Sustaining children’s attention in complex educational digital environments remains a major challenge. We present an adaptive video-learning system for children aged 6–10 that couples two functions: estimating each child’s attentional state from gaze behaviour, and generating new story content that is adapted to that state. While a child watches an animated story about ocean life (the learning task), the system monitors gaze on predefined areas of interest; once the gaze dwells on a story anchor for a validated threshold of 2.37 s, a generative artificial-intelligence pipeline composes a new 30-second episode that continues the story from the attended element, so the pace, focus, and narrative of the video change as a direct function of the child’s gaze. Facial expressions recorded during viewing are analysed frame by frame with the open-source Py-Feat facial-expression analysis toolbox to quantify emotional engagement. In a within-subjects experiment (N = 60), the gaze-adaptive condition was compared with traditional linear video viewing on five measures; a separate non-randomised pilot cohort (N = 24) viewing a conventional live-action documentary served as a descriptive baseline. Expert ratings of attentiveness showed high inter-rater reliability (intraclass correlation = .945). The gaze-adaptive condition outperformed traditional viewing on all five measures—fill-in-the-blank comprehension, story retelling, self-reported interest, expert-rated attentiveness, and gaze-based focus duration (all p<.001 ; Cohen’s d_z = 0.46-−1.00)—and gains relative to the documentary baseline remained large. Gaze-driven generation thus shifts video watching from passive reception toward active, attention-guided interaction and offers a methodological framework for attention-aware adaptive media for children.
Prior research suggests that the attitude of virtual humans, in situations such as job interviews, can influence users’ stress levels. However, the mechanisms through which designers can precisely modulate stress levels in these settings remain unclear. We propose a mechanism for modulating users’ stress levels through multimodal social interactions with virtual humans. These interactions are modulated along two dimensions: friendliness (friendly vs. hostile) and dominance (dominant vs. non-dominant). To test this mechanism, we conducted two studies in which participants interacted with virtual humans that varied in friendliness and dominance. Friendliness was expressed through cues of appreciation or dislike, such as facial expressions (e.g., smiling or frowning) and speech content (e.g., “that’s good!”), whereas dominance was expressed through cues of control or lack of control, such as social rank (i.e., supervisor vs. subordinate) and assertiveness (i.e., giving instructions vs. asking permission). The results indicate that the behavior of the virtual humans plays a role in shaping stress responses; however, the findings reveal a complex relationship and therefore need to be interpreted with caution. In particular, the findings showed that the dominant virtual human was perceived as less friendly than the non-dominant virtual human. Future research should further investigate this effect and consider alternative models that provide more clearly distinguishable and independent control parameters for modulating stress responses in social virtual environments.
Enhancing the user interaction experience in computer-aided design (CAD) modeling through natural and intuitive interfaces has become an active research area. Eye tracking, as a novel physiological signal-based interaction mode, presents notable advantages for CAD model selection, which has a dominant influence on the modeling experience. Current eye tracking-based CAD model selection methods are largely limited by the challenge of Midas touch problem. Existing solutions, such as multimodal gaze-gesture interfaces, often introduce excessive cognitive load and operational complexity, while dwell time-based methods compromise interaction efficiency. Consequently, there is a critical need for a natural, intuitive gaze-only solution capable of inferring user selection intentions without relying on auxiliary modalities or predefined fixation rules. In this study, we attempt to discriminate the user’s selection intentions by investigating the correlation between eye movement features and model selection states. To this aim, we conducted statistical analysis to eye movement data collected under different model selection intentions in a behavioral experiment and examined the relevant features for revealing the selection intentions. With the preliminary eye movement features determined by statistical analysis, a feature selection with an elastic net was employed to further identify the most effective features for discriminating users’ model selection intention. Given the imbalanced nature of the collected eye movement dataset, we developed a support vector machine (SVM) equipped with an adaptive conformal transformation (ACT) kernel to identify users’ CAD model selection intentions based on their eye movements. The experiment results demonstrated that the developed classifier can effectively discriminate between different CAD model selection intentions with an average accuracy of 91.71
There is extensive research on measuring Cognitive Load (CL) using both behavioural and physiological indicators. However, we do not know whether existing datasets used to measure CL are transferable across contexts. We systematically reviewed non-invasive measurement methods in naturalistic settings using PRISMA guidelines, searching five databases for studies published between 2015-2025. We analysed 24 studies focusing on non-invasive measurement methods and found that multimodal approaches reported the highest accuracy (up to 95.4
Collective intelligence refers to a team’s general ability to perform well across a wide variety of tasks, a stable team-level analogue of individual intelligence. Although it fundamentally emerges from interaction, it is still largely treated as a static property or as an aggregate performance outcome. This article proposes a computational proof of concept that operationalizes a dynamic, process-oriented model of collective intelligence as a multilayer pipeline instantiated in immersive virtual reality. The model articulates three complementary views: collective intelligence as a measurable capacity (the C-factor), as an emergent state produced by three transactive subsystems, that is, shared processes through which teammates jointly manage memory (“who knows what”), attention (what the team focuses on), and reasoning (how goals are aligned), and as a performance gain over individual baselines, all embedded within a recursive Input-Mediator-Output-Input cycle. A multi-user VR collaborative task built on role interdependence and distributed information enables the synchronized capture of each participant’s multimodal behavioral traces (speech, gaze, facial expressions). These signals are transformed into windowed features and fused into subsystem-level indicators, whose recursive updates formalize collective intelligence as a temporally unfolding process rather than a fixed score. A pilot study (six triads) demonstrates the technical feasibility and user acceptability of the system, as well as meaningful variability in performance and interaction dynamics; higher-performing teams exhibit multimodal patterns consistent with more efficient attentional and regulatory coordination. This work bridges these perspectives within a unified computational framework and lays the groundwork for the longitudinal study of collective intelligence and for multimodal systems designed to support it.
