intervention, the researchers found that half of the participating students reported translating problem statements literally, following the order of the words without reorganizing the mathematical relationships, and a quarter admitted reading the statements hastily without identifying which quantity was being compared to which.

The gamified tool, built in Unity3D and running on ordinary Android tablets, deliberately avoided the narrative structure and fictional worlds of commercial video games and serious games. Instead, it offered a guided learning experience organized as a path across an island, with eight levels of ten exercises each. Each exercise unfolded through interactive phases: reading and understanding a comparative word problem, formulating an equation with a calculator-style interface, and then receiving immediate feedback. When a reversal error was detected, the tool first applied an arithmetic validation strategy, substituting concrete numerical values into the student’s equation so that the failed equality became visible. If the error persisted, a second corrective layer engaged: a visual representation using containers of different sizes as physical metaphors for the relative magnitudes of the two quantities, prompting the student to decide which variable needed to be multiplied, divided, added to or subtracted from so that both containers reached the same size.

From level four onward, the scaffolding deliberately faded. The tool stopped displaying the relative container sizes explicitly and instead asked participants to determine for themselves which quantity should be larger, choose the appropriate operator and decide which variable it should apply to. The designers also introduced a coin-based reward system at this point, granting one virtual coin per correct answer and up to fifty coins in total, a motivational layer added after pilot studies showed concentration flagging around level three. An inverse task, in which students constructed a word problem from an equation they had written, drew on Piagetian ideas about reversible thinking to deepen structural understanding. Explanatory messages addressed the semantics of phrases like “times more than” and “times less than,” directly targeting the operator confusions documented in prior research.

The quantitative results were striking. After roughly three weeks of training in two to three supervised sessions per week, correct responses increased significantly in both the timed task, with a large effect size of r = .87, and the untimed task, with r = .86. Reversal errors fell by 87 percent, operator errors dropped by 87.5 percent, and hybrid errors vanished entirely. Half of the intervention participants earned the maximum fifty coins, and the overwhelming majority exceeded the forty-five-coin threshold tied to a small financial incentive. Wilcoxon signed-rank tests confirmed that these gains were statistically robust at p < .001 across both assessment formats.

Equally notable were the affective findings. Drawing on flow theory, the researchers measured enjoyment, relaxation and concentration at the end of each level on six-point Likert scales, and administered a five-item retrospective flow questionnaire after the final level. Enjoyment remained stable and high throughout, while relaxation and concentration rose significantly as students progressed, particularly after the coin rewards were introduced. A principal component analysis of the three emotional measures yielded a single factor explaining 83 percent of the variance, interpreted as a unified dimension of positive emotional engagement. The average perceived flow score reached 4.57 out of 6, and flow correlated strongly with sustained emotional engagement, r = 0.661. Notably, students who had made reversal errors only under time pressure and those who made them consistently reported equally high engagement, suggesting the tool works across ability profiles.

Qualitative responses collected after the training illuminated how the learning happened. Forty-four percent of participants said they had come to understand that an equation expresses an equality between two expressions, with one student explaining that they now understood equations “in terms of quantities, not in the order of the sentence.” Others reported improved interpretation of comparative idioms, strategic self-correction through repeated feedback, and growing fluency and confidence. Many described the tool as having a metacognitive function, acting as a catalyst for recognizing their own error patterns and adjusting strategies accordingly, a hallmark of self-regulated learning. When asked about pedagogical value, nearly a third of participants said the tool was directly applicable in classrooms, particularly for introducing equations and equalities to younger learners.

The authors are careful about the limits of the work. The sample was small, purposively selected and non-probabilistic; the emotional data relied on self-report; and the design cannot fully disentangle the contribution of the gamified motivational layer from the instructional core. Future validation in authentic secondary school classrooms, with larger and more diverse samples, remains necessary, as does investigation of whether the cognitive and emotional gains persist over time. Nevertheless, the study represents one of the few instructional interventions ever shown to substantially reduce a misconception that has resisted traditional abstraction-based teaching since it was first documented in 1979. The message for educators is twofold: misconceptions rooted in the translation of everyday language into algebra are remediable with the right combination of visual grounding, immediate feedback and faded scaffolding, and the emotional states that sustain learning, enjoyment, calm and focused attention, can be designed for rather than left to chance.

Subject of Research: Reducing reversal errors in comparative algebra word problem solving through a gamified interactive educational tool

Article Title: A gamified interactive educational tool to support algebraic thinking: reducing reversal errors in comparative word problem solving

Article References: Moreno-Rus, A., Ventura, M., Ventura-Campos, N., Stefanescu, D. G., & Falomir, Z. (2026). A gamified interactive educational tool to support algebraic thinking: reducing reversal errors in comparative word problem solving. Journal of New Approaches in Educational Research, 15(1), Article 17. https://doi.org/10.1007/s44322-026-00066-z

Image Credits: AI Generated

DOI: 10.1007/s44322-026-00066-z

Keywords: reversal error, algebraic thinking, comparative word problems, gamification, educational technology, flow theory, mathematics education, equation translation, emotional engagement, interactive learning environment, gamified, interactive