MobiCom 2026 Papers

Optically active substances such as sugars and amino acids are critical quality indicators in biomanufacturing and food safety, yet they typically coexist as multi-component mixtures. Existing monitoring techniques either require expensive instruments (e.g., HPLC) with per-sample costs and turnaround times ill-suited for online control, or rely on optical methods that cannot resolve individual component concentrations from a mixture. We present Polarix, a training-free polarimetric platform built from $100 in hardware that exploits a fundamental physical property: the specific optical rotation of each optically active substance varies with temperature at a distinct, substance-dependent rate. By sweeping the sample temperature across a bounded 15°C window, Polarix transforms the underdetermined problem into a solvable system, mathematically disentangling each component's concentration. Its optical design improves robustness against real-world disturbances including turbidity, flow, and mechanical drift. We validate Polarix on three real-world tasks: amino-acid fermentation tracking, trace byproduct detection, and honey adulteration screening against rice syrup and HFCS-55. Polarix resolves individual component concentrations with a mean relative error of 12.1%.

The complexity, interdependence, and rapid evolution of 3GPP specifications present fundamental challenges for ensuring the security of modern cellular networks. Manual reviews and existing automated approaches, which often depend on rule-based parsing or small sets of manually crafted security requirements, fail to capture deep semantic dependencies, cross-sentence/clause relationships, and evolving specification behaviors. In this work, we present CellSecInspector, an automated framework for security analysis of 3GPP specifications. CellSecInspector extracts structured state-condition-action (SCA) representations, models mobile network procedures with comprehensive function chains, systematically validates them against 9 foundational security properties under 4 adversarial scenarios, and automatically generates test cases. This end-to-end approach enables the automated discovery of vulnerabilities without relying on manually predefined security requirements or rules. Applying CellSecInspector to the well-studied 5G and 4G NAS and RRC specifications and selected sections of TS 23.501 and TS 24.229, it discovers 43 vulnerabilities, 7 of which are previously unreported. Our findings show that CellSecInspector is a scalable, adaptive, and effective solution to assess 3GPP specifications for safeguarding operational and next-generation cellular networks.

The vision of ambient LoRa backscatter is to achieve long-range, pervasive connectivity by leveraging existing LoRa signals. However, requiring transmitter-side coordination limits widespread deployment because LoRa infrastructure needs to generate tag-specific traffic or allocate airtime and spectrum resources, thereby disrupting its primary services. We present AmbientRider, an infrastructure-independent LoRa backscatter system that operates over the signals without any transmitter-side coordination. Our key insight is that commodity LoRa chirps contain exploitable timing and phase structure that enables blind synchronization, channel adaptation, and interference cancellation, thereby supporting infrastructure-independent LoRa backscatter. We implement AmbientRider with off-the-shelf components and evaluate it over uncoordinated commercial LoRa transmitters. The system achieves up to 19.0 kbps throughput with the wake-up sensitivity of -24 dBm when uncooperative ambient gateways transmit across the entire 2.4 GHz ISM band with randomly varying spreading factors (SFs) and channels.

Mixture-of-Experts (MoE) inference on memory-constrained edge platforms is often dominated by expert-weight movement rather than computation. Although each token activates only a small subset of experts, routing is input-dependent, so all experts must remain accessible and memory movement dominates latency. Existing approaches such as uniform pruning and expert dropping overlook two properties of MoE execution: expert combinations are highly skewed in frequency, and experts contribute asymmetrically within each combination. We present CRISP, a combination-aware framework that co-designs pruning and runtime scheduling for efficient MoE serving on memory-constrained edge platforms. CRISP reorders neurons by importance so one stored expert can support multiple execution widths, assigns differentiated execution to hot, warm, and cold combinations based on frequency and expert role, and exploits heterogeneous widths to overlap expert loading with computation. Across Mixtral-8x7B and DeepSeek-MoE-16B, CRISP improves throughput by up to 2.1× at the same memory budget while limiting accuracy loss to 3.6% relative to full-width inference. Our results highlight that efficient edge MoE serving requires jointly optimizing which expert capacity is preserved and how expert execution is scheduled.

High-quality large language model inference is increasingly used on mobile edge, from satellites to smart homes, where limited connectivity or privacy requirements motivate local model inference. In many such settings, token generation is inherently compute-limited and added latency may be acceptable, but output quality, memory feasibility and crash-free execution are essential. However, on today's mobile edge with unified-memory architectures, running a large model near the device limit can trigger non-deterministic out-of-memory failures. This is because deep learning frameworks retain freed buffers inside caching pools, thus continuing to occupy shared DRAM, which the OS cannot reclaim. During inference, fragmentation and conservative reuse can produce growing memory usage, eventually exceeding the memory availability. In this paper, we present ExoMem, an OS-governed middleware that restores OS authority over physical pages without sacrificing framework compatibility. ExoMem keeps model topology and operators in the framework, while anchoring weights in OS-managed mappings, and introduces a dormant state, whose virtual addresses remain stable for correctness but whose physical pages are immediately reclaimable when needed, leveraging standard OS page caching as an implicit multi-tier hierarchy. Implemented as a plug-and-play extension, ExoMem enables stable local inference for substantially larger LLMs on mobile edge, including 72B-class models on 64GB NVIDIA Jetson devices, while preserving headroom for concurrent workloads.

On-device LLM inference is attractive for privacy and responsiveness, but remains challenging on mobile and embedded devices because model weights far exceed available DRAM. Prior systems exploit activation sparsity and offload weights to SSD or flash storage, but face a fundamental systems tradeoff: accurate sparse execution decisions require the latest context, whereas efficient computation-I/O overlap requires early prediction. As a result, existing designs either serialize execution or incur redundant weight fetches, extra computation, and large cache overheads. We present LeanStream, a streaming speculate-and-refine framework for efficient on-device LLM inference. LeanStream progressively refines computation, loading, and cache-retention priorities using partial GPU results, enabling fine-grained overlap between GPU execution and storage I/O. We implement LeanStream on both mobile and embedded platforms. Compared with prior on-device LLM inference systems, LeanStream reduces memory usage by 4.8×-7.5× at the best throughput achieved by prior work, while further improving token generation throughput by 1.6×-2.1×.

Large-kernel Convolutional Neural Networks (CNNs) deliver remarkable performance in vision tasks by significantly expanding receptive fields, yet their quadratic parameter growth critically impedes storage-efficient edge deployment. While existing efficient architectures adopt parameter-efficient depthwise separable convolution backbones that leverage techniques like low-rank approximation and weight sharing to compress depthwise convolutions, we identify a critical oversight: pointwise convolutions dominate parameter volume (>87% in models like RepLKNet-31B) and constitute the primary deployment bottleneck on resource-constrained edge devices. This results in prohibitive storage costs and severe memory-loading constraints on resource-limited devices (e.g., smartphones with 4–12 GB Random Access Memory (RAM)). To overcome this, we propose Channel Group-Shared (CGS) low-rank approximation, a novel Singular Value Decomposition (SVD)-based parameter-sharing strategy. CGS constructs a structured low-rank paradigm isomorphic to SVD decomposition, comprising shared (high-parameter-cost) down/up-projection matrices across channel groups within a layer and channel-group-specific (low-parameter-cost) scalable diagonal matrices. This group-sharing design achieves significant parameter reduction. Extensive experiments demonstrate that large-kernel CNNs (RepLKNet, ConvNeXt, SLaK) enhanced with CGS strike an empirically favorable balance between competitive performance and substantially reduced storage costs. Crucially, by alleviating storage constraints, reducing memory bandwidth pressure during loading, and minimizing model loading latency, CGS enables the feasible deployment of pre-trained large-kernel CNN models on edge devices, thereby bridging the gap between high-performance vision models and practical edge deployment.

Modern mobile applications (e.g., AR/XR, assistant robots) increasingly rely on multiple concurrent DNN models, yet their inference latency becomes highly unpredictable under dynamic resource contention from concurrent workloads. In mobile SoCs with heterogeneous CPU architectures, such as big. LITTLE designs, the latency is affected by a complex interplay of factors, including heterogeneous core performance, cache hierarchies and shared memory bandwidth. Existing latency prediction methods either fail to capture runtime dynamics or overlook the heterogeneity in mobile processors. In this paper, we propose nn-Live that can predict live inference latency for DNN models on mobile CPUs under dynamically changing operating conditions. nn-Live leverages per-core hardware counter statistics to quantify the interference from concurrent tasks, and utilizes a lightweight linear model to estimate each core's interference contribution to a specific task, enabling robust latency prediction across dynamic execution conditions. Given the large pool of DNN models and huge potential task placement configurations, nn-Live further adopts an adaptive hierarchical sampling strategy to reduce profiling cost while preserving configuration diversity and representativeness. We build a large-scale interference analysis dataset1 across three mainstream SoCs. Evaluation results show that nn-Live achieves high prediction accuracy under diverse runtime environments, with a clear advantage over existing methods. We further demonstrate the extensibility of nn-Live to mobile accelerators (e.g., GPU/NPU).

On-device embodied AI allows local processing of visual data and prompt responses to user requests. Applying embodied AI to more complex domains requires spatial cognition, to properly comprehend the 3D environment and geometric relations between objects. However, existing vision-language models (VLMs) are weak in spatial cognition due to their limitations to 2D image understandings. In this paper, we present SpatialMind, a new on-device embodied AI technique that bridges the gap between 2D visual inputs and 3D physical reality, by injecting 3D spatial knowledge into on-device VLMs in a sparse form of 2D allocentric spatial memory. This memory is constructed by systematically aligning cross-frame visual features to a unified global viewpoint, and we prune the scope of on-device computation to maximize the compute efficiency in memory construction and VLM reasoning. Experiment results in diverse indoor environments show that SpatialMind greatly enhances the performance of spatially aware on-device embodied AI tasks, with high adaptability, robustness and compute efficiency.

Reconfigurable intelligent surfaces (RIS) offer a promising way to control wireless propagation by engineering reflected paths. However, realizing this vision requires real-time knowledge of users' location relative to the RIS. This is challenging because RIS lack the capability to sense the channel, while existing approaches rely on beam search with high overhead. This paper presents MORSE, a novel one-shot localization framework that leverages the inherent frequency-dependent radiation patterns of RIS to create spatial-spectral encoding, where each spatial direction corresponds to a unique spectral signature. A user can then infer its angular position relative to the RIS by observing the received power spectrum. To ensure robustness in multipath environments, MORSE excites the RIS using an FMCW chirp, allowing the receiver to separate propagation paths by delay and isolate the spectral signature of the direct RIS-UE path. We implement MORSE on a 12×12 RIS prototype operating from 3 to 6 GHz and evaluate it through extensive over-the-air experiments in three environments. Our results demonstrate accurate single-shot joint azimuth and elevation angle estimation and robustness to multipath interference, enabling RIS adaptation in dynamic wireless environments.

Next-generation extended reality (XR) networks rely on mmWave communication for multi-gigabit throughput, yet highly directional links are vulnerable to user mobility and blockages, causing frequent outages under reactive beam management. Emerging neural radio fields can predict radio propagation, but prior work remains limited to offline channel reconstruction. We introduce RadioSight, a real-time multi-modal radio field system for predictive mmWave optimization and proactive Multi-User MIMO beamforming. RadioSight combines backward beam-tracing with real-time semantic object synchronization to anticipate RF geometry changes without full model retraining. Implemented as an edge-executable pipeline for commercial 28 GHz arrays, RadioSight determines each scheduling window's beams during the preceding window without exhaustive beam sweeps. Experiments show that RadioSight reduces beam-search error by up to ~50%, improves median throughput by 2×, and enhances link stability.

The photoacoustic effect converts light into sound, bridging optical excitation and acoustic generation. However, existing photoacoustic systems remain largely confined to specialized applications because the generated signals are weak and typically require expensive lasers and delicate materials. In this paper, we demonstrate for the first time that natural butterfly wings can function as highly effective photoacoustic materials, enabling strong and practical photoacoustic generation using an ordinary LED. The key principle is that butterfly wings are not only excellent light absorbers but also support efficient light-heat-acoustic conversion, owing to their unique ultra-thin and lightweight membrane structures. In this work, we establish a model of the butterfly photoacoustic effect, providing a theoretical foundation for future research. We address critical technical challenges, including weak signal, as well as practical challenges such as environmental interference, realizing the first end-to-end butterfly photoacoustic system. Results show that our system achieves strong photoacoustic outputs exceeding 70 dB—comparable to a speaking voice—using off-the-shelf LEDs and naturally deceased butterfly specimens. We demonstrate the application potential of the proposed butterfly photoacoustic effect by using the butterfly as a bio-speaker for audible information delivery and for fine-grained ultrasound sensing. Our system harnesses this intriguing cross-energy phenomenon for practical applications and opens new opportunities for bio-inspired designs.

Robots increasingly must perform dexterous object manipulation in real-world environments, yet existing VTS offer rich camera-based spatial readout while rigid optics limit force sensitivity, force range, and active compliance regulation. We present MoiréSkin, an inflatable visuo-tactile skin using flexible Moiré optics to unify ultra-sensitive tactile sensing and active compliance regulation. Dual-sided submillimeter gratings in a soft membrane amplify microscopic contact-induced strain into large, high-contrast Moiré fringe variations for tactile readout. To balance reliability and sensitivity, our tailored soft-material grating pipeline combines internal illumination with fluorescent doping to suppress ambient-light interference. Self-supervised calibration compensates for air-pressure-dependent sensing drift during pneumatic regulation, while a neural pipeline uses pre- and post-contact fringes to estimate contact position and force. Experiments show that MoiréSkin achieves 0.094 mm position and 0.0038 N contact force MAE, outperforming prior VTS by an order of magnitude. It also reaches 99.7% 10-shot classification accuracy over 7 ultra-soft materials spanning 8.73–21.72 kPa Young's modulus, and measures shear-force-related liquid mass variation with ~4 g precision, demonstrating a new path toward ultra-sensitive sensor-actuator visuo-tactile skins.

Human pose estimation using inertial measurement units (IMUs) has gained increasing attention due to its ability to enable camera-free motion capture for numerous human-centered applications. However, estimating full-body pose from only one to three daily carried commercial-off-the-shelf IMU devices (e.g., smartphones, smartwatches, and earphones) is highly under-constrained, as sparse sensor coverage and noisy signals often lead to unstable predictions and physically implausible body configurations. To address this challenge, we propose ViPoser, a lightweight pose estimation framework that transfers human-structure priors from a large vision foundation model into an IMU-based model through knowledge distillation. Specifically, we analyze the human-centric vision model Sapiens to identify layers rich in biomechanical and skeletal knowledge, distill these representations into a compact student Transformer, and integrate the distilled priors into a multi-stage pose estimation pipeline that performs local joint estimation, full-body masked completion, and global motion recovery. This design enables the model to infer anatomically plausible full-body poses even under sparse IMU observations while remaining suitable for real-time edge deployment. Experiments on multiple real-world datasets show that ViPoser consistently outperforms state-of-the-art IMU-based methods such as IMUPoser and MobilePoser, achieving 10~30% lower pose estimation error. In addition, ViPoser produces more anatomically realistic poses, significantly reducing failure cases observed in prior methods where predicted limbs intersect the body or violate biomechanical constraints.

Diversified wireless sensing and localization tasks impose distinct requirements on the underlying perception platform. However, existing wireless platforms are typically fabricated with a fixed array configuration, requiring substantial hardware redesign to adapt to different applications. In this paper, we present UltraLEGO, a modular ultra-wideband phasedarray architecture that enables flexible array reconfiguration without circuit redesign while simultaneously supporting heterogeneous sensing and localization paradigms. To address phase incoherence among modularized array elements, we leverage multipath correlations in wireless channel estimation and incorporate the mechanism into a unified sensing and localization framework. Extensive experiments demonstrate that UltraLEGO adapts efficiently to diverse tasks while achieving state-of-the-art perception accuracy.

To support real-time and domain-customized edge intelligent applications, LLMs are increasingly deployed on edge computing servers. However, we observe that current MoE-based LLM serving systems, inherited from the cloud-based architectures, face a critical sparsity-vanishing problem when serving at the edge: tokens within an input batch often activate a large fraction of experts, which incurs significant I/O overhead from frequent expert swapping between CPU and GPU memory, ultimately degrading system throughput. In this work, we propose Saber, a sparsity-aware scheduling framework optimized for efficient MoE-based LLM serving on edge servers. The core idea of Saber is to introduce a group-level batching strategy that dynamically identifies requests with similar activated experts, and coalesces them to construct sparsity-preserving input batches. To ensure efficient online execution, activation patterns are precomputed and clustered offline, enabling fast request grouping through a lightweight classifier at runtime. Additionally, Saber incorporates a utility-driven priority scheduling mechanism to maximize inference throughput while maintaining low latency by intelligently adjusting group priorities. Experiments on representative MoE-based LLMs demonstrate that Saber significantly outperforms existing sparsity-agnostic schedulers, achieving up to 3.8× lower batch execution time and 3.27× higher throughput at comparable latency levels.

Wireless power transfer (WPT) enables contactless power delivery, with magnetic resonance coupling (MRC-WPT) as a leading approach. However, existing MRC-WPT systems rely on fixed transmitter (TX) geometries, limiting their tolerance to receiver (RX) displacement. While circuit-level tuning can improve power flow, it operates within fixed magnetic pathways and cannot reconfigure the field at the source, thus offering limited adaptability. This paper presents Shape-Reconfigurable Charger (SRC), an in-plane MRC-WPT charging system that senses its pose, morphs its structure to optimize power pathways, and adapts its configuration. SRC employs silicone tubes filled with liquid metal (LM) as coils and attaches liquid crystal elastomer (LCE) fibers as tendons, enabling morphable TX coils. We formulate online coil configuration as a dynamic black-box optimization and develop a distributed Bayesian optimization framework that runs on each TX, requiring no RX communication or inter-TX coordination. A hardware prototype demonstrates our performance across diverse RX placements: SRC improves power delivered to load (PDL) by 174% over frequency tuning and 96% over motorized rotation on average. By reframing the TX design as adaptive field shaping, SRC moves MRC-WPT beyond rigid transmission toward programmable magnetic pathways, advancing the vision of truly unconstrained wireless charging.

The disaggregation of the 5G Radio Access Network (RAN) decomposes traditional base stations into distributed components such as the centralized unit (CU) and distributed unit (DU). While this architecture improves flexibility and vendor interoperability, it also turns many control procedures into distributed workflows executed by independent state machines. We show that such executions can lead to Disaggregation Faults (DF), where specification-compliant procedure interactions drive different RAN components into inconsistent control-plane states. We present D-Fault, a testing framework that systematically uncovers DF bugs by generating adversarial interleavings of control procedures and initiating them with targeted timing. A model-guided exploration reduces the intractable search space, while verification based on externally observable behavior detects inconsistencies without access to internal states. We implement D-Fault and evaluate it on SDR-based testbeds, uncovering 17 unique DF bugs that lead to serious effects such as resource corruption and denial of service.

Nocturnal blood pressure (BP) is a stronger predictor of cardiovascular risk than daytime measurements, yet its unobtrusive monitoring remains a formidable challenge. For practical convenience, existing cuffless methods typically rely on daytime calibration. However, through a retrospective analysis of a 1,328-patient ICU dataset, we reveal a pronounced day-night domain gap: vascular states undergo highly personalized, non-stationary drifts during sleep. Consequently, conventional daytime calibrations fail to capture nocturnal hemodynamic shifts, leading to substantial measurement errors. To address this, we propose VasTrack, a mmWave wearable system that bridges the day-night gap via a novel “Anchor-and-Track” architecture. Specifically, to establish a vascular anchor, we design a full-spectrum pressure scan method to extract the user's intrinsic vascular compliance during a standard cuff measurement, serving as a personalized biomechanical prior for nocturnal BP tracking. To track nocturnal vascular dynamics, our multi-granular tracking network transforms temporal signals into 2D phase-aligned physiological manifolds. It then employs biomechanical-based rhythmic gating to dynamically compensate for vascular tone variations using cardiopulmonary rhythms as proxies. Evaluation on 43 participants spanning healthy individuals and clinical patients over 340 hours shows that VasTrack achieves a mean error and standard deviation of 0.24±7.31 mmHg for systolic blood pressure and 0.10±4.87 mmHg for diastolic blood pressure. It also captures acute BP fluctuations exceeding 30 mmHg and a preliminary six-month evaluation suggests feasibility for extended monitoring periods, offering a potential pathway toward long-term cardiovascular health management.