arXiv subjects

Yaxin Li

Publications and source records attributed to Yaxin Li.

At least 19 recordsLinked to original sources

arXiv2026

Devil in the Lens: Analyzing and Defending Physical Prompt Injection Against Vision-Language Models on Wearable Devices

Vision-Language Models (VLMs) are rapidly deployed on human-facing wearable devices such as smart glasses to enable multimodal perception and AI-assisted decision-making. While prior research has demonstrated the risks of visual prompt injection into digital image inputs of VLMs, the unique security challenges posed by the increasing integration between physical environments and wearable intelligence, such as those embodied in VLM-enabled AI glasses, remain underexplored. Toward understanding and modeling such threats, our work characterizes how malicious textual information embedded in physical environments introduces a high-priority visual channel for indirect prompt injection, where scene texts that hinder or evade human perception could hijack VLM models' behavior. Such \textit{Physical Prompt Injection Attacks} can not only disrupt normal tasks of VLM-enabled wearable devices, but also steer models to produce profane, biased, or even untruthful outputs. Using physically captured photos from AI glasses in over 200 real-world environments, our analysis identifies 6 representative threat vectors of physically injected prompts, and further evaluates their impacts on 12 VLM models. Results show that these attacks consistently manipulate model outputs across integrity- and safety-critical tasks, achieving attack success rates of up to 96\% and 60\% in simulated and real-world settings. Our analysis confirms that multiple models exhibit excessive blind trust in environmental text, ignoring the actual visual context and producing completely opposite summaries or directives. We further propose two targeted defense strategies, including a mask-based external filter and a semantic-vector-based internal detector, to effectively reduce the success rate and safety impact of these attacks.

cs.CR↗

arXiv2026

Noncyclic geometric phase in three-level Ramsey interferometry for enhanced metrology

In a standard two-level Ramsey interferometer, the accumulated signal phase is linearly mapped to the readout phase. Here, we introduce three-level Ramsey interferometry, in which projected interference between internal paths reshapes this mapping through a noncyclic geometric phase response. Near a geodesic-closure transition, a small accumulated signal phase produces a sharply amplified readout-phase shift. We quantify the accompanying gain--visibility tradeoff and identify a finite operating window in which the amplified response can yield a net signal-to-noise-ratio gain in the presence of additional technical phase noise. A controllable Ramsey phase offset further positions this high-slope response at a chosen operating point, making the local enhancement accessible without requiring a large signal-induced phase. More broadly, these results establish a multilevel Ramsey route to enhanced phase sensitivity in quantum platforms, in which two signal-collecting internal paths interfere to produce a controllable noncyclic geometric response.

quant-ph↗

arXiv2026

UniPinRec: Unifying Generative Retrieval and Ranking at Pinterest Scale

Modern recommendation systems predominantly train retrieval and ranking as separate models despite both increasingly relying on large transformers encoding the same user behavior data, duplicating parameters, compute, and serving cost. Prior work unifies the model architecture but not the full pipeline: input formats, training procedures, and serving stacks remain fragmented across stages. We present UniPinRec, which achieves full-stack unification of retrieval and ranking at Pinterest: one input format, one model, one training stage, deployed within existing serving infrastructure. A shared transformer encodes the user action sequence into candidate-independent representations that branch into retrieval (ANN dot-product) and ranking (cross-attention) via task-specific heads. Three ideas make this work: (1) Masked Action Modeling (MAM) eliminates interleaving, enabling weight sharing without doubling context length; (2) Blended training examples pair action sequences with feedview impression slates to satisfy both objectives jointly; (3) Cross-stage KV cache sharing reuses user-history computation from retrieval for ranking, reducing total FLOPs versus serving two independent models. Deployed in the Pinterest core surfaces, UniPinRec delivers approximately +1% online engagement lift while cutting end-to-end serving latency by 11.1% and lifting QPS by 63.6%. To our knowledge, this is the first full-stack unification of retrieval and ranking, covering inputs, model, training and serving, deployed in a production recommendation system.

cs.IR↗

arXiv2025

Distillation-Guided Structural Transfer for Continual Learning Beyond Sparse Distributed Memory

Sparse neural systems are gaining traction for efficient continual learning due to their modularity and low interference. Architectures such as Sparse Distributed Memory Multi-Layer Perceptrons (SDMLP) construct task-specific subnetworks via Top-K activation and have shown resilience against catastrophic forgetting. However, their rigid modularity limits cross-task knowledge reuse and leads to performance degradation under high sparsity. We propose Selective Subnetwork Distillation (SSD), a structurally guided continual learning framework that treats distillation not as a regularizer but as a topology-aligned information conduit. SSD identifies neurons with high activation frequency and selectively distills knowledge within previous Top-K subnetworks and output logits, without requiring replay or task labels. This enables structural realignment while preserving sparse modularity. Experiments on Split CIFAR-10, CIFAR-100, and MNIST demonstrate that SSD improves accuracy, retention, and representation coverage, offering a structurally grounded solution for sparse continual learning.

cs.LG↗

arXiv2025

High-Fidelity Microwave-Polarization Control in a Rydberg-Ensemble Experiment

Control of the polarization of microwave fields is a key experimental capability for a number of atomic physics platforms. However, producing high-fidelity microwaves requires a well-controlled microwave environment, where reflections that distort the polarization must be avoided or well characterized, a constraint that often conflicts with other experimental design considerations. Here we demonstrate a microwave control system capable of producing high-fidelity microwave polarizations in a Rydberg-ensemble experiment. We use three in-vacuum DC electrodes, repurposed as microwave antennae, to produce imperfect and initially unknown polarizations. Each source is driven with independent phase and amplitude control to generate the desired microwave fields. We probe the fields produced at the position of the atoms using Rydberg-EIT spectroscopy of the microwave-induced avoided crossings. We produce $\sigma_-$, $\pi$, and $\sigma_+$ polarized microwaves with > 99 % fidelity and generate their combinations. We extend our purification techniques to frequencies away from Rydberg resonances by utilizing an auxiliary microwave field, generating two-photon microwave resonances. The techniques developed here will facilitate the engineering of dipolar interactions in atomic and molecular physics experiments.

physics.atom-ph↗

arXiv2025

Toward Engineering AGI: Benchmarking the Engineering Design Capabilities of LLMs

Modern engineering, spanning electrical, mechanical, aerospace, civil, and computer disciplines, stands as a cornerstone of human civilization and the foundation of our society. However, engineering design poses a fundamentally different challenge for large language models (LLMs) compared with traditional textbook-style problem solving or factual question answering. Although existing benchmarks have driven progress in areas such as language understanding, code synthesis, and scientific problem solving, real-world engineering design demands the synthesis of domain knowledge, navigation of complex trade-offs, and management of the tedious processes that consume much of practicing engineers' time. Despite these shared challenges across engineering disciplines, no benchmark currently captures the unique demands of engineering design work. In this work, we introduce EngDesign, an Engineering Design benchmark that evaluates LLMs' abilities to perform practical design tasks across nine engineering domains. Unlike existing benchmarks that focus on factual recall or question answering, EngDesign uniquely emphasizes LLMs' ability to synthesize domain knowledge, reason under constraints, and generate functional, objective-oriented engineering designs. Each task in EngDesign represents a real-world engineering design problem, accompanied by a detailed task description specifying design goals, constraints, and performance requirements. EngDesign pioneers a simulation-based evaluation paradigm that moves beyond textbook knowledge to assess genuine engineering design capabilities and shifts evaluation from static answer checking to dynamic, simulation-driven functional verification, marking a crucial step toward realizing the vision of engineering Artificial General Intelligence (AGI).

cs.CE↗

arXiv2025

Legal Mathematical Reasoning with LLMs: Procedural Alignment through Two-Stage Reinforcement Learning

Legal mathematical reasoning is essential for applying large language models (LLMs) in high-stakes legal contexts, where outputs must be both mathematically accurate and procedurally compliant. However, existing legal LLMs lack structured numerical reasoning, and open-domain models, though capable of calculations, often overlook mandatory legal steps. To address this, we present LexNum, the first Chinese legal mathematical reasoning benchmark, covering three representative scenarios where each instance reflects legally grounded procedural flows. We further propose LexPam, a two-stage reinforcement learning framework for efficient legal reasoning training. Leveraging curriculum learning, we use a stronger teacher model to partition data into basic and challenging subsets. A lightweight 1.5B student model is then fine-tuned with Group Relative Policy Optimization, which avoids costly value networks and enables stable training from sparse, end-of-sequence rewards. The first stage improves accuracy and format; the second introduces a novel reward to guide procedural alignment via task-specific legal elements. Experiments show that existing models perform poorly on LexNum, while LexPam enhances both mathematical accuracy and legal coherence, and generalizes effectively across tasks and domains.

cs.CL↗

arXiv2025

Artificial Intelligence-derived Photoplethysmography Age as a Digital Biomarker for Cardiovascular Health

Background: Photoplethysmography (PPG), increasingly available through wearable devices, provides a non-invasive means of monitoring human hemodynamics. In this study, we introduce artificial intelligence-derived photoplethysmography (AI-PPG) age, a deep learning-based estimate of biological age from raw PPG signals, and evaluate its potential as a digital biomarker for cardiovascular health. Methods: We developed a deep learning model with a distribution-aware loss function to reduce bias from imbalanced data. The model was trained and evaluated on the UK Biobank cohort (N = 212,231). We analyzed the association between the AI-PPG age gap (AI-PPG age minus calendar age) and multiple cardiovascular and metabolic outcomes, assessed its longitudinal value using serial PPG measurements, and externally validated its generalizability in an independent MIMIC-III-derived cohort (N = 2,343). Results: After adjusting for key confounders, participants with an AI-PPG age gap greater than 9 years have a significantly higher risk of major adverse cardiovascular and cerebrovascular events (hazard ratio of 2.37, p = 8.46x10$^{-80}$), as well as seven secondary outcomes including coronary heart disease and myocardial infarction (all p < 0.005). Conversely, those with a gap below -9 years show a lower risk profile. Longitudinal analysis demonstrates that changes in AI-PPG age add predictive value over time. In the external validation cohort, each one-year increase in AI-PPG age gap is associated with higher in-hospital mortality (odds ratio of 1.02, p = 0.01). Conclusions: AI-PPG age is a scalable, non-invasive biomarker for cardiovascular health assessment. Integrated with wearable devices, it may enable population-level screening, personalized monitoring, and early intervention.

eess.SP↗

arXiv2024

Large enhancement of nonlinear optical response of graphene nanoribbon heterojunctions with multiple topological interface states

We investigate the nonlinear optical response of graphene nanoribbon (GNR) heterojunctions both without and with one or multiple topological interface states. By implementing a distant-neighbor quantum-mechanical (DNQM) method, we demonstrate a pronounced enhancement of the nonlinear optical response of GNR heterojunctions as the number of topological states at their interfaces increases. Specifically, we find that GNR heterojunctions with multiple topological interface states exhibit a notably stronger third-order nonlinear optical response in comparison with the similarly sized counterparts with a single topological interface state or without such states. Furthermore, we observe that the presence of topological interface states in GNR heterojunctions can induce a significant red-shift in their quantum plasmon frequency. Our results reveal the potential to enhance the nonlinear optical response at the nanoscale by increasing the number of topological interface states in graphene nanostructures or other topological systems.

cond-mat.mes-hall↗

arXiv2024

Enhancement of Rydberg Blockade via Microwave Dressing

Experimental control over the strength and angular dependence of interactions between atoms is a key capability for advancing quantum technologies. Here, we use microwave dressing to manipulate and enhance Rydberg-Rydberg interactions in an atomic ensemble. By varying the cloud length relative to the blockade radius and measuring the statistics of the light retrieved from the ensemble, we demonstrate a clear enhancement of the interaction strength due to microwave dressing. These results are successfully captured by a theoretical model that accounts for the excitation dynamics, atomic density distribution, and the phase-matched retrieval efficiency. Our approach offers a versatile platform for further engineering interactions by exploiting additional features of the microwave fields, such as polarization and detuning, opening pathways for new quantum control strategies.

physics.atom-ph↗

arXiv2024

Deep Inertia $L_p$ Half-Quadratic Splitting Unrolling Network for Sparse View CT Reconstruction

Sparse view computed tomography (CT) reconstruction poses a challenging ill-posed inverse problem, necessitating effective regularization techniques. In this letter, we employ $L_p$-norm ($0<p<1$) regularization to induce sparsity and introduce inertial steps, leading to the development of the inertial $L_p$-norm half-quadratic splitting algorithm. We rigorously prove the convergence of this algorithm. Furthermore, we leverage deep learning to initialize the conjugate gradient method, resulting in a deep unrolling network with theoretical guarantees. Our extensive numerical experiments demonstrate that our proposed algorithm surpasses existing methods, particularly excelling in fewer scanned views and complex noise conditions.

eess.IV↗

arXiv2024

Towards Efficient Deep Spiking Neural Networks Construction with Spiking Activity based Pruning

The emergence of deep and large-scale spiking neural networks (SNNs) exhibiting high performance across diverse complex datasets has led to a need for compressing network models due to the presence of a significant number of redundant structural units, aiming to more effectively leverage their low-power consumption and biological interpretability advantages. Currently, most model compression techniques for SNNs are based on unstructured pruning of individual connections, which requires specific hardware support. Hence, we propose a structured pruning approach based on the activity levels of convolutional kernels named Spiking Channel Activity-based (SCA) network pruning framework. Inspired by synaptic plasticity mechanisms, our method dynamically adjusts the network's structure by pruning and regenerating convolutional kernels during training, enhancing the model's adaptation to the current target task. While maintaining model performance, this approach refines the network architecture, ultimately reducing computational load and accelerating the inference process. This indicates that structured dynamic sparse learning methods can better facilitate the application of deep SNNs in low-power and high-efficiency scenarios.

cs.NE↗

arXiv2024

Unveiling and Mitigating Memorization in Text-to-image Diffusion Models through Cross Attention

Recent advancements in text-to-image diffusion models have demonstrated their remarkable capability to generate high-quality images from textual prompts. However, increasing research indicates that these models memorize and replicate images from their training data, raising tremendous concerns about potential copyright infringement and privacy risks. In our study, we provide a novel perspective to understand this memorization phenomenon by examining its relationship with cross-attention mechanisms. We reveal that during memorization, the cross-attention tends to focus disproportionately on the embeddings of specific tokens. The diffusion model is overfitted to these token embeddings, memorizing corresponding training images. To elucidate this phenomenon, we further identify and discuss various intrinsic findings of cross-attention that contribute to memorization. Building on these insights, we introduce an innovative approach to detect and mitigate memorization in diffusion models. The advantage of our proposed method is that it will not compromise the speed of either the training or the inference processes in these models while preserving the quality of generated images. Our code is available at https://github.com/renjie3/MemAttn .

cs.CV↗

arXiv2024

Enhancing Adaptive History Reserving by Spiking Convolutional Block Attention Module in Recurrent Neural Networks

Spiking neural networks (SNNs) serve as one type of efficient model to process spatio-temporal patterns in time series, such as the Address-Event Representation data collected from Dynamic Vision Sensor (DVS). Although convolutional SNNs have achieved remarkable performance on these AER datasets, benefiting from the predominant spatial feature extraction ability of convolutional structure, they ignore temporal features related to sequential time points. In this paper, we develop a recurrent spiking neural network (RSNN) model embedded with an advanced spiking convolutional block attention module (SCBAM) component to combine both spatial and temporal features of spatio-temporal patterns. It invokes the history information in spatial and temporal channels adaptively through SCBAM, which brings the advantages of efficient memory calling and history redundancy elimination. The performance of our model was evaluated in DVS128-Gesture dataset and other time-series datasets. The experimental results show that the proposed SRNN-SCBAM model makes better use of the history information in spatial and temporal dimensions with less memory space, and achieves higher accuracy compared to other models.

cs.NE↗

arXiv2023

Exploring Memorization in Fine-tuned Language Models

Large language models (LLMs) have shown great capabilities in various tasks but also exhibited memorization of training data, raising tremendous privacy and copyright concerns. While prior works have studied memorization during pre-training, the exploration of memorization during fine-tuning is rather limited. Compared to pre-training, fine-tuning typically involves more sensitive data and diverse objectives, thus may bring distinct privacy risks and unique memorization behaviors. In this work, we conduct the first comprehensive analysis to explore language models' (LMs) memorization during fine-tuning across tasks. Our studies with open-sourced and our own fine-tuned LMs across various tasks indicate that memorization presents a strong disparity among different fine-tuning tasks. We provide an intuitive explanation of this task disparity via sparse coding theory and unveil a strong correlation between memorization and attention score distribution.

cs.AI↗

arXiv2023

Number-State Reconstruction with a Single Single-Photon Avalanche Detector

Single-photon avalanche detectors (SPADs) are crucial sensors of light for many fields and applications. However, they are not able to resolve photon number, so typically more complex and more expensive experimental setups or devices must be used to measure the number of photons in a pulse. Here, we present a methodology for performing photon number-state reconstruction with only one SPAD. The methodology, which is cost-effective and easy to implement, uses maximum-likelihood techniques with a detector model whose parameters are measurable. We achieve excellent agreement between known input pulses and their reconstructions for coherent states with up to $\approx$ 10 photons and peak input photon rates up to several Mcounts/s. When detector imperfections are small, we maintain good agreement for coherent pulses with peak input photon rates of over 40 Mcounts/s, greater than one photon per detector dead time. For anti-bunched light, the reconstructed and independently measured pulse-averaged values of $g^{(2)}(0)$ are also consistent with one another. Our algorithm is applicable to light pulses whose pulse width and correlation time scales are both at least a few detector dead times. These results, achieved with single commercially available SPADs, provide an inexpensive number-state reconstruction method and expand the capabilities of single-photon detectors.

quant-ph↗

arXiv2023

3D Reconstruction of Spherical Images based on Incremental Structure from Motion

3D reconstruction plays an increasingly important role in modern photogrammetric systems. Conventional satellite or aerial-based remote sensing (RS) platforms can provide the necessary data sources for the 3D reconstruction of large-scale landforms and cities. Even with low-altitude UAVs (Unmanned Aerial Vehicles), 3D reconstruction in complicated situations, such as urban canyons and indoor scenes, is challenging due to the frequent tracking failures between camera frames and high data collection costs. Recently, spherical images have been extensively exploited due to the capability of recording surrounding environments from one camera exposure. Classical 3D reconstruction pipelines, however, cannot be used for spherical images. Besides, there exist few software packages for 3D reconstruction of spherical images. Based on the imaging geometry of spherical cameras, this study investigates the algorithms for the relative orientation using spherical correspondences, absolute orientation using 3D correspondences between scene and spherical points, and the cost functions for BA (bundle adjustment) optimization. In addition, an incremental SfM (Structure from Motion) workflow has been proposed for spherical images using the above-mentioned algorithms. The proposed solution is finally verified by using three spherical datasets captured by both consumer-grade and professional spherical cameras. The results demonstrate that the proposed SfM workflow can achieve the successful 3D reconstruction of complex scenes and provide useful clues for the implementation in open-source software packages. The source code of the designed SfM workflow would be made publicly available.

cs.CV↗

arXiv2023

3D reconstruction from spherical images: A review of techniques, applications, and prospects

3D reconstruction plays an increasingly important role in modern photogrammetric systems. Conventional satellite or aerial-based remote sensing (RS) platforms can provide the necessary data sources for the 3D reconstruction of large-scale landforms and cities. Even with low-altitude UAVs (Unmanned Aerial Vehicles), 3D reconstruction in complicated situations, such as urban canyons and indoor scenes, is challenging due to frequent tracking failures between camera frames and high data collection costs. Recently, spherical images have been extensively used due to the capability of recording surrounding environments from one camera exposure. In contrast to perspective images with limited FOV (Field of View), spherical images can cover the whole scene with full horizontal and vertical FOV and facilitate camera tracking and data acquisition in these complex scenes. With the rapid evolution and extensive use of professional and consumer-grade spherical cameras, spherical images show great potential for the 3D modeling of urban and indoor scenes. Classical 3D reconstruction pipelines, however, cannot be directly used for spherical images. Besides, there exist few software packages that are designed for the 3D reconstruction of spherical images. As a result, this research provides a thorough survey of the state-of-the-art for 3D reconstruction of spherical images in terms of data acquisition, feature detection and matching, image orientation, and dense matching as well as presenting promising applications and discussing potential prospects. We anticipate that this study offers insightful clues to direct future research.

cs.CV↗