AI-assisted decision-making is subject to AI model uncertainty. Prior works proposed to make this uncertainty explicit for increasing trust and transparency, but its behavioral role was rarely treated. This position paper argues, from a game-theoretic perspective, that human–AI decision support should be viewed as a repeated mechanism in which AI uncertainty functions as a strategic signal that shapes how users adopt reliance policies over time. We formalize a framework in which the interface specifies uncertainty signals, user response such as accepting versus verifying, and the resulting policy-shaping consequences. These repeated steps are used to characterize near-separating reliance regimes. A first pilot study conducted with 180 participants supports our proposition: Our game-theoretic mechanism increased verification and sharply reduced blind acceptance of wrong AI outputs. These initial results support treating human–AI interaction as a game-theoretic mechanism with uncertainty as a strategic signal, rather than a static model property or purely informational label.

Conformal prediction converts point predictions into set-valued predictions with coverage guarantees under exchangeability between calibration and deployment data. We study , where this requirement fails because labeled calibration is available only in a source space, while prediction sets are needed in a target space linked to the source through observations (e.g., paired modalities or sensor changes). We propose Transported Conformal Calibration (TCC): we transport labeled source calibration into the target space using the paired data, and then correct residual post-transport mismatch using only unlabeled target inputs. We instantiate this correction with two complementary methods: , which uses a label-free uncertainty surrogate to detect mismatch and adjust calibration conservatively, and , which reweights transported calibration toward the target domain for improved efficiency when weights are stable. We provide finite-sample target-domain coverage guarantees that adapt to an observable measure of mismatch. Across CIFAR-100-C, Tiny-ImageNet-C, and SEN12MS, we show reliable target-domain coverage transfer without labeled target calibration data, with label-free diagnostics that predict when correction is needed.

Large Language Models (LLMs) show growing promise in autonomous driving by reasoning over complex traffic scenarios to generate path plans. However, their tendencies toward overconfidence, and hallucinations raise critical safety concerns. We introduce SafePath, a modular framework that augments LLM-based path planning with formal safety guarantees using conformal prediction. SafePath operates in three stages. In the first stage, we use an LLM that generates a set of diverse candidate paths, exploring possible trajectories based on agent behaviors and environmental cues. In the second stage, SafePath filters out high-risk trajectories while guaranteeing that at least one safe option is included with a user-defined probability, through a multiple-choice question-answering formulation that integrates conformal prediction. In the final stage, our approach selects the path with the lowest expected collision risk when uncertainty is low or delegates control to a human when uncertainty is high. We theoretically prove that SafePath guarantees a safe trajectory with a user-defined probability, and we show how its human delegation rate can be tuned to balance autonomy and safety. Extensive experiments on nuScenes and Highway-env show that SafePath reduces planning uncertainty by 77% and collision rates by up to 70%, demonstrating effectiveness in making LLM-driven path planning more safer.

Active learning (AL) reduces annotation costs by selectively querying humans for uncertain data samples, thus improving data efficiency. In traditional AL, annotators simply label the data without gaining a deeper understanding of the reasoning of the model. However, incorporating model-generated explanations can enhance annotation quality by helping annotators interpret decisions and detect errors, a crucial need in safety-critical tasks, such as path prediction, where reliable annotations are vital for autonomous navigation. We propose Active Explainable Learning (AEXL), a novel AL framework that integrates natural language explanations via large language models (LLMs), enabling annotators to interpret predicted paths and identify errors more effectively. Our preliminary evaluations indicate that AEXL reaches 99% of the anticipated performance using only 60% of the data. The results of our preliminary user study show that natural language explanations boost annotator understanding, trust, and reduce cognitive load compared to approaches that lack explanatory guidance. Our work aims to foster future research on explanation-driven AL to improve model performance and human-AI collaboration.

AI-based mission-critical software exposes a blessing and a curse: its inherent statistical nature allows for flexibility in result quality, yet the mission-critical importance demands adherence to stringent constraints such as execution deadlines. This creates a space for trade-offs between the Quality of Result (QoR)-a metric that quantifies the quality of a computational outcome-and other application attributes like execution time and energy, particularly in real-time scenarios. Fluctuating resource constraints, such as data transfer to a remote server over unstable network connections, are prevalent in mobile and edge computing environments-encompassing use cases like Vehicle-to-Everything, drone swarms, or social-VR scenarios. We introduce a novel approach that enables software engineers to easily specify alternative AI service chains-sequences of AI services encapsulated in microservices aiming to achieve a predefined goal-with varying QoR and resource requirements. Our methodology facilitates dynamic optimization at runtime, which is automatically driven by the MARQ framework. Our evaluations show that MARQ can be used effectively for the dynamic selection of AI service chains in real-time while maintaining the required application constraints of mission-critical AI software. Notably, our approach achieves a 100x acceleration in service chain selection and an average 10% improvement in QoR compared to existing methods.

Effective communication is vital in emergency response scenarios, where clarity and speed can save lives. Traditional systems often struggle under the chaotic conditions of real-world emergencies, leading to breakdowns in communication and task management. In this paper we introduce the CLEAR (Coordinated Listening, Extraction, and Analysis for Emergency Response)-Command system, which leverages Large Language Models (LLMs) to enhance emergency communications. CLEARCommand automates the transcription, summarization, and task extraction from live radio communications of emergency first responders using the OpenAI Whisper API for transcription and Chat-GPT 4 for summarization and task extraction. We decided for ChatGPT 4 after conducting an expert pre-study that showed it to be the most accurate LLM in terms of task extraction for our case. To evaluate our system, we conducted a user study with 13 participants. Our results show that CLEAR-Command significantly outperforms traditional radio communication in terms of clarity, trust, and correctness of task extraction. The link to a live demo website of our system is https://clear- command. vercel.app. The video demonstrating our system can be found on https://youtu.be/ ZF3HMMUEq9o. All project details are presented in our Gitlab page https://gitlab. com/achref.d/clear-command.

Deep learning models are pivotal in enhancing driver assistance systems and improving environmental perception. However, the tendency of neural networks towards overconfident predictions poses a risk of inaccurate predictions, potentially compromising driver safety in adverse conditions. To mitigate this issue, we introduce AR-CP, an uncertainty-aware framework designed to augment driver perception in scenarios characterized by adverse weather and insufficient lighting, through the integration of conformal prediction and augmented reality (AR). Our framework initiates with a conformal prediction step that produces an uncertainty-aware prediction set including potential object classes at a predefined probability level. Subsequently, AR is used to provide a simplified and informative visualization of the closest common parent class of the classes in the prediction set, thereby reducing the likelihood of misinformation. We provide a principled formulation and theoretical analysis of our framework. We evaluate AR-CP on the ROAD dataset, a large dataset containing different difficult situations that induce high uncertainty during prediction time. The results show that our framework outperforms state-of-the-art approaches in providing smaller prediction sets while holding the theoretical guarantees, ensuring an uncertainty-aware prediction, and reducing user confusion. We conduct an immersive user study with 15 participants to investigate the effects of our concept on the quality of perception, situation awareness, and mental load of participants. The results show that our concept facilitates a safer driving experience while holding the mental load low and the situation awareness high.

Recent advancements in deep learning have significantly enhanced the development of efficient models for multi-modal path prediction within urban environments, offering approaches to navigate complex environments accurately. Despite their performance, models grounded in deep learning techniques frequently encounter challenges related to interpretability. This limitation not only hampers their practical application but also complicates the process of diagnosing and rectifying errors within these systems, which is a critical factor for ensuring reliability and safety in real-world deployments. In this paper we propose NeSyMoF, a Neuro-Symbolic model for Motion Forecasting, to address this critical gap by combining the predictive power of deep neural networks with the interpretable logic inherent in symbolic reasoning. Data processing in NeSyMoF involves extracting pertinent features from the agent's environment and channeling them into a neuro-symbolic reasoning module. The neuro-symbolic reasoning module generates first-order logic rules that describe and condition the path prediction process, thereby providing clear explanations and intentions behind the forecasts of the model. We evaluate our model with the Argoverse benchmark for path forecasting, as it includes challenging driving situations, necessary to extensively evaluate our model. The results of our evaluation show that NeSyMoF outperforms state-of-the-art interpretable models for single-mode predictions while providing logic-based explanations for its forecasts, that articulate the reasoning behind predictions, making NeSyMoF more adapted for human-centric applications.

We introduce Knowledge-Refined Prediction Sets (KRPS), a novel approach that performs semantically-aware uncertainty quantification for multitask-based autonomous perception in urban environments. KRPS extends conformal prediction (CP) to ensure 2 properties not typically addressed by CP frameworks: semantic label consistency and true label coverage, across multiple perception tasks. We elucidate the capability of KRPS through high-level classification tasks crucial for semantically-aware autonomous perception in urban environments, including agent classification, agent location classification, and agent action classification. In a theoretical analysis, we introduce the concept of semantic label consistency among tasks and prove the semantic consistency and marginal coverage properties of the produced sets by KRPS. The results of our evaluation on the ROAD dataset and the Waymo/ROAD++ dataset show that KRPS outperforms state-of-the-art CP methods in reducing uncertainty by up to 80% and increasing the semantic consistency by up to 30%, while maintaining the coverage guarantees.

The recent advances in deep learning have paved the way for autonomous vehicles (AVs) to take charge of more complex tasks in the navigation process. However, predictions of deep learning models are subject to different types of uncertainty that may put the user and the surrounding environment in danger. In this paper, we investigate the effects that AR-based visualizations of 3 types of uncertainties in deep learning modules for path planning in AVs may have on drivers. The uncertainty types of the deep learning models that we consider are: the waypoint uncertainty, the situation uncertainty, and the path uncertainty. We propose 3 concepts to visualize the 3 uncertainty types on a Windshield display. We evaluate our AR-based concepts with a user study $(\mathrm{N}=20)$ using a VR-based immersive environment, to ensure the security of the participants. The results of our evaluation reveal that the absence of uncertainty visualization leads to lower driver engagement. More importantly, the combination of situation uncertainty and path uncertainty visualizations leads to higher driver engagement, and higher trust in the automated vehicle, while inducing an acceptable mental load for the drive.

The annotation of 3D point cloud datasets is an expensive and tedious task. To optimize the annotation process, recent works have proposed the use of environments with higher levels of immersion in combination with different types of visual aids. However, two problems remain unresolved. First, the proposed environments limit the user to a unique level of immersion and a fixed hardware setup. Second, their design overlooks the interaction effects between the level of immersion and the visual aids on the quality of the annotation process. To address these issues, we propose PointCloudLab, an environment for 3D point cloud annotation that allows the use of different levels of immersion that work in combination with visual aids. Using PointCloudLab, we conducted a controlled experiment (N=20) to investigate the effects of levels of immersion and visual aids on the annotation process. Our findings reveal that higher levels of immersion combined with object-based visual aids lead to a faster and more accurate annotation. Furthermore, we found significant interaction effects between the levels of immersion and the visual aids on the accuracy of the annotation.

The recent advances in smart city infrastructure have provided support for a higher adoption of surveillance cameras as a mainstream crime prevention measure. However, a consequent massive deployment raises concerns about privacy issues among citizens. In this paper, we present VR-Surv, a VR-based privacy aware surveillance system for large scale urban environments. Our concept is based on conveying the semantics of the scene uniquely, without revealing the identity of the individuals or the contextual details that might violate the privacy of the entities present in the surveillance area. For this, we create a virtual replica of the areas of interest, in real-time, through the combination of procedurally generated environments and markerless motion capture models. The results of our preliminary evaluation revealed that our system successfully conceals privacy-sensitive data, while preserving the semantics of the scene. Furthermore, participants in our user study expressed higher acceptance to being surveilled through the proposed system.

We present Immersive-Labeler, an environment for the annotation of large-scale 3D point cloud scenes of urban environments. Our concept is based on the full immersion of the user in a VR-based environment that represents the 3D point cloud scene while offering adapted visual aids and intuitive interaction and navigation modalities. Through a user-centric design, we aim to improve the annotation experience and thus reduce its costs. For the preliminary evaluation of our environment, we conduct a user study (N=20) to quantify the effect of higher levels of immersion in combination with the visual aids we implemented on the annotation process. Our findings reveal that higher levels of immersion combined with object-based visual aids lead to a faster and more engaging annotation process.

The reconstruction of city landmarks is central to creating recognizable virtual environments representing real cities. Despite the recent advances, it is still not clear what level of detail (LOD) to adopt when reconstructing those landmarks for their correct recognition, and if particular architectural styles represent specific challenges in this respect. In this paper, we investigate the effect of LOD on landmark recognition, generally, and on some architectural styles, specifically. The results of our user study show that higher LOD lead to a better landmark identification. Particularly, Neoclassical-style buildings need more details to be individually distinguished from similar ones.

Augmented reality see-through applications rely mostly on machine learning models to detect and localize occluded objects. In this case, the user is usually presented the result with the highest probability without taking into account the uncertainty of the model. However, the uncertainty plays a vital role when considering applications where a critical decision-making process relies heavily on the predictions of the model, such as in the case where occluded cars are shown to a driver. In this work, we conduct an investigation of the effects of communicating the uncertainty of machine learning models to users in AR-based see-through applications. Through a controlled user study, we compare three visualization modes: no visualization, most probable output, and probability distribution. The results of our evaluation reveal that when considering the visualizations, each of them lead to comparable results in terms of speed and accuracy of the decision-making process. A relevant finding is that participants considered uncertainty as a substantial part of the output of machine learning models and needs to be delivered with the results. An additional important conclusion is that the preference of users over a specific visualization is strongly dependent on the particular use case.