In group discussions, silence conveys important intentions, yet its invisible nature frequently causes awkwardness and misinterpretation. While visualizing the motives behind silence helps mutual understanding, sharing this information in an identifiable format creates psychological pressure on participants. To address this, we investigated an intervention that aggregates self-reported silence motives into an anonymized distribution, with its presentation timing automatically controlled based on speech pauses. We evaluated this approach through a within-subjects experiment involving four-person groups engaged in consensus-building tasks. The results indicated that anonymization and neutral labeling contributed to lowering psychological barriers to self-disclosure and enhancing mutual understanding. However, strictly visualizing silence based on a predefined temporal threshold inadvertently induced an “alerting effect,” suggesting that stripping away natural conversational buffers may amplify perceived awkwardness and hinder discussion fluency. These findings provide crucial design implications: while state-collection inputs must ensure psychological safety, the presentation of such information must prioritize the natural rhythm and social acceptability of human communication over strict real-time responsiveness.
Universities seek to promote interdisciplinary research, but internal collaboration networks often remain clustered around established teams, making cross-group collaborations difficult to identify and activate. This paper treats collaboration recommendation as an institution-level decision-support problem. Using OpenAlex publication metadata combined with administrative affiliation records, we reconstruct the internal co-authorship network of Universitat Pompeu Fabra (UPF), covering 5,529 authors and 27,818 co-authorship links from 2014–2024, and enrich it with a topic-similarity layer. We evaluate single-layer Adamic–Adar and a multiplex extension combining co-authorship and topical proximity using a temporal split between 2014–2019 and 2020–2024 on 701 active authors. We assess both predictive accuracy and the structural effects of adding recommended links, including changes in clustering and degree assortativity. We show that topical similarity, modeled as a separate multiplex layer, changes the type of collaborations recommended and reveals a trade-off between preserving local cohesion and producing more bridge-like structural effects. Among the evaluated variants, the multiplex recommender achieves the highest predictive performance. A blinded validation study with 16 UPF researchers further indicates that multiplex recommendations are more often perceived as relevant and accurate than a co-authorship-only baseline.
In individual learning environments such as e-learning, the difficulty of maintaining concentration for extended periods causes a decline in learning effectiveness. To address this issue, various support methods have been explored in which a system detects a learner’s decline in concentration and intervenes. However, existing methods face a trade-off problem. Specifically, making the intention of the intervention explicit tends to induce psychological reactance, whereas concealing the intention makes it difficult to encourage spontaneous behavioral correction. To resolve this issue, we have proposed a learning support method based on objective self-awareness theory. This method physically actuates the monitor itself, which serves as the display for the learning task, to externalize the learner’s unfocused state through a dozing motion. In this paper, we investigate how differences in motion parameters affect user perceptions of the physical monitor agent’s dozing motion. An experiment with 12 participants revealed that perceived task hindrance was significantly lower for a modulated motion, consisting of a slow forward motion and a high-speed backward tilt, than for a constant high-speed motion profile. Additionally, a marginally significant difference was observed between the high-speed forward and backward motion and the slow forward and backward motion. These results suggest that combining a slow forward tilt with a fast backward return may reduce perceived task hindrance in the dozing motion of a physical monitor agent.
This work explores the Wellbeing Journey (WB Journey), an educational application developed to enhance student well-being in higher education through personalized recommendations that target individual and collective well-being awareness. The paper provides a comprehensive walkthrough of the tool and assesses its perceived value and usability through a multi-stage evaluation involving higher education students. This evaluation includes a thematic analysis of value with 125 participants, followed by two distinct UX studies with 25 and 22 students, respectively. The results indicate that the WB Journey shows potential to support self-regulation, consistent with SDT principles. Future research will focus on iterative UX improvements and the development of a best-practice guide to assist educators in the strategic implementation of the tool within academic environments.
This exploratory study investigates the role of personality-based personalisation of LLMs on human workload during human–AI collaboration. To examine whether personalisation effects are consistent across different collaborative problem-solving contexts, two structurally distinct tasks were used: a Sudoku task (n = 24) and a Zebra puzzle (n = 27). Personality-based personalisation was operationalised through personality traits derived from the EPQR-A, resulting in three agent conditions: personality-aligned, personality-inverted, and non-personalised (control). Workload was assessed using the NASA Task Load Index (NASA-TLX). Epistemic Network Analysis (ENA) was used to model human–AI problem-solving process. Findings from NASA-TLX suggest a trend toward higher mental demand, effort, and frustration alongside a relatively lower performance in the non-personalised condition compared to both personality-aligned and personality-inverted conditions. ENA showed that the non-personalised condition was characterised by agent-dominated interactions, whereas the personality-aligned condition reflected greater collaborative engagement compared to personality-inverted condition. Overall, the findings suggest that while personality-aligned interactions can enhance collaborative engagement, the absence of personalisation may increase cognitive burden, thereby highlighting personality-based personalisation of LLMs as a relevant design factor in human–AI collaborative systems.
