MobiCom 2026 Papers

On-device mobile Large Language Model (LLM) inference is gaining significant attention. However, mobile devices operate in highly dynamic multitasking environments where users frequently switch between applications. This creates memory pressure, forcing LLM memory—model weights and KV cache—to be evicted by the operating system. When a new inference request arrives, the inference system must restore the evicted memory through slow storage reads or recompute the entire KV cache, severely degrading responsiveness. To address this, we present mzCache, an on-device LLM inference system with specialized memory management for multitasking environments. Under unpredictable memory pressure, mzCache elastically evicts LLM memory and leverages the unified memory of mobile SoCs to enable zero-wait inference on the GPU with concurrent CPU-side restoration. mzCache realizes this through restoration-oriented memory management: LLM memory is partitioned into fine-grained shared buffers to enable partial eviction and restoration with concurrent cross-processor access, while hybrid swap and backward-out eviction policies ensure low-latency restoration from any eviction state. Implemented on llama.cpp and deployed as an Android application, mzCache achieves 2.1–5.5× reduction in Time-to-First-Token compared to storage-backed partial offload and demonstrates its effectiveness in real multitasking scenarios.

Real-time object detection is essential for mobile augmented reality (AR) applications; however, modern detectors based on Vision Transformers (ViTs) incur significant computational overhead. Although edge offloading can alleviate the limited compute resources on mobile devices, the inference latency of ViT-based detection on edge servers remains a major bottleneck, causing detection results to arrive after the camera viewpoint has changed and degrading AR experiences. To address this problem, we propose DeltaAR, an edge-assisted AR object detection system that leverages AR geometry (depth maps and 6-DoF camera poses) to reduce the cost of ViT-based detection. Different from existing AR systems, DeltaAR uses sensory data available on modern AR devices to accurately identify newly exposed regions between consecutive frames. Based on this information, we design a region-aware ViT that reduces computation by processing tokens required for detecting objects in newly exposed regions and skipping redundant tokens from old regions. To preserve detection accuracy, we further introduce a lightweight token selector that retains a small set of old-region tokens needed for cross-boundary objects and contextual information. Experiments show that DeltaAR significantly reduces edge inference cost and latency while maintaining high detection accuracy for real-time AR applications.

Voice calls are a critical communication mode of interpersonal interaction. However, one of its most significant limitations is susceptibility to ambient noise interference. Current approaches to mitigate noise typically rely on advanced algorithmic processing or supplementary modal signals to enhance the original speech. Nevertheless, when environmental noise reduces the signal-to-noise ratio (SNR) of the original speech below a certain threshold, these methods become entirely ineffective. In this paper, we propose AccVoice, a novel approach that relies solely on the smartphone's builtin accelerometer to capture facial vibration signals, enabling a general speech reconstruction system for voice calls. We first propose a triaxial fusion method to eliminate variations in accelerometer data caused by user-induced micro-motion noise and enhance the feature representation of the signal. Then, we introduce for the first time, to the best of our knowledge, the concept of removing voice timbre to enable cross-user general speech reconstruction capabilities of accelerometer signals. Finally, by employing a large-scale pre-trained speech model combined with a spectrum-separation-based reconstruction approach, we achieve high-quality speech generation. Besides, we use the one-shot sample to re-infuse the timbre into the reconstructed speech to maintain the authenticity of voice calls. We conduct extensive experiments in a cross-user scenario. The results demonstrate that AccVoice achieves high-quality speech reconstruction across different impact factors.

Multimodal target tracking estimates a moving target's trajectory by fusing heterogeneous per-anchor feature streams from sensing modalities. Existing tracking pipelines are usually built for a fixed feature set and anchor layout. They become brittle when streams drop out, new feature families are introduced, or the deployment changes. We present PATHS, a deployment-aware token interface for plug-and-play tracking. PATHS converts per-anchor feature streams into a shared, fixed-width token interface, allowing a unified tracking model to operate as the available feature mix changes. To add a feature family, PATHS trains only a new encoder with a frozen reference decoder, aligning its tokens to the shared interface without paired cross-modal data. At runtime, set attention fuses any available tokens, and a shared decoder reconstructs the trajectory. A reference-frame mapping adapts PATHS to different spatial scales without retraining. Across seven feature families in simulation and three real deployments, PATHS achieves mean tracking errors of 0.13–0.49 m and reduces median error by up to 33% relative to fixed-input baselines under stream quality fluctuations.

Human body point clouds are a versatile representation for AI-enabled human sensing. However, existing methods using LiDAR, radar, and depth cameras suffer from inherent drawbacks in high cost, sparse reconstruction, and privacy concerns, etc. In this paper, we exploit low-cost thermal arrays and present TAP3D, the first system to reconstruct 3D human point clouds from body heat signatures, offering significant advantages in cost, density, human sensitivity, and privacy. To overcome major challenges in depth estimation, thermal interference, and multi-person separation, we propose a novel physics-informed design, which integrates a forward thermal physics model with two distinct modules: multi-primitive estimation for self-supervised joint recovery of depth and other thermal properties, and geometric perspective fusion for suppressing interference and disentangling multiple people. We implement TAP3D using a single commodity thermal array sensor and build a large-scale dataset (160K samples, 8 environments, 11 users) for evaluation. TAP3D achieves remarkable accuracy for dense point cloud generation, enabling downstream tasks like fall detection (91.46%), indoor tracking (21.86 cm MAE), and human mesh recovery (4.87 cm error). By transforming body heat into point clouds for the first time, TAP3D pioneers a new paradigm for privacy-first, fully passive human sensing for many applications. TAP3D is open-sourced at https://github.com/aiot-lab/TAP3D.

We present tinyGist, an efficient decentralized federated learning framework designed for tiny microcontrollers. Instead of treating all model parameters equally, each device exchanges a portion of them, effectively sharing a “gist” of the information. Our major contributions are: (1) segment model parameters based on their importance, while identifying the most impactful updates, (2) share the gist probabilistically, maintaining model diversity while ensuring fast knowledge propagation across peers, (3) update models based on the information merit received from other neighbors, and (4) adaptive learning. We implement tinyGist across various simulation experiments, large-scale realistic emulation with hundreds of devices, and real deployment on homogeneous and heterogeneous physical microcontroller clusters. Evaluated on eight machine learning models and 15 tasks, tinyGist outperforms existing baselines in convergence speed by 1.35–3.3× and accuracy by up to 17%, while reducing communication overhead by ≈3× compared to the full-model baselines. We further validate our lightweight importance proxy against gradient-, Fisher-, and Hessian-based measures, and characterize how tinyGist interacts with differential privacy and secure aggregation.

Underwater optical sensing is severely degraded by backscattered light from suspended particles. While polarization-based image enhancement methods can suppress this backscattering effect, existing methods are either costly, or require sequential image captures that do not handle dynamic scenes. Recently, learning-based enhancement methods have shown promising results, but they depend on large paired datasets that are difficult to obtain underwater and introduce noise or artifacts due to domain gaps.

Rapidly growing low-Earth-orbit (LEO) constellations increasingly operate in the spectrum shared with radio astronomy services (RAS), creating escalating interference risks for sensitive scientific observations. Existing mitigation mechanisms rely on reactive beam steering, or avoidance near observatories but fail to account for aggregate sidelobe emissions—leading to residual interference and substantial, unnecessary capacity loss. We present SkyShare, a constellation-wide sky-sharing system that enables predictive, interference-aware spot beam scheduling to protect radio astronomy while preserving network coverage. SkyShare integrates high-fidelity orbital prediction with International Telecommunication Union (ITU)-compliant Equivalent Power Flux Density (EPFD) modeling, and realtime observatory data via Operational Data Sharing (ODS) to jointly optimize beam-cell assignments over observation windows. To make constellation-scale coordination tractable, we introduce a concept of EPFD-budgeted Region-of-Interest (RoI) that bounds residual sidelobe interference while confining optimization to a minimal, provably sufficient set of cells. Building on RoI, we formulate LEO-RAS coexistence as a scalable scheduling problem and design SkySched, a flow-based algorithm that is optimal in special cases and yields scalable near-optimal solutions in the general NP-hard setting. SkyShare operates entirely in the control plane and requires no satellite hardware changes. Using real Starlink constellation geometries, we evaluate SkyShare across 25 single-dish, Ku-band RAS sites worldwide. Compared to Star-link boresight avoidance, SkyShare reduces unserved cells by up to 90.68% while remaining within EPFD limits.

We present mmFHE, the first system that executes the entire cloud-side mmWave sensing pipeline including the DSP and ML inference under fully homomorphic encryption (FHE). mmFHE encrypts range profiles on an edge device after lightweight plaintext preprocessing and executes the entire mmWave signal-processing and ML inference pipeline homomorphically on a semi-honest cloud that operates exclusively on ciphertexts. At the core of mmFHE is a library of seven composable, data-oblivious FHE kernels that replace standard DSP routines with fixed arithmetic circuits for different application-specific pipelines. We demonstrate this approach on two representative tasks: vital-sign monitoring and gesture recognition. We formally prove two cryptographic guarantees for any pipeline assembled from this library: input privacy and data obliviousness. These guarantees effectively neutralize various supervised and unsupervised privacy attacks on raw data, including re-identification and data-dependent privacy leakage. Evaluation on three public radar datasets shows that encryption introduces negligible error versus the plaintext pipeline, with 84.5% gesture accuracy (vs. 84.7%). End-to-end cloud GPU latency is 1.21 s per 10 s vital-sign window and 5.76 s per 3 s gesture window. These results establish the initial feasibility of end-to-end mmWave sensing under FHE on commodity hardware.

Bowel sounds are a clinically valuable non-invasive indicator of gastrointestinal motility, yet current systems remain tethered to dedicated auscultation hardware, support only coarse event detection, and do not provide actionable diagnostic interpretation. We present Ventrilo, an end-to-end bowel-sound diagnostic system that runs on both electronic stethoscopes and commodity smartphones. First, a cross-device acquisition pipeline couples active noise cancellation with learned spectral transfer, suppressing ambient interference and aligning smartphone-stethoscope frequency responses to within 2 dB. Second, Listen-BowelNet performs joint segmentation and classification of bowel-sound events, achieving 0.984 Event F1 across three clinically defined event types. Third, an LLM-driven diagnostic engine combines clinical-guideline retrieval with structured evidence reasoning to reach 0.934 diagnostic accuracy at a 0.013 false-alarm rate. The engine outputs structured clinical reports covering four GI diagnostic categories with risk-stratified conclusions, evidence traces, and actionable recommendations. We evaluate Ventrilo on 1,678 recordings from 712 subjects across two medical centres and home environments, using five smartphone models and three stethoscope models. In clinical deployments across acute, chronic, and post-discharge settings, Ventrilo produces actionable alerts from 3.15 hours ahead in-clinic to 1.75 days ahead at home, relative to corresponding physician decisions.

This paper reports our three-year experience building 5GAudit, an LLM-based agentic framework for discovering vulnerabilities in 5G protocol interactions. We review the evolution of 5GAudit through successive redesigns, each driven by fundamental deficiencies uncovered in real deployment. We refine 5GAudit to enable it to conduct agentic cross-protocol/domain/system analysis, perform automatic verifier-in-the-loop validations of vulnerabilities, and go beyond predefined threat models by leveraging CWE (Common Weakness Enumeration) theory to enumerate new vulnerability patterns. Our system successfully discovers 42 0-day vulnerabilities of 5G protocol interactions in open-source 5G software, 20 of which have been assigned CVE numbers to date. We demonstrate 5GAudit as an example to share our lessons learned on applying LLM-based agents to 5G-specific security analysis with the community.

Programmable reflective metasurfaces enable flexible wave manipulation, but conventional resonance-based designs suffer from high cost, narrow bandwidth, complex calibration, and limited deployability at high frequencies. We present MetaX, a low-cost mechanically programmable metasurface that controls wavefronts by continuously deforming a reflective membrane with only 64 micro-motors. Geometric path-length modulation enables broadband, calibration-free operation across different wave domains. We further develop a differentiable model that maps actuator commands to 3D surface deformation and wave responses, and use gradient-based inverse design to generate mechanical code-books for wide-angle single-beam steering and concurrent multi-beam generation. Experiments with a 64-actuator prototype demonstrate stable RF beam steering over 77-81 GHz with a maximum frequency-induced pointing deviation of only 2.5° and support for four concurrent beams with at most 1.07 dB average inter-user power imbalance. The experiments further demonstrate that adjusting the motor spacing enables the MetaX prototype to support beam steering across other frequency bands within 20-85 GHz. The complete prototype costs approximately $81, demonstrating low-cost, broadband, and multi-beam wavefront control for high frequencies.

Multi-user extended reality (XR) applications require accurate spatial synchronization across devices to support real-time interactions such as content sharing. Existing approaches primarily rely on centralized infrastructure and exhaustive cross-user visual feature matching, resulting in substantial communication and computational overhead. In this paper, we present ProxLink, a fully decentralized synchronization framework that eliminates the need for a global map server and enables lightweight peer-to-peer spatial coordination without centralized control. To reduce bandwidth and computation overhead, ProxLink introduces a multi-stage filtering pipeline before visual feature matching, combining RF-IMU distance estimation to identify nearby users, overlapping map-point density prediction to estimate potential view overlap, and graph-based structured matching to narrow the map-point candidate space before final feature-based relative pose estimation. Implemented on commodity XR devices without server support, ProxLink reduces bandwidth consumption by 10-15× and initial pairwise synchronization latency by over 3× compared with Map++, the state-of-the-art centralized baseline, and over 2× compared with FreeAR (a decentralized scheme), while keeping bandwidth below 6 Mbps for 15 users and achieving comparable or better synchronization accuracy.

As on-device large language model (LLM) applications like mobile agents become increasingly popular, long-context inference has become a key challenge. The key-value (KV) cache is the foundation of efficient autoregressive decoding, but its memory footprint grows rapidly with context length, often exceeding the memory limits of mobile devices. Existing KV cache management methods, such as compression and storage offloading, often lead to performance degradation or unpredictable stall latency. We find that in the prevalent CPU/GPU-decoding mobile pipelines, the NPU remains largely idle during decoding and can recompute missing KV cache more efficiently than storage offloading. We therefore propose NPUGen, whose KV recomputation engine generates missing KV cache on the idle NPU via a prefill-like forward pass, avoiding unstable storage I/O, CPU intervention, and unnecessary data movement. NPUGen further introduces a hidden-state anchor mechanism that eliminates redundant prefix recomputation, and turns KV recomputation into a schedulable background service to cope with resource contention and guarantee foreground QoS. Experiments on smartphones with Qualcomm SoCs across multiple long-context tasks show that NPUGen significantly reduces decoding latency and fluctuations, providing fast and stable long-context inference on resource-constrained mobile devices.

Wi-Fi has emerged as a promising medium for device-free human sensing, yet its capability in fine-grained sensing tasks remains rather limited. In particular, existing solutions to human pose estimation (HPE) demand extensive hardware elevations (e.g., multiple antennas or transceivers) directly upon Wi-Fi home deployments, typically involving very few APs and clients. In this paper, we present MetaVision, a system that utilizes programmable metasurfaces as external add-on to create multiple virtual viewpoints, thus substantially extending the spatial diversity of a single pair of AP and client. MetaVision organizes its metasurfaces into a master-slave hierarchy to tackle inter-surfaces synchronization. It further separates individual signals from these homogeneous metasurfaces, exploiting a novel temporal encoding scheme that effectively decouples superimposed signals into independent signal spectrograms. Fusing these multi-view spectral representations, MetaVision employs a neural network to model inter-joint dependencies and to accordingly reconstruct 3D human poses. Extensive experiments demonstrate that MetaVision achieves high-resolution 3D HPE and reduces error by 61% compared with baseline approaches.

Dense 3D depth perception fails under smoke, fog, and darkness because optical sensors cannot penetrate airborne particulates. mmWave radar remains usable and measures range accurately under these conditions, but its small aperture limits angular resolution. We present GRADE, which grounds a pretrained generative prior in single-frame radar geometry to estimate high-fidelity metric depth. GRADE first maps raw 4D radar spectra to coarse metric depth. A latent diffusion backbone then recovers structural detail while conditioning every denoising step on this estimate. A pixel-space adapter uses residual camera cues when available and is trained across clear, smoke-degraded, and occluded inputs so the full output approaches the radar-conditioned path as visibility degrades. Trained and evaluated on ~95K frames across 12 buildings with real smoke, GRADE achieves an MAE of 0.303 m in clear scenes and 0.313 m under smoke, outperforming existing baselines. Code and datasets are available at https://phi-lab-rice.github.io/GRADE.

Smart glasses are the next mobile platform. Unlike phones, glasses operate as sensing-first devices that route data through a multi-hop wireless pipeline: a constrained glass to companion puck or a phone short-range edge link and cloud AI backends. This architecture exposes an entirely new class of wireless and mobile system problems that the community has not yet confronted. We present the first in-depth measurement study across several glasses and core XR and AI applications. Our findings open up rich problems for the community in rethinking glass-phone-cloud architecture and reimagining how smartphones can serve as efficient edge servers for the next generation of wearable computing.