Training professionals in high-stakes, trauma-informed communication is critical across domains such as law enforcement, healthcare, and counseling. While live role-play with trained actors remains the gold standard, it is resource-intensive and emotionally demanding. Generative AI offers scalable alternatives, but what is gained or lost when training shifts to AI? We developed an AI-powered sexual assault victim interview training system and conducted a mixed-methods study with 35 police recruits, each completing both an AI-based and a live, actor-based training session. By varying the sequence (AI-first vs. human-first), we examined differences in self-efficacy, perceptions of the AI system, and perceived learning experience. Although both modalities supported learning, the order in which they were experienced significantly shaped learners’ emotional engagement, sense of preparedness, and interpretation of each simulation’s role. Building on these insights, we introduce a conceptual design framework that identifies social–emotional, temporal, and embodied distance as key pedagogical dimensions, and we offer implications for sequencing hybrid simulations to scaffold preparation, performance, and reflection. Our findings position AI not as a replacement for human realism, but as a complementary modality that expands opportunities for safe, scalable practice in sensitive communication training.

As AI support tools become more common in immersive medical training, designers face a critical interaction-design question: When should an AI facilitator take initiative, and when should it wait for the learner? To investigate this design tension, we compared two versions of an AI facilitator in a mixed-reality (XR) lumbar puncture simulator training conditions: one in which the AI proactively initiated guidance and encouragement, and another in which the AI responded only when prompted. Using a mixed-methods approach, we examined how medical students (n=22) engaged with, interpreted, and reacted to these two facilitation styles. We found no significant differences in learning outcomes, interaction frequency, or overall experience ratings. However, interviews and behavioral analyses revealed nuanced differences in how learners perceived AI interventions across distinct task phases. AI-initiated support was seen as helpful in some moments and disruptive in others, depending on task phase, cognitive load, and personal preferences. Based on these findings, we contribute a boundary framework which offers actionable design guidance for calibrating AI proactivity in immersive training systems, and extends HCI research on proactive agents and human–AI collaboration within high-cognitive-load environments.

What if future dining involved eating robots? We explore this question through a playful and poetic experiential dinner theater: a tangible design fiction staged as a 2052 Paris restaurant where diners consume a biohybrid flying robot in place of the banned delicacy of ortolan bunting. Moving beyond textual or visual speculation, our “dinner-in-the-drama” combined performance, ritual, and multisensory immersion to provoke reflection on sustainability, ethics, and cultural identity. Six participants from creative industries engaged as diners and role-players, responding with curiosity, discomfort, and philosophical debate. They imagined biohybrids as both plausible and unsettling—raising questions of sentience, symbolism, and technology adoption that extend beyond conventional sustainability framings of synthetic meat. Our contributions to HCI are threefold: (i) a speculative artifact that stages robots as food, (ii) empirical insights into how people negotiate cultural and ethical boundaries in post-natural eating, and (iii) a methodological advance in embodied, multisensory design fiction.

Decomposition, the process of breaking down complex problems into manageable parts, is a fundamental component of computational thinking (CT) but remains challenging for novice learners. We present Spark, a multimodal programming interface that supports the externalization of decomposition thinking by organizing user-articulated goals into structured steps and enacting them through a tangible robot, making decomposition visible and open to inspection. The design of Spark is theory-driven, with its user interface aligned to three decomposition rationales: substantive, relational, and functional decomposition. In a study with 20 adult novices, we compared Spark with Scratch, an educational block-based visual programming interface. While both systems were associated with improvements in participants’ self-reported decomposition skills, only Spark was associated with measurable gains on objective assessments and significantly higher task success when experienced first, while maintaining comparable workload and completion times. Participants reported that externalizing the otherwise hidden process of decomposition made programming more tangible and motivating. These findings, highlighting the complementary roles of multimodal interaction in shaping novices’ decomposition experiences, inform the design of interactive programming interfaces that aim to support the externalization of reasoning processes. More broadly, our work contributes to the field of human-AI interaction in learning, illustrating how multimodal interaction can be responsibly integrated to scaffold reasoning, support reflection, and promote equitable participation in computing.

Peer Reviewed

Physics-Informed Neural Networks (PINNs) are a kind of deep-learning-based numerical solvers for partial differential equations (PDEs). Existing PINNs often suffer from failure modes of being unable to propagate patterns of initial conditions. We discover that these failure modes are caused by the simplicity bias of neural networks and the mismatch between PDE’s continuity and PINN’s discrete sampling. We reveal that the State Space Model (SSM) can be a continuous-discrete articulation allowing initial condition propagation, and that simplicity bias can be eliminated by aligning a sequence of moderate granularity. Accordingly, we propose PINNMamba, a novel framework that introduces sub-sequence modeling with SSM. Experimental results show that PINNMamba can reduce errors by up to 86.3% compared with state-of-the-art architecture. Our code is available at Supplementary Material.

Identifying effective interventions from the scientific literature is challenging due to the high volume of publications, specialized terminology, and inconsistent reporting formats, making manual curation laborious and prone to oversight. To address this challenge, this paper proposes a novel framework leveraging large language models (LLMs), which integrates a progressive ontology prompting (POP) algorithm with a dual-agent system, named LLM-Duo. On the one hand, the POP algorithm conducts a prioritized breadth-first search (BFS) across a predefined ontology, generating structured prompt templates and action sequences to guide the automatic annotation process. On the other hand, the LLM-Duo system features two specialized LLM agents, an explorer and an evaluator, working collaboratively and adversarially to continuously refine annotation quality. We showcase the real-world applicability of our framework through a case study focused on speech-language intervention discovery. Experimental results show that our approach surpasses advanced baselines, achieving more accurate and comprehensive annotations through a fully automated process. Our approach successfully identified 2,421 interventions from a corpus of 64,177 research articles in the speech-language pathology domain, culminating in the creation of a publicly accessible intervention knowledge base with great potential to benefit the speech-language pathology community.

Understanding emerging safety risks in nuanced social VR spaces and how existing safety features are used is crucial for the future development of safe and inclusive 3D social worlds. Prior research on safety risks in social VR is mainly based on interview or survey data about social VR users' experiences and opinions, which lacks "in-situ observations" of how individuals react to these risks. Using two empirical studies, this paper seeks to understand safety risks and safety design in social VR. In Study 1, we investigated 212 YouTube videos and their transcripts that document social VR users' immediate experiences of safety risks as victims, attackers, or bystanders. We also analyzed spectators' reactions to these risks shown in comments to the videos. In Study 2, we summarized 13 safety features across various social VR platforms and mapped how each existing safety feature in social VR can mitigate the risks identified in Study 1. Based on the uniqueness of social VR interaction dynamics and users' multi-modal simulated reactions, we call for further re-thinking and re-approaching safety designs for future social VR environments and propose potential design implications for future safety protection mechanisms in social VR.

Early conversational agents (CAs) focused on dyadic human-AI interaction between humans and the CAs, followed by the increasing popularity of polyadic human-AI interaction, in which CAs are designed to mediate human-human interactions. CAs for polyadic interactions are unique because they encompass hybrid social interactions, i.e., human-CA, human-to-human, and human-to-group behaviors. However, research on polyadic CAs is scattered across different fields, making it challenging to identify, compare, and accumulate existing knowledge. To promote the future design of CA systems, we conducted a literature review of ACM publications and identified a set of works that conducted UX (user experience) research. We qualitatively synthesized the effects of polyadic CAs into four aspects of human-human interactions, i.e., communication, engagement, connection, and relationship maintenance. Through a mixed-method analysis of the selected polyadic and dyadic CA studies, we developed a suite of evaluation measurements on the effects. Our findings show that designing with social boundaries, such as privacy, disclosure, and identification, is crucial for ethical polyadic CAs. Future research should also advance usability testing methods and trust-building guidelines for conversational AI.

Many couples experience long-distance relationships (LDRs), and "couple technologies" have been designed to influence certain relational practices or maintain them in challenging situations. Chatbots show great potential in mediating people's interactions. However, little is known about whether and how chatbots can be desirable and effective for mediating LDRs. In this paper, we conducted a two-phase study to design and evaluate a chatbot, PocketBot, that aims to provide effective interventions for LDRs. In Phase I, we adopted an iterative design process through conducting need-finding interviews to formulate design ideas and piloted the implemented PocketBot with 11 participants. In Phase II, we evaluated PocketBot with eighteen participants (nine LDR couples)in a week-long field trial followed by exit interviews, which yielded empirical understandings of the feasibility, effectiveness, and potential pitfalls of using PocketBot. First, a knock-on-the-door feature allowed couples to know when to resume an interaction after evading a conflict; this feature was preferred by certain participants (e.g., participants with stoic personalities). Second, a humor feature was introduced to spice up couples' conversations. This feature was favored by all participants, although some couples' perceptions of the feature varied due to their different cultural or language backgrounds. Third, a deep talk feature enabled couples at different relational stages to conduct opportunistic conversations about sensitive topics for exploring unknowns about each other, which resulted in surprising discoveries between couples who have been in relationships for years. Our findings provide inspiration for future conversational-based couple technologies that support emotional communication.