Mo Li

Linshan Jiang, Qun Song, Rui Tan et al.

Mobile cloud offloading is indispensable for inference tasks based on large-scale deep models. However, transmitting privacy-rich inference data to the cloud incurs concerns. This paper presents the design of a system called PriMask, in which the mobile device uses a secret small-scale neural network called MaskNet to mask the data before transmission. PriMask significantly weakens the cloud's capability to recover the data or extract certain private attributes. The MaskNet is em cascadable in that the mobile can opt in to or out of its use seamlessly without any modifications to the cloud's inference service. Moreover, the mobiles use different MaskNets, such that the collusion between the cloud and some mobiles does not weaken the protection for other mobiles. We devise a {\em split adversarial learning} method to train a neural network that generates a new MaskNet quickly (within two seconds) at run time. We apply PriMask to three mobile sensing applications with diverse modalities and complexities, i.e., human activity recognition, urban environment crowdsensing, and driver behavior recognition. Results show PriMask's effectiveness in all three applications.

Sihan Yang, Runsen Xu, Yiman Xie et al. · pku

Spatial intelligence is essential for multimodal large language models (MLLMs) operating in the complex physical world. Existing benchmarks, however, probe only single-image relations and thus fail to assess the multi-image spatial reasoning that real-world deployments demand. We introduce MMSI-Bench, a VQA benchmark dedicated to multi-image spatial intelligence. Six 3D-vision researchers spent more than 300 hours meticulously crafting 1,000 challenging, unambiguous multiple-choice questions from over 120,000 images, each paired with carefully designed distractors and a step-by-step reasoning process. We conduct extensive experiments and thoroughly evaluate 34 open-source and proprietary MLLMs, observing a wide gap: the strongest open-source model attains roughly 30% accuracy and OpenAI's o3 reasoning model reaches 40%, while humans score 97%. These results underscore the challenging nature of MMSI-Bench and the substantial headroom for future research. Leveraging the annotated reasoning processes, we also provide an automated error analysis pipeline that diagnoses four dominant failure modes, including (1) grounding errors, (2) overlap-matching and scene-reconstruction errors, (3) situation-transformation reasoning errors, and (4) spatial-logic errors, offering valuable insights for advancing multi-image spatial intelligence. Project page: https://runsenxu.com/projects/MMSI_Bench .

Gaole Dai, Shiqi Jiang, Ting Cao et al.

We propose V-Droid, a mobile GUI task automation agent. Unlike previous mobile agents that utilize Large Language Models (LLMs) as generators to directly generate actions at each step, V-Droid employs LLMs as verifiers to evaluate candidate actions before making final decisions. To realize this novel paradigm, we introduce a comprehensive framework for constructing verifier-driven mobile agents: the discretized action space construction coupled with the prefilling-only workflow to accelerate the verification process, the pair-wise progress preference training to significantly enhance the verifier's decision-making capabilities, and the scalable human-agent joint annotation scheme to efficiently collect the necessary data at scale. V-Droid obtains a substantial task success rate across several public mobile task automation benchmarks: 59.5% on AndroidWorld, 38.3% on AndroidLab, and 49% on MobileAgentBench, surpassing existing agents by 5.2%, 2.1%, and 9%, respectively. Furthermore, V-Droid achieves a remarkably low latency of 4.3s per step, which is 6.1x faster compared with existing mobile agents. The source code is available at https://github.com/V-Droid-Agent/V-Droid.

Sijie Ji, Mo Li

CSI feedback is an important problem of Massive multiple-input multiple-output (MIMO) technology because the feedback overhead is proportional to the number of sub-channels and the number of antennas, both of which scale with the size of the Massive MIMO system. Deep learning-based CSI feedback methods have been widely adopted recently owing to their superior performance. Despite the success, current approaches have not fully exploited the relationship between the characteristics of CSI data and the deep learning framework. In this paper, we propose a jigsaw puzzles aided training strategy (JPTS) to enhance the deep learning-based Massive MIMO CSI feedback approaches by maximizing mutual information between the original CSI and the compressed CSI. We apply JPTS on top of existing state-of-the-art methods. Experimental results show that by adopting this training strategy, the accuracy can be boosted by 12.07% and 7.01% on average in indoor and outdoor environments, respectively. The proposed method is ready to adopt to existing deep learning frameworks of Massive MIMO CSI feedback. Codes of JPTS are available on GitHub for use.

Shenghong Dai, Shiqi Jiang, Yifan Yang et al.

This paper presents Babel, the expandable modality alignment model, specially designed for multi-modal sensing. While there has been considerable work on multi-modality alignment, they all struggle to effectively incorporate multiple sensing modalities due to the data scarcity constraints. How to utilize multi-modal data with partial pairings in sensing remains an unresolved challenge. Babel tackles this challenge by introducing the concept of expandable modality alignment. The key idea involves transforming the N-modality alignment into a series of binary-modality alignments. Novel techniques are also proposed to further mitigate data scarcity issue and balance the contribution of the newly incorporated modality with the previously established modality alignment during the expandable alignment process. We provide the comprehensive implementation. In the pre-training phase, Babel currently aligns 6 sensing modalities, namely Wi-Fi, mmWave, IMU, LiDAR, video, and depth. For the deployment phase, as a foundation model, any single or combination of aligned modalities could be selected from Babel and applied to downstream tasks. Evaluation demonstrates Babel's outstanding performance on eight human activity recognition datasets, compared to a broad range of baselines e.g., the SOTA single-modal sensing networks, multi-modal sensing framework, and multi-modal large language models. Babel not only improves the performance of individual modality sensing (12% averaged accuracy improvement), but also effectively fuses multiple available modalities (up to 22% accuracy increase). Case studies also highlight emerging application scenarios empowered by Babel, including cross-modality retrieval (i.e., sensing imaging), and bridging LLM for sensing comprehension.

Huatao Xu, Yan Zhang, Wei Gao et al.

This paper presents the first nationwide deployment of human activity recognition (HAR) technology in the on-demand food delivery industry. We successfully adapted the state-of-the-art LIMU-BERT foundation model to the delivery platform. Spanning three phases over two years, the deployment progresses from a feasibility study in Yangzhou City to nationwide adoption involving 500,000 couriers across 367 cities in China. The adoption enables a series of downstream applications, and large-scale tests demonstrate its significant operational and economic benefits, showcasing the transformative potential of HAR technology in real-world applications. Additionally, we share lessons learned from this deployment and open-source our LIMU-BERT pretrained with millions of hours of sensor data.

Maosong Cao, Taolin Zhang, Mo Li et al. · pku

The quality of Supervised Fine-Tuning (SFT) data plays a critical role in enhancing the conversational capabilities of Large Language Models (LLMs). However, as LLMs become more advanced, the availability of high-quality human-annotated SFT data has become a significant bottleneck, necessitating a greater reliance on synthetic training data. In this work, we introduce Condor, a novel two-stage synthetic data generation framework that incorporates World Knowledge Tree and Self-Reflection Refinement to produce high-quality SFT data at scale. Our experimental results demonstrate that a base model fine-tuned on only 20K Condor-generated samples achieves superior performance compared to counterparts. The additional refinement stage in Condor further enables iterative self-improvement for LLMs at various scales (up to 72B), validating the effectiveness of our approach. Furthermore, our investigation into the scaling for synthetic data in post-training reveals substantial unexplored potential for performance improvements, opening promising avenues for future research.

Gaole Dai, Huatao Xu, Yifan Yang et al.

Expanding existing learning systems to provide high-quality customized models for more domains, such as new users, is challenged by the limited labeled data and the data and device heterogeneities. While knowledge distillation methods could overcome label scarcity and device heterogeneity, they assume the teachers are fully reliable and overlook the data heterogeneity, which prevents the direct adoption of existing models. To address this problem, this paper proposes a framework, HaT, to expand learning systems. It first selects multiple high-quality models from the system at a low cost and then fuses their knowledge by assigning sample-wise weights to their predictions. Later, the fused knowledge is selectively injected into the customized models based on the knowledge quality. Extensive experiments on different tasks, modalities, and settings show that HaT outperforms state-of-the-art baselines by up to 16.5% accuracy and saves up to 39% communication traffic.

Hyungjun Yoon, Jaehyun Kwak, Biniyam Aschalew Tolera et al.

Self-supervised learning has emerged as a method for utilizing massive unlabeled data for pre-training models, providing an effective feature extractor for various mobile sensing applications. However, when deployed to end-users, these models encounter significant domain shifts attributed to user diversity. We investigate the performance degradation that occurs when self-supervised models are fine-tuned in heterogeneous domains. To address the issue, we propose SelfReplay, a few-shot domain adaptation framework for personalizing self-supervised models. SelfReplay proposes self-supervised meta-learning for initial model pre-training, followed by a user-side model adaptation by replaying the self-supervision with user-specific data. This allows models to adjust their pre-trained representations to the user with only a few samples. Evaluation with four benchmarks demonstrates that SelfReplay outperforms existing baselines by an average F1-score of 8.8%p. Our on-device computational overhead analysis on a commodity off-the-shelf (COTS) smartphone shows that SelfReplay completes adaptation within an unobtrusive latency (in three minutes) with only a 9.54% memory consumption, demonstrating the computational efficiency of the proposed method.

Chaojie Gu, Linshan Jiang, Rui Tan et al.

Low-power wide-area network technologies such as LoRaWAN are promising for collecting low-rate monitoring data from geographically distributed sensors, in which timestamping the sensor data is a critical system function. This paper considers a synchronization-free approach to timestamping LoRaWAN uplink data based on signal arrival time at the gateway, which well matches LoRaWAN's one-hop star topology and releases bandwidth from transmitting timestamps and synchronizing end devices' clocks at all times. However, we show that this approach is susceptible to a {\em frame delay attack} consisting of malicious frame collision and delayed replay. Real experiments show that the attack can affect the end devices in large areas up to about $50,000\,\text{m}^2$. In a broader sense, the attack threatens any system functions requiring timely deliveries of LoRaWAN frames. To address this threat, we propose a $\mathsf{LoRaTS}$ gateway design that integrates a commodity LoRaWAN gateway and a low-power software-defined radio receiver to track the inherent frequency biases of the end devices. Based on an analytic model of LoRa's chirp spread spectrum modulation, we develop signal processing algorithms to estimate the frequency biases with high accuracy beyond that achieved by LoRa's default demodulation. The accurate frequency bias tracking capability enables the detection of the attack that introduces additional frequency biases. We also investigate and implement a more crafty attack that uses advanced radio apparatuses to eliminate the frequency biases. To address this crafty attack, we propose a pseudorandom interval hopping scheme to enhance our frequency bias tracking approach. Extensive experiments show the effectiveness of our approach in deployments with real affecting factors such as temperature variations.