Millimeter-wave (mmWave) radar presents a compelling solution for privacy-preserving Human Activity Recognition (HAR), yet bridging the gap between abstract radio signals and high-level semantics remains a fundamental challenge. The intrinsic sparsity, noise, and environmental sensitivity of radar point clouds often hinder the extraction of robust features suitable for cross-domain deployment. To address these inherent signal limitations, we propose a unified paradigm that bridges physics-based sensing and the semantic embedding space of Multimodal Large Language Models (MLLMs). Specifically, we design an energy-aware point cloud generation pipeline coupled with a hierarchical spatio-temporal encoder to capture continuous kinematic dynamics from sparse radar returns. Furthermore, to bridge these physical features with semantic knowledge, we introduce a cross-modal projector and a decoupled two-stage training strategy, enabling the direct projection of wireless signals into the latent space of MLLMs. We evaluate mmLM on the large-scale XRF55 dataset. Extensive experiments demonstrate that mmLM establishes a new state-of-the-art across both in-domain and cross-domain settings. Specifically, the system achieves an accuracy of 93.56
Standardized self-report scales provide efficient assessment, but their abstract and repetitive items often offer limited contextual support for respondents. Contextualized narrative assessment can make assessment tasks more concrete and engaging, yet direct scale-to-narrative transformation may increase reading, interaction, and workload. This paper presents SituAssessor, an LLM-assisted workflow for user-centered narrative assessment task design. The workflow connects narrative task generation, simulated response-based screening, statistical task reduction, and task reorganization to balance construct coverage with assessment manageability. We evaluated SituAssessor using the Motivated Strategies for Learning Questionnaire (MSLQ) as a reference case. The workflow generated a full 81-task narrative version and then produced a reduced 49-task version. A domain-evaluator quality review showed that the generated task pool provided an acceptable content-quality foundation. In a 40-participant within-subjects study comparing the traditional scale, the full narrative version, and the reduced narrative version, the reduced version preserved substantial score correspondence with both narrative and scale-based formats. It also showed lower scores on several workload dimensions and a shorter completion time than the full narrative assessment, while receiving higher engagement-related user-experience ratings than the traditional scale. These findings suggest that LLM-assisted assessment workflows should jointly support contextualized task construction, task manageability, and engaging assessment experiences.
Accurate liquid level monitoring is crucial in many industrial processes. While non-contact radar-based level meters overcome issues like corrosion encountered in contact-based methods, they require prior knowledge of the distance between the radar and the container’s bottom, which is often unavailable in practice. To address this, we propose DualMeasure, a novel calibration-free system utilizing a standard FMCW radar for both liquid dielectric constant and level measurement. It captures the dual echoes from the air-liquid and liquid-bottom interfaces. By jointly analyzing the time delay and phase difference ratio between these echoes, the system estimates the liquid’s dielectric constant in real-time and directly calculates the liquid level without any prior knowledge of container dimensions. Experimental results demonstrate that the system achieves millimeter-level precision under various conditions, validating its practical applicability for direct and reliable liquid parameter sensing.
Data sparsity is a central challenge in pervasive computing and human–machine interactive systems, where observations arise unevenly across users, devices, algorithms, and environments. Across recommender interfaces, healthcare monitoring, smart-city mobility, and language-enabled interaction, sparsity can reflect incomplete sensing, constrained exposure, intermittent participation, privacy limits, and heterogeneous event generation. These mechanisms can produce missing or weakly supported instances as well as observable structures such as excess zeros and long-tail frequency distributions; conflating mechanisms with their manifestations can lead to mismatched modelling and evaluation. This survey synthesises 134 studies and develops a source-to-structure diagnostic framework. It distinguishes two mechanism-oriented perspectives—data collection oriented sparsity and user–system interaction oriented sparsity—from a distributional structure oriented perspective that records observable zero-inflated and long-tail patterns without assuming their causes. Six operational forms are connected through a support-based chain from mechanism to observational support, observable structure, and intervention. The roadmap further distinguishes evidence acquisition, bias correction, robust learning under fixed sparsity, and decentralised utilisation, and specifies cross-domain transfer conditions, mechanism- and structure-aware evaluation, and human-centred trade-offs for conditional method selection.
In pervasive computing environments, virtual reality (VR) systems equipped with eye-tracking capabilities enable gaze-based content understanding, defined as generating descriptions for the region the user is looking at, which is crucial for accurately capturing user attention and intention. With the rapid advancement of Large Vision-Language Models (LVLMs), their powerful visual understanding capabilities present new opportunities for these tasks. However, existing segmentation-based methods, which generally rely on precomputed segmentation, face critical limitations: their results typically cover entire objects, thereby losing fine-grained gaze details, while their analysis granularity remains rigid and unable to adapt to varying user intent. To address these challenges, this paper proposes Gaze-MIRL, a multi-granularity instruction-based reinforcement learning method designed to achieve controllable and precise gaze-based content understanding. Specifically, a multi-granularity instruction fine-tuning approach is introduced, which constructs multi-level instruction templates to enable LVLMs to autonomously adjust visual perception granularity based on user instructions, generating descriptions across different levels (part-level, object-level, etc.). Simultaneously, a group relative policy optimization (GRPO) method with location-based rewards is proposed to drive LVLMs to focus on the image regions indicated by the input gaze coordinates during description generation. Finally, to support controlled evaluation of this task, we contribute MGaze, an image-based proxy benchmark with synthetically generated gaze points and multi-granularity annotations. Extensive experiments demonstrate that Gaze-MIRL outperforms existing baseline methods in both gaze-based description accuracy and controllability, suggesting its potential for future VR interaction studies.
