Efective human-AI teaming heavily depends on swift trust, particularly in high-stakes scenarios such as emergency response, where timely and accurate decision-making is critical. In these time-sensitive and cognitively demanding settings, adaptive explainability is essential for fostering trust between human operators and AI systems. However, existing explainable AI (XAI) approaches typically ofer uniform explanations and rely heavily on explicit feedback mechanisms, which are often impractical in such high-pressure scenarios. To address this gap, we propose a conceptual framework for adaptive XAI that operates non-intrusively by responding to users' real-time cognitive and emotional states through implicit feedback, thereby enhancing swift trust in high-stakes environments. The proposed adaptive explainability trust framework (AXTF) leverages physiological and behavioral signals, such as EEG, ECG, and eye tracking, to infer user states and support explanation adaptation. At its core is a multi-objective, personalized trust estimation model that maps workload, stress, and emotion to dynamic trust estimates. These estimates guide the modulation of explanation features enabling responsive and personalized support that promotes swift trust in human-AI collaboration. This conceptual framework establishes a foundation for developing adaptive, non-intrusive XAI systems tailored to the rigorous demands of high-pressure, time-sensitive environments.
In high stakes domains such as emergency response [
Explainability has emerged as a central mechanism for cultivating trust in AI, enabling humans to
understand and anticipate AI behavior [
To overcome these gaps, incorporating implicit human feedback is essential for advancing
explainable AI (XAI) systems, especially in high-stakes environments. Non-invasive technologies, such as
wearable sensors [
This work presents adaptive explainability trust framework (AXTF), a conceptual framework designed to advance human-AI teaming in high-stakes, time-sensitive domains by enabling adaptive, non-intrusive explainability driven by multi-objective trust estimation model. It outlines a foundational approach that combines implicit human feedback, AI performance metrics, and situational awareness to infer the evolving trust state of the user. At the core of this framework is a personalized trust inference model that integrates the user’s cognitive and emotional state along with situational awareness to infer dynamic trust levels. These estimates guide the adaptive modulation of explanation features such as timing, granularity, content, and presentation mode, enabling dynamic, context-aware explanation strategies that foster swift trust. Importantly, this approach reflects a neurosymbolic paradigm by combining low-level physiological sensing with interpretable symbolic reasoning via fuzzy rules. This conceptual approach lays the foundation for future research and the practical development of trust-sensitive, non-intrusive XAI systems tailored to the demands of time-critical, high-pressure domains.
Unlike task specific or opaque AI models [ 16], our framework is generalizable across high stakes domains, interpretable by design, and adaptable in real time. It supports collaboration by recognizing the cognitive and afective constraints of human operators and responding accordingly closing the loop between trust inference, explanation adaptation, and mission performance. In the following sections, we review prior work, detail our conceptual model, and outline its application to real-world high-pressure settings such as emergency response.
This section introduces key foundations for our proposed framework by synthesizing prior work across four core areas. First, we examine the role of implicit feedback in high-stakes human-AI collaboration, emphasizing the need for implicit, real-time cues such as physiological and behavioral signals to support decision-making under stress and cognitive load. Next, we explore the construct of swift trust, outlining its determinants, reliability, predictability, competence, transparency, and adaptability and their sensitivity to fluctuating user states. We then review the limitations of adaptive explainability, highlighting gaps in existing XAI systems that fail to adjust explanations to changing user conditions. Finally, we discuss user-centric and afect-aware XAI, emphasizing emerging evidence for modeling trust through real-time physiological inference, and outlining the need for integrated models that dynamically adapt explanation features to maintain trust and cognitive eficiency in time-sensitive contexts. This background frames the motivation and design of our proposed conceptual model.
In high stakes domains, human AI collaboration is often task critical, requiring both agents to operate
in close coordination under intense cognitive and temporal pressure. The efectiveness of this teaming
rests heavily on the human operator’s ability to trust the AI system to understand its role, predict its
behavior, and rely on its outputs in moments of uncertainty. Studies across domains such as emergency
response and autonomous operations show that trust in AI systems enhances decision making eficiency,
reduces cognitive burden, and improves overall team performance [
Implicit feedback, measured through physiological and behavioral signals, ofers a promising
foundation for adaptive support in human-machine teams (HMTs). Unlike explicit feedback, implicit indicators
are passively observable and continuous, providing a non-intrusive method for assessing the user’s
cognitive and emotional state. A growing body of research supports the viability of using such signals
for trust estimation. For instance, EEG has been shown to correlate with cognitive load and attention
[
By grounding our model in these implicit signals, we enable AI systems to dynamically assess the operator’s cognitive and emotional state and adapt their behavior accordingly. In the context of XAI, this means tailoring explanations to be more concise, timely, or expressive depending on the inferred user state. For example, a spike in physiological arousal following an AI action may indicate confusion or concern, prompting the system to proactively issue a clarifying explanation. Similarly, indicators of high trust and low load might invite more detailed, exploratory explanations to support learning or calibration. In this way, implicit feedback enables real time, user sensitive adaptation, forming a critical bridge between human trust dynamics and machine explainability. It allows the AI system to act as an responsive teammate not just explaining what it does, but choosing when and how to explain based on the operator’s needs. This perspective forms the foundation for the next section, where we examine the elements of swift trust, and how they intersect with user state and explainability in high stake teams.
In high stakes human AI collaboration, trust must be formed quickly often in the absence of prolonged
interaction or past performance history. This phenomenon, referred to as swift trust [
A large body of empirical work (e.g.[17, 26, 18]) confirms that trust in AI systems is closely linked to
perceived performance, system reliability and predictability, as well as the user’s workload, stress, and
emotional state. Hancock et al. [
To model and support swift trust efectively, it is useful to decompose it into key elements, as outlined in Table 1 and commonly cited in the literature [28, 29, 19]. Among these, adaptability plays a critical role in high-stakes settings where user state and task demand shift rapidly. It reflects the AI system’s responsiveness to physiological, cognitive, and contextual signals, allowing it to adjust its explanations (e.g., simplifying content during stress) and behaviors (e.g., increasing feedback frequency during uncertainty) to maintain trust. As Cho et al. [29] and Seong and Bisantz [30] note, adaptive systems promote more accurate and timely trust calibration, allowing the user to rapidly align trust with the actual performance of AI.
These elements are not static; rather, they are dynamically influenced by both user state and system behavior. An efective trust-supporting system must monitor changes in stress, workload, and emotional valence and adjust its communication strategy accordingly. Further, these elements can be modulated by explainability features, but only if the system is responsive to the underlying user state. For instance, low predictability can be improved by proactive explanations. Low transparency can be mitigated with “why” explanations about intent. Low reliability perception under stress may be best addressed with brief confidence statements (e.g., “High certainty: obstacle detected”) [ 19]. In the next section, we explore how explainability mechanisms can be leveraged to modulate these elements and support the dynamic formation and maintenance of swift trust.
Explainability has long been recognized as a mechanism for fostering trust in AI systems. However, most existing approaches are static and context agnostic, ofering fixed explanations that do not adjust to the user’s state or task environment. Recent advances have begun to explore adaptive explanation strategies for instance, using reinforcement learning or partially observable Markov decision processes (POMDPs) to tailor explanations based on user type or task progression [31]. However, these approaches often rely on predefined user profiles or require explicit feedback, making them dificult to apply in high stakes, real time environments. Model reconciliation approaches [31] align AI explanations with human mental models, but typically assume static trust misalignment and do not account for fluctuating physiological or afective states. Floyd et al. [ 32] introduced trust guided transparency, where the AI modulates its behavior based on estimated user trust, but their system relied on explicit interaction logs and performance scores, rather than physiological signals.
Recent research has demonstrated that trust related cognitive and emotional states can be inferred
using physiological and behavioral signals such as heart rate variability (HRV), electrodermal activity
(EDA), facial expressions, and gaze [
Moreover, trust calibration aligning user trust with system capability is especially critical in high-risk
or time sensitive domains. Studies in aviation, medicine, and robotics have shown that miscalibrated
trust leads to automation bias or disuse [
Early work on user aware XAI [16] focused on clustering users by behavior patterns to personalize
explanations. However, most approaches lack real time responsiveness, and few incorporate afective
signals to dynamically adjust content. Ali et al. [
behavior [17] Unfamiliar task (Exper- Increases mental efort; risks trust erosion tise) [18] Disengagement
In Human Machine Teams (HMTs), trust and explainability are dynamically interlinked, unfolding as
a sequence of cause-and-efect interactions. AI behaviors whether task related actions, feedback, or
navigation decisions directly influence the human operator’s physiological state, emotional response,
and cognitive processing. These changes, in turn, shape the operator’s perception of trust, impacting
collaboration quality and task performance [
AI Behavior and Trust Perception. The observable behavior of AI such as decision making, task
execution, or error handling shapes the operator’s perception of its competence, reliability, and intent.
When the AI behaves transparently and contextually, users are more likely to perceive it as trustworthy.
In contrast, opaque or inconsistent behaviors introduce uncertainty and distrust. Shin [
Emotional and Physiological Responses. Trust perceptions are further mediated by afective responses, which manifest physiologically through changes in heart rate (HR), heart rate variability (HRV), electrodermal activity, or neural activity (EEG) [35]. For example, unexpected or ambiguous AI actions may trigger stress or frustration, while cooperative and predictable behavior fosters engagement and calm. These physiological markers serve as real time, implicit indicators of trust state [19], enabling continuous user monitoring without explicit intervention.
Cognitive Load and Information Processing. AI outputs that are complex, ambiguous, or mistimed can impose a high cognitive burden, impairing the user’s ability to process information and make timely decisions. This efect is especially pronounced in emergency response scenarios, where attention is divided, time is limited, and errors are costly. Prior studies [18, 19] show that cognitive overload negatively afects both trust and performance in collaborative human AI systems. Adaptive explainability can address this by modulating explanation complexity, timing, and content helping reduce overload, maintain attention, and recalibrate trust in real time.
However, most XAI systems do not modulate their explanations based on implicit human signals such as stress, workload, or emotion, nor do they account for rapidly changing contextual cues. This limitation is particularly consequential in high-stakes environments, where cognitive overload and uncertainty diminish the operator’s ability to process static or overly generic explanations. There remains a significant gap in designing closed-loop adaptive XAI systems that leverage real-time physiological and contextual data to both interpret user states and dynamically tailor explanation strategies. Such systems would not only reflect the user’s cognitive and emotional state but also influence it over time supporting trust formation, cognitive eficiency, and collaborative resilience under pressure.
The findings presented in Section 2 lead to a critical insight: explainability is not merely a communication tool, but a dynamic mechanism for shaping and stabilizing swift trust. To support this, we propose, Adaptive Explainability Trust Framework (AXTF), a conceptual framework designed to enhance human decision-making within human-AI teams by improving swift trust and team performance through dynamic adaptation of AI explanations based on real-time assessments of user states and environmental factors. Specifically, our framework enables AI explanations to be continuously tailored according to the user’s cognitive load, stress, and emotional state, as well as the task context, including urgency, goals, and environmental complexity, particularly in high-stakes domains such as emergency response.
To support swift trust and efective human AI teaming in high stakes environments, our proposed conceptual model links real time physiological and behavioral indicators of human state to adaptive explainability mechanisms. The model integrates three interconnected components: (1) multimodal feedback sensing and inference, (2) multi-objective trust modeling, and (3) explanation feature adaptation, with environmental inputs to ensure contextual relevance.
The proposed framework (Fig. 1) forms a closed-loop pipeline that integrates real-time physiological and behavioral signals (e.g., EEG, ECG, heart rate variability—HRV) with environmental data such as task goals, urgency, and state to assess the user’s cognitive load (W), stress (S), and emotional valence (E) levels [18, 36]. Based on these user states and environmental performance metrics (e.g., task errors, success rates), dynamic trust estimation is performed using a multi-objective neurofuzzy rule-based inference engine. The framework then adapts key explanation features, including timing, duration, granularity, and mode of delivery, by mapping these trust estimates and contextual knowledge to reduce cognitive overload, enhance trust, and guide the subsequent human action (+1). For example, in scenarios characterized by low trust and high cognitive load, short, moderate steps, and reactive explanations are more efective for fostering trust compared to lengthy, detailed ones. While some lfuctuations in cognitive load ( ) and trust () are expected, the ultimate goal of these adaptive explanations is to increase trust (T) and decrease cognitive load (W) over time, thereby promoting swift trust and improving situational awareness, decision-making, and overall team performance. This feedback-driven framework is designed to support temporal situational awareness, workload balancing, and trust resilience in high-pressure environments where explicit communication may be limited, but implicit signals provide actionable insight. The following subsections detail each component of the framework, including multimodal feedback sensing and inference, multi-objective trust modeling, and explanation feature adaptation.
The system continuously monitors a range of physiological and behavioral signals to infer latent cognitive and emotional states that are critical for trust assessment, including: • EEG and pupillometry for evaluating cognitive workload, • ECG, galvanic skin response (GSR), and heart rate variability (HRV) to detect stress and arousal levels, • Facial expressions, gaze tracking, and voice features to infer emotional valence and user engagement.
Physiological signals are first processed using personalized machine learning models that learn
discriminative representations directly from raw or minimally preprocessed data [
To translate these human state metrics into actionable trust estimates, we suggest a multiobjective neurofuzzy inference method. This hybrid approach exemplifies a neurosymbolic architecture, where neural signal processing informs symbolic reasoning. It combines the perceptual strength of physiological sensing with interpretable symbolic decision logic to support transparent trust calibration. While current fuzzy membership functions are fixed for conceptual clarity, the framework is designed to support personalization and future learning-based adaptation where parameters can be tuned per user or task. This approach serves as an interpretable guardrail that encodes literature-grounded, rule-based mappings (see Table 3), helping filter and structure physiological and contextual signals before any downstream reasoning (e.g., via large language models) is performed.
Trust is estimated as a categorical variable with three levels Low, Medium, or High based on a combination of user physiological and afective states and system performance metrics. The model enables interpretable reasoning that links observed human state and AI behavior to well-established trust dimensions such as reliability, competence, predictability, transparency, and adaptability. The model takes the following inputs: • Input Variables – Workload (W): categorized as Low, Medium, or High; derived from EEG features, gaze data, or behavioral task-switching patterns. – Stress level (S): categorized as Low, Medium, or High; inferred from ECG, GSR, and HRV metrics. – Emotion valence (E): categorized as Negative, Neutral, or Positive; estimated from facial expressions, tone of voice, or afective models. – System performance score (P): a normalized score between 0 and 1; computed from task metrics such as success rate and error frequency, allowing generalization across domains. • Trust Output
– Trust (T): classified as Low, Medium, or High.
The fuzzy trust inference system maps normalized input variables into fuzzy linguistic categories using membership functions. Each input (e.g., workload, stress, emotion valence, performance) is associated with three fuzzy sets: Low, Medium, and High, except for emotion valence, which uses Negative, Neutral, and Positive.
1. Workload (W), Stress (S)
Both workload and stress are defined over the domain [
⎧⎪1, ≤ 0.2 Low() = ⎨0.05.−3 , 0.2 < ≤ 0.5
⎪⎩0, > 0.5 2. Emotion Valence (E)
⎧⎪0, ≤ 0.2 or ≥ 0.8 Medium() = ⎨−00..23 , 0.2 < ≤ 0.5 ⎪⎩0.08.−3 , 0.5 < < 0.8
⎧⎪0, ≤ 0.5
High() = ⎨−00..53 , 0.5 < ≤ 0.8
⎪⎩1, > 0.8
Emotion valence is modeled over the range [
These functions are derived from literature indicating that increased workload and stress reduce
trust in automation [
The model focuses on seven core explainability features that shape user experience and collaboration.
We define these key features as follows:
1. Timing: Timely delivery of explanations is vital in high-pressure situations. Explanations can
be delivered proactively, ahead of an AI action, to minimize potential confusion (e.g., "Avoiding
unstable debris ahead"), or reactively, triggered by user hesitation or unexpected AI behavior.
In an emergency context, timing must align with both the task phase and the user’s attentional
bandwidth to avoid distraction or delay [
High-level summaries, such as "Scanning lower level first," help reduce the user’s information processing demands. In contrast, detailed step-by-step explanations, for example, "Entering sector B → mapping → thermal anomaly detected," are better suited for experienced users or situations where trust is high. The granularity should be adapted based on factors such as user familiarity, workload, and trust levels [18, 33, 38] to ensure explanations remain cognitively accessible while still informative. 4. Content: Explanation content is chosen based on task relevance and the user’s current focus.
Contextual or local explanations, such as "Rerouting due to obstacle," emphasize immediate actions or environmental conditions. In contrast, hierarchical explanations, for example, "Prioritizing lower floors due to heat signature density," communicate broader planning strategies. In emergency scenarios, providing contextual content enhances temporal situational awareness and responsiveness [39, 30]. 5. Transparency: Transparency shapes the user’s understanding of the AI’s decision-making process. "How" transparency, such as "Based on heatmap and terrain risk, path updated," helps users evaluate the system’s methods. "Why" transparency, for example, "Avoiding risk to maximize coverage," clarifies the AI’s intent and goal reasoning. Both types support mental model alignment and trust; however, the appropriate level and form of transparency should be adapted based on the user’s state and the context [40, 28, 32]. 6. Adaptability: Adaptability acts as the central mechanism that dynamically modulates all other explainability features in real time. This capability allows AI systems to selectively tailor explanations to align with the user’s current state and task objectives. Under conditions of high workload or stress, the system simplifies explanations and employs lower-detail modes to reduce cognitive load. Conversely, when users are calm and trust levels are high, the system can provide more complex and interactive explanations. This adaptability ensures that explanations remain both informative and sustainable, especially in high-pressure environments [41, 28, 32, 16]. 7. Mode of Delivery: The medium used to deliver explanations significantly influences user comprehension and cognitive load. Visual formats such as maps, trajectory overlays, or alert icons are ideal when the user’s auditory attention is available. Textual explanations, including on-screen summaries or status updates, are more suitable during quieter moments or post-task phases. Auditory delivery through spoken instructions works best when the user’s hands and eyes are occupied. Combining multiple channels in a multimodal approach—such as spoken plus visual alerts—enhances system resilience and inclusivity. Research by Adadi and Berrada [42] demonstrates that multimodal explanations improve user understanding, particularly when users are multitasking or experiencing stress.
The inferred trust level is used to modulate these key explainability features. This trust adaptive mechanism supports continuous calibration, allowing the system to respond not just to performance, but to how the human feels and functions during collaboration.
While explainability is often discussed as a means of improving user understanding or meeting regulatory requirements, in high stakes human AI teaming, it serves a deeper role: calibrating trust in real time. As trust is highly sensitive to changes in workload, stress, and afect, static or misaligned explanations may unintentionally erode confidence or overload the user. In contrast, adaptive explainability tuned to the user’s current cognitive and emotional state can serve as a powerful tool for trust repair, reinforcement, and regulation. Consider a search and rescue drone system operating in post-disaster environments. A human operator under high stress and workload may be overwhelmed by frequent decision updates. The system detects high HRV, elevated EEG load, and negative valence, infers low trust, and adapts by issuing short, confidence-framed explanations through audio (“Clear path detected. High certainty.”). If later states show reduced load and increased engagement, it shifts to detailed, interactive visualizations for planning and collaboration.
While this work establishes a conceptual foundation for adaptive explainability in high-stakes human-AI teaming, several important avenues remain for future investigation. First, two-way communication between human and AI teammates should be more explicitly modeled not only to adapt AI explanations to user states, but also to enable reciprocal influence where the AI dynamically adjusts its behavior based on user reactions and evolving task demands. Such bi-directional interaction will allow the AI to both respond to human needs and evaluate and refine its own performance, closing the loop between perception, explanation, and behavior. Second, cultural and dispositional factors play a foundational role in trust formation but remain underexplored in adaptive XAI. Future research should investigate how traits such as uncertainty avoidance, communication preferences, and prior experience influence trust dynamics and explanation preferences, enabling more inclusive and culturally-aware explanation strategies. Third, implementation and evaluation in interactive, dynamic environments, such as simulation-based emergency response scenarios, will be essential to validate the framework. These settings ofer controllable, high-fidelity contexts for assessing how real-time physiological and behavioral feedback impacts explanation efectiveness, trust calibration, and team performance under pressure. Fourth, advances in generative agent simulations [43] open opportunities for large-scale validation using synthetic populations embedded with memory, social reasoning, and behavioral diversity. These agent-based testbeds can be used to examine long-term trust trajectories and crossprofile adaptation strategies in simulated high-stakes team settings. Finally, to ensure AI safety in high-pressure decision environments, future work must incorporate safeguards that prevent explanation misuse or cognitive overload. Adaptive explanation systems must remain transparent, interpretable, and bounded by safety constraints that prevent miscalibration of trust—especially under uncertainty or stress. Embedding safety-aware logic into adaptation rules (e.g., thresholds on explanation complexity or delivery timing) will help maintain alignment with human cognitive capacity, trust boundaries, and ethical standards in mission-critical operations.
In conclusion, advancing adaptive explainability through bi-directional interaction, cultural awareness, scalable simulation, and embedded safety principles will help realize the next generation of trust-sensitive, cognitively aligned, and ethically grounded human-AI systems. While this work is conceptual, validation eforts are underway. Preliminary pilot studies have shown that latency in explanation delivery can negatively impact trust[17], reinforcing the need for real-time adaptation. The use of multimodal signals (EEG, ECG, GSR, etc.) allows the system to remain robust even if individual channels degrade. Future work will explore how conflicting or missing signals are resolved through confidence-based weighting and ensemble logic.
Finally, ethical considerations are critical when designing adaptive explanation mechanisms. While tailoring explanations to user states can improve usability and decision support, there is also a risk that such adaptations could artificially inflate user trust beyond what the system’s actual competence warrants. For example, if the system simplifies or frames its reasoning too persuasively under highstress conditions, users may place undue reliance on it, leading to overtrust and potential harm in safety-critical settings.
This work contributes to afective, situated, and trustworthy AI by introducing a conceptual framework for real-time adaptation of explainability to support swift trust and efective teamwork in high-stakes environments. The framework integrates multimodal implicit feedback physiological, behavioral, environmental, and contextual signals to infer user states such as workload, stress, and emotional valence. These inferred states inform the dynamic adjustment of explanation features (e.g., timing, granularity, modality), enabling alignment with the user’s cognitive and afective demands in pursuit of time-critical task goals. Explainability is reframed as a multi-objective adaptive function balancing transparency, cognitive eficiency, and trust calibration. The framework is designed to be model-agnostic and extensible, supporting the integration of diverse trust modeling and learning mechanisms. It enables AI systems to act as responsive teammates fostering trust, maintaining collaboration under pressure, and supporting decision-making when explicit communication is limited. This lays a foundation for generalizable real-world deployment of adaptive XAI in high-stakes domains such as emergency response, medical operations, and mission-critical decision support, where human-machine teaming must remain transparent, afect-aware, and cognitively eficient under uncertainty.
Unlike existing XAI models that rely on predefined user profiles or static inference pipelines[ 18, 31], our framework incorporates real-time, user-specific trust estimation and links these inferences directly to explanation adaptation. It addresses a critical gap in adaptive XAI literature by treating explainability as a dynamic, multi-objective function responsive to workload, stress, and afective cues. This personalized and continuous adjustment diferentiates our approach from prior eforts that lack temporal responsiveness or user-state integration.
This research was funded by the Asian Ofice of Aerospace Research and Development through Grant 23IOA087.
During the preparation of this work, the author(s) used GPT-4 in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content. for culturally adaptive emotional intelligence in hci, 2025. URL: https://arxiv.org/abs/2506.14166. arXiv:2506.14166. [14] X. Hao, B. Nakisa, M. N. Rastgoo, G. Pang, Bcr-drl: Behavior- and context-aware reward for deep reinforcement learning in human-ai coordination, 2025. URL: https://arxiv.org/abs/2408.07877. arXiv:2408.07877. [15] D. Spina, J. Gwizdka, K. Ji, Y. Moshfeghi, J. Mostafa, T. Ruotsalo, M. Zhang, A. Ahmad, S. F. D.
Al Lawati, N. Boonprakong, N. Fernando, J. He, O. Hoeber, G. Jayawardena, B.-G. Lee, H. Liu, M. Pike, A. Pirmoradi, B. Nakisa, M. N. Rastgoo, F. D. Salim, F. Scott, S. Sun, H. Tang, D. Towey, M. L. Wilson, Report on the 3rd workshop on neurophysiological approaches for interactive information retrieval (neurophysiir 2025) at sigir chiir 2025, SIGIR Forum 59 (2025) 1–26. [16] U. Soni, S. Sreedharan, S. Kambhampati, Not all users are the same: Providing personalized explanations for sequential decision making problems, in: 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2021, pp. 6240–6247. URL: https://doi.org/10.1109/ IROS51168.2021.9636331. doi:10.1109/IROS51168.2021.9636331. [17] S. Milivojevic, M. Sobhani, N. Webb, Z. Madin, J. Ward, S. Yusuf, C. Baber, E. R. Hunt, Swift trust in mobile ad hoc human-robot teams, 2024. URL: https://doi.org/10.1145/3686038.3686057. doi:10.1145/3686038.3686057. [18] R. Paleja, M. Ghuy, N. R. Arachchige, R. Jensen, M. Gombolay, The utility of explainable ai in ad hoc human-machine teaming, 2021. URL: https://dl.acm.org/doi/10.5555/3540261.3540308, https://github.com/CORE-Robotics-Lab/Utility-of-Explainable-AI-NeurIPS2021. [19] M. R. Endsley, Supporting human-ai teams:transparency, explainability, and situation awareness, Computers in Human Behavior 140 (2023) 107574. URL: https://www.sciencedirect.com/science/ article/pii/S0747563222003946. doi:https://doi.org/10.1016/j.chb.2022.107574. [20] F. Shafer, J. P. Ginsberg, An overview of heart rate variability metrics and norms, Front Public Health 5 (2017) 258. doi:10.3389/fpubh.2017.00258, 2296-2565 Shafer, Fred Ginsberg, J P Journal Article Review Switzerland 2017/10/17 Front Public Health. 2017 Sep 28;5:258. doi: 10.3389/fpubh.2017.00258. eCollection 2017. [21] H. N. Green, T. Iqbal, Using physiological measures, gaze, and facial expressions to model human trust in a robot partner, 2025. URL: https://arxiv.org/abs/2504.05291. [22] A. Aygun, H. Ghasemzadeh, R. Jafari, Robust interbeat interval and heart rate variability estimation method from various morphological features using wearable sensors, IEEE J Biomed Health Inform 24 (2020) 2238–2250. doi:10.1109/jbhi.2019.2962627, 2168-2208 Aygun, Ayca Ghasemzadeh, Hassan Jafari, Roozbeh R01 EB028106/EB/NIBIB NIH HHS/United States Journal Article Research Support, N.I.H., Extramural Research Support, U.S. Gov’t, Non-P.H.S. United States 2020/01/04 IEEE J Biomed Health Inform. 2020 Aug;24(8):2238-2250. doi: 10.1109/JBHI.2019.2962627. Epub 2019 Dec 27. [23] M. N. Rastgoo, B. Nakisa, R. Andry, M. Frederic, , V. Chandran, Driver stress levels detection system using hyperparameter optimization, Journal of Intelligent Transportation Systems 28 (2024) 443–458. URL: https://doi.org/10.1080/15472450.2022.2140046. doi:10.1080/15472450. 2022.2140046, doi: 10.1080/15472450.2022.2140046. [24] B. Nakisa, M. N. Rastgoo, D. Tjondronegoro, V. Chandran, Evolutionary computation algorithms for feature selection of eeg-based emotion recognition using mobile sensors, Expert Systems with Applications 93 (2018) 143–155. URL: https://www.sciencedirect.com/science/article/pii/ S0957417417306747. doi:https://doi.org/10.1016/j.eswa.2017.09.062. [25] H. M. Khalid, L. W. Shiung, P. Nooralishahi, Z. Rasool, M. G. Helander, L. C. Kiong, C. Ai-vyrn, Exploring psycho-physiological correlates to trust:implications for human-robot-human interaction, Proceedings of the Human Factors and Ergonomics Society Annual Meeting 60 (2016) 697–701. URL: https://journals.sagepub.com/doi/abs/10.1177/1541931213601160. doi:10.1177/ 1541931213601160. [26] S. C. Kohn, E. J. de Visser, E. Wiese, Y. C. Lee, T. H. Shaw, Measurement of trust in automation: A narrative review and reference guide, Front Psychol 12 (2021) 604977. URL: https://doi.org/10.3389/ fpsyg.2021.604977. doi:10.3389/fpsyg.2021.604977, 1664-1078 Kohn, Spencer C de Visser, ezproxy-f.deakin.edu.au/10.1109/ACCESS.2023.3294569. doi:10.1109/ACCESS.2023.3294569. [41] J. D. Lee, K. A. See, Trust in automation: Designing for appropriate reliance, Human Factors 46 (2004) 50–80. URL: https://journals.sagepub.com/doi/abs/10.1518/hfes.46.1.50_30392. doi:10. 1518/hfes.46.1.50_30392. [42] A. Adadi, M. Berrada, Peeking inside the black-box: A survey on explainable artificial intelligence (xai), IEEE Access 6 (2018) 52138–52160. URL: https://doi.org/10.1109/ACCESS.2018.2870052. doi:10.1109/ACCESS.2018.2870052. [43] J. S. Park, C. Q. Zou, A. Shaw, B. M. Hill, C. Cai, M. R. Morris, R. Willer, P. Liang, M. S. Bernstein, Generative agent simulations of 1,000 people, 2024. URL: https://arxiv.org/abs/2411.10109.