arXiv subjects

Jie Zhang

Publications and source records attributed to Jie Zhang.

At least 451 records · Page 25Linked to original sources

arXiv2025

CORBA: Contagious Recursive Blocking Attacks on Multi-Agent Systems Based on Large Language Models

Large Language Model-based Multi-Agent Systems (LLM-MASs) have demonstrated remarkable real-world capabilities, effectively collaborating to complete complex tasks. While these systems are designed with safety mechanisms, such as rejecting harmful instructions through alignment, their security remains largely unexplored. This gap leaves LLM-MASs vulnerable to targeted disruptions. In this paper, we introduce Contagious Recursive Blocking Attacks (Corba), a novel and simple yet highly effective attack that disrupts interactions between agents within an LLM-MAS. Corba leverages two key properties: its contagious nature allows it to propagate across arbitrary network topologies, while its recursive property enables sustained depletion of computational resources. Notably, these blocking attacks often involve seemingly benign instructions, making them particularly challenging to mitigate using conventional alignment methods. We evaluate Corba on two widely-used LLM-MASs, namely, AutoGen and Camel across various topologies and commercial models. Additionally, we conduct more extensive experiments in open-ended interactive LLM-MASs, demonstrating the effectiveness of Corba in complex topology structures and open-source models. Our code is available at: https://github.com/zhrli324/Corba.

cs.CL↗

arXiv2025

Does Training with Synthetic Data Truly Protect Privacy?

As synthetic data becomes increasingly popular in machine learning tasks, numerous methods--without formal differential privacy guarantees--use synthetic data for training. These methods often claim, either explicitly or implicitly, to protect the privacy of the original training data. In this work, we explore four different training paradigms: coreset selection, dataset distillation, data-free knowledge distillation, and synthetic data generated from diffusion models. While all these methods utilize synthetic data for training, they lead to vastly different conclusions regarding privacy preservation. We caution that empirical approaches to preserving data privacy require careful and rigorous evaluation; otherwise, they risk providing a false sense of privacy.

cs.CR↗

arXiv2025

Progress of the TianQin project

TianQin is a future space-based gravitational wave observatory targeting the frequency window of $10^{-4}$ Hz $\sim 1$ Hz. A large variety of gravitational wave sources are expected in this frequency band, including the merger of massive black hole binaries, the inspiral of extreme/intermediate mass ratio systems, stellar-mass black hole binaries, Galactic compact binaries, and so on. TianQin will consist of three Earth orbiting satellites on nearly identical orbits with orbital radii of about $10^5$ km. The satellites will form a normal triangle constellation whose plane is nearly perpendicular to the ecliptic plane. The TianQin project has been progressing smoothly following the ``0123" technology roadmap. In step ``0", the TianQin laser ranging station has been constructed and it has successfully ranged to all the five retro-reflectors on the Moon. In step ``1", the drag-free control technology has been tested and demonstrated using the TianQin-1 satellite. In step ``2", the inter-satellite laser interferometry technology will be tested using the pair of TianQin-2 satellites. The TianQin-2 mission has been officially approved and the satellites will be launched around 2026. In step ``3", i.e., the TianQin-3 mission, three identical satellites will be launched around 2035 to form the space-based gravitational wave detector, TianQin, and to start gravitational wave detection in space.

gr-qc↗

arXiv2025

Observation of quantum temporal correlations well beyond Luders bound

Quantum correlations, both spatial and temporal, are the central pillars of quantum mechanics. Over the last two decades, a big breakthrough in quantum physics is its complex extension to the non-Hermitian realm, and dizzying varieties of novel phenomena and applications beyond the Hermitian framework have been uncovered. However, unique features of non-Hermitian quantum correlations, especially in the time domain, still remain to be explored. Here, for the first time, we experimentally achieve this goal by using a parity-time (PT)-symmetric trapped-ion system. The upper limit of temporal quantum correlations, known as the algebraic bound, which has so far not been achieved in the standard measurement scenario, is reached here by approaching the exceptional point (EP), thus showing the unexpected ability of EPs in tuning temporal quantum correlation effects. Our study, unveiling the fundamental interplay of non-Hermiticity, nonlinearity, and temporal quantum correlations, provides the first step towards exploring and utilizing various non-Hermitian temporal quantum effects by operating a wide range of EP devices, which are important for both fundamental studies and applications of quantum EP systems.

quant-ph↗

arXiv2025

Towards the generation of petawatt near-infrared few-cycle light pulses via forward Raman amplification in plasma

Light amplification towards extremely high power in the infrared regime remains a significant challenge due to the lack of suitable gain media. Here we propose a new scheme to amplify a laser pulse with tunable wavelengths towards extremely high power via forward Raman amplification in plasma. Different from those previously proposed schemes based upon backward Raman or Brillouin amplification, our scheme involves a pump pulse and a seed pulse co-propagating in moderate density plasma, with the phase matching conditions for forward Raman scattering fulfilled. Due to their group velocity difference in plasma, the pump with a shorter wavelength and longer duration will chase the seed and transfer energy to the latter efficiently. Analytical models both for linear and nonlinear stages of amplification as well as particle-in-cell simulation show that by employing a 1.0 $\mathrm{μm}$ pump laser, a 1.8 $\mathrm{μm}$ seed pulse can be amplified $10^4$ times in its intensity, and then self-compressed to near-single-cycle. Our scheme shows the merits of high efficiency, high compactness, and relatively easy implementation with the co-propagating configuration, which may provide a unique route towards the petawatt few-cycle infrared laser pulses.

physics.plasm-ph↗

arXiv2025

Memory Offloading for Large Language Model Inference with Latency SLO Guarantees

Offloading large language models (LLMs) state to host memory during inference promises to reduce operational costs by supporting larger models, longer inputs, and larger batch sizes. However, the design of existing memory offloading mechanisms does not take latency service-level objectives (SLOs) into consideration. As a result, they either lead to frequent SLO violations or underutilize host memory, thereby incurring economic loss and thus defeating the purpose of memory offloading. This paper presents Select-N, a latency-SLO-aware memory offloading system for LLM serving. A key challenge in designing Select-N is to reconcile the tension between meeting SLOs and maximizing host memory usage. Select-N overcomes it by exploiting a unique characteristic of modern LLMs: during serving, the computation time of each decoder layer is deterministic. Leveraging this, Select-N introduces offloading interval, an internal tunable knob that captures the tradeoff between SLOs and host memory usage, thereby reducing the aforementioned challenge to pick an optimal offloading interval. With that, Select-N proposes a two-stage approach to automatically pick the offloading interval. The first stage is offline that generates the range of optimal offloading interval, while the second stage adjusts offloading interval at the granularity of inference iteration based on runtime hardware status. Our evaluation shows that Select-N consistently meets SLOs and improves the serving throughput over existing mechanisms by 1.85X due to maximizing the use of host memory.

cs.DC↗

arXiv2025

Spindle: Efficient Distributed Training of Multi-Task Large Models via Wavefront Scheduling

Recent foundation models are capable of handling multiple tasks and multiple data modalities with the unified base model structure and several specialized model components. However, efficient training of such multi-task (MT) multi-modal (MM) models poses significant system challenges due to the sophisticated model architecture and the heterogeneous workloads of different tasks and modalities. In this paper, we propose Spindle, a brand new training system tailored for resource-efficient and high-performance training of MT MM models via wavefront scheduling. The key idea of Spindle is to decompose the model execution into waves and address the joint optimization problem sequentially, including both heterogeneity-aware workload parallelization and dependency-driven execution scheduling. We build our system and evaluate it on various MT MM models. Experiments demonstrate the superior performance and efficiency of Spindle, with speedup ratio up to 71% compared to state-of-the-art training systems.

cs.DC↗

arXiv2025

Related Knowledge Perturbation Matters: Rethinking Multiple Pieces of Knowledge Editing in Same-Subject

Knowledge editing has become a promising approach for efficiently and precisely updating knowledge embedded in large language models (LLMs). In this work, we focus on Same-Subject Editing, which involves modifying multiple attributes of a single entity to ensure comprehensive and consistent updates to entity-centric knowledge. Through preliminary observation, we identify a significant challenge: Current state-of-the-art editing methods struggle when tasked with editing multiple related knowledge pieces for the same subject. To address the lack of relevant editing data for identical subjects in traditional benchmarks, we introduce the $\text{S}^2\text{RKE}$(Same-Subject Related Knowledge Editing) benchmark. Our extensive experiments reveal that only mainstream locate-then-edit methods, such as ROME and MEMIT, exhibit "related knowledge perturbation," where subsequent edits interfere with earlier ones. Further analysis reveals that these methods over-rely on subject information, neglecting other critical factors, resulting in reduced editing effectiveness.

cs.CL↗

arXiv2025

Evaluating the Robustness of the "Ensemble Everything Everywhere" Defense

Ensemble everything everywhere is a defense to adversarial examples that was recently proposed to make image classifiers robust. This defense works by ensembling a model's intermediate representations at multiple noisy image resolutions, producing a single robust classification. This defense was shown to be effective against multiple state-of-the-art attacks. Perhaps even more convincingly, it was shown that the model's gradients are perceptually aligned: attacks against the model produce noise that perceptually resembles the targeted class. In this short note, we show that this defense is not robust to adversarial attack. We first show that the defense's randomness and ensembling method cause severe gradient masking. We then use standard adaptive attack techniques to reduce the defense's robust accuracy from 48% to 14% on CIFAR-100 and from 62% to 11% on CIFAR-10, under the $\ell_\infty$-norm threat model with $\varepsilon=8/255$.

cs.LG↗

arXiv2025

Warfare:Breaking the Watermark Protection of AI-Generated Content

AI-Generated Content (AIGC) is rapidly expanding, with services using advanced generative models to create realistic images and fluent text. Regulating such content is crucial to prevent policy violations, such as unauthorized commercialization or unsafe content distribution. Watermarking is a promising solution for content attribution and verification, but we demonstrate its vulnerability to two key attacks: (1) Watermark removal, where adversaries erase embedded marks to evade regulation, and (2) Watermark forging, where they generate illicit content with forged watermarks, leading to misattribution. We propose Warfare, a unified attack framework leveraging a pre-trained diffusion model for content processing and a generative adversarial network for watermark manipulation. Evaluations across datasets and embedding setups show that Warfare achieves high success rates while preserving content quality. We further introduce Warfare-Plus, which enhances efficiency without compromising effectiveness. The code can be found in https://github.com/GuanlinLee/warfare.

cs.CV↗

arXiv2025

Picky LLMs and Unreliable RMs: An Empirical Study on Safety Alignment after Instruction Tuning

Large language models (LLMs) have emerged as powerful tools for addressing a wide range of general inquiries and tasks. Despite this, fine-tuning aligned LLMs on smaller, domain-specific datasets, critical to adapting them to specialized tasks, can inadvertently degrade their safety alignment, even when the datasets are benign. This phenomenon makes models more susceptible to providing inappropriate responses. In this study, we systematically examine the factors contributing to safety alignment degradation in benign fine-tuning scenarios. Our analysis identifies three critical factors affecting aligned LLMs: answer structure, identity calibration, and role-play. Additionally, we evaluate the reliability of state-of-the-art reward models (RMs), which are often used to guide alignment processes. Our findings reveal that these RMs frequently fail to accurately reflect human preferences regarding safety, underscoring their limitations in practical applications. By uncovering these challenges, our work highlights the complexities of maintaining safety alignment during fine-tuning and offers guidance to help developers balance utility and safety in LLMs. Datasets and fine-tuning code used in our experiments can be found in https://github.com/GuanlinLee/llm_instruction_tuning.

cs.AI↗

arXiv2025

Magnetic Field-Constrained Ensemble Image Segmentation of Coronal Holes in Chromospheric Observations

Coronal Holes (CHs) are large-scale, low-density regions in the solar atmosphere which may expel high-speed solar wind streams that incite hazardous, geomagnetic storms. Coronal and solar wind models can predict these high-speed streams and the performance of the coronal model can be validated against segmented CH boundaries. We present a novel method named Sub-Transition Region Identification of Ensemble Coronal Holes (STRIDE-CH) to address prominent challenges in segmenting CHs with Extreme Ultraviolet (EUV) imagery. Ground-based, chromospheric He I 10830 Å line imagery and underlying Fe I photospheric magnetograms are revisited to disambiguate CHs from filaments and quiet Sun, overcome obscuration by coronal loops, and complement established methods in the community which use space-borne, coronal EUV observations. Classical computer vision techniques are applied to constrain the radiative and magnetic properties of detected CHs, produce an ensemble of boundaries, and compile these boundaries in a confidence map that quantifies the likelihood of CH presence throughout the solar disk. This method is science-enabling towards future studies of CH formation and variability from a mid-atmospheric perspective.

astro-ph.SR↗

arXiv2025

Learning-Based Joint Antenna Selection and Precoding Design for Cell-Free MIMO Networks

This paper considers a downlink cell-free multiple-input multiple-output (MIMO) network in which multiple multi-antenna access points (APs) serve multiple users via coherent joint transmission. In order to reduce the energy consumption by radio frequency components, each AP selects a subset of antennas for downlink data transmission after estimating the channel state information (CSI). We aim to maximize the sum spectral efficiency by jointly optimizing the antenna selection and precoding design. To alleviate the fronthaul overhead and enable real-time network operation, we propose a distributed scalable machine learning algorithm. In particular, at each AP, we deploy a convolutional neural network (CNN) for antenna selection and a graph neural network (GNN) for precoding design. Different from conventional centralized solutions that require a large amount of CSI and signaling exchange among the APs, the proposed distributed machine learning algorithm takes only locally estimated CSI as input. With well-trained learning models, it is shown that the proposed algorithm significantly outperforms the distributed baseline schemes and achieves a sum spectral efficiency comparable to its centralized counterpart.

cs.IT↗

arXiv2025

Collaboratively Self-supervised Video Representation Learning for Action Recognition

Considering the close connection between action recognition and human pose estimation, we design a Collaboratively Self-supervised Video Representation (CSVR) learning framework specific to action recognition by jointly factoring in generative pose prediction and discriminative context matching as pretext tasks. Specifically, our CSVR consists of three branches: a generative pose prediction branch, a discriminative context matching branch, and a video generating branch. Among them, the first one encodes dynamic motion feature by utilizing Conditional-GAN to predict the human poses of future frames, and the second branch extracts static context features by contrasting positive and negative video feature and I-frame feature pairs. The third branch is designed to generate both current and future video frames, for the purpose of collaboratively improving dynamic motion features and static context features. Extensive experiments demonstrate that our method achieves state-of-the-art performance on multiple popular video datasets.

cs.CV↗

arXiv2025

Leveraging Surgical Activity Grammar for Primary Intention Prediction in Laparoscopy Procedures

Surgical procedures are inherently complex and dynamic, with intricate dependencies and various execution paths. Accurate identification of the intentions behind critical actions, referred to as Primary Intentions (PIs), is crucial to understanding and planning the procedure. This paper presents a novel framework that advances PI recognition in instructional videos by combining top-down grammatical structure with bottom-up visual cues. The grammatical structure is based on a rich corpus of surgical procedures, offering a hierarchical perspective on surgical activities. A grammar parser, utilizing the surgical activity grammar, processes visual data obtained from laparoscopic images through surgical action detectors, ensuring a more precise interpretation of the visual information. Experimental results on the benchmark dataset demonstrate that our method outperforms existing surgical activity detectors that rely solely on visual features. Our research provides a promising foundation for developing advanced robotic surgical systems with enhanced planning and automation capabilities.

cs.RO↗

arXiv2025

Lateral migration and bouncing of a deformable bubble rising near a vertical wall. Part 2. Highly inertial regimes

The fate of deformable buoyancy-driven bubbles rising near a vertical wall under highly inertial conditions is investigated numerically. In the absence of path instability, simulations reveal that when the Galilei number, $Ga$, which represents the buoyancy-to-viscous force ratio, exceeds a critical value, bubbles escape from the near-wall region after one to two rounds of bouncing, while at smaller $Ga$ they perform periodic bounces without escaping. The escape mechanism is rooted in the vigorous rotational flow that forms around a bubble during its bounce at high enough $Ga$, resulting in a Magnus-like repulsive force capable of driving it away from the wall. Path instability takes place with bubbles whose Bond number, the buoyancy-to-capillary force ratio, exceeds a critical $Ga$-dependent value. Such bubbles may or may not escape from the wall region, depending on the competition between the classical repulsive wake-wall interaction mechanism and a specific wall-ward trapping mechanism. The latter results from the reduction of the bubble oblateness caused by the abrupt drop of the rise speed when the bubble-wall gap becomes very thin. Owing to this transient shape variation, bubbles exhibiting zigzagging motions with a large enough amplitude experience larger transverse drag and virtual mass forces when departing from the wall than when returning to it. With moderately oblate bubbles, i.e. in an intermediate Bond number range, this effect is large enough to counteract the repulsive interaction force, forcing such bubbles to perform a periodic zigzagging-like motion at a constant distance from the wall.

physics.flu-dyn↗

arXiv2025

Pre-train and Fine-tune: Recommenders as Large Models

In reality, users have different interests in different periods, regions, scenes, etc. Such changes in interest are so drastic that they are difficult to be captured by recommenders. Existing multi-domain learning can alleviate this problem. However, the structure of the industrial recommendation system is complex, the amount of data is huge, and the training cost is extremely high, so it is difficult to modify the structure of the industrial recommender and re-train it. To fill this gap, we consider recommenders as large pre-trained models and fine-tune them. We first propose the theory of the information bottleneck for fine-tuning and present an explanation for the fine-tuning technique in recommenders. To tailor for recommendation, we design an information-aware adaptive kernel (IAK) technique to fine-tune the pre-trained recommender. Specifically, we define fine-tuning as two phases: knowledge compression and knowledge matching and let the training stage of IAK explicitly approximate these two phases. Our proposed approach designed from the essence of fine-tuning is well interpretable. Extensive online and offline experiments show the superiority of our proposed method. Besides, we also share unique and important lessons we learned when deploying the method in a large-scale online platform. We also present the potential issues of fine-tuning techniques in recommendation systems and the corresponding solutions. The recommender with IAK technique has been deployed on the homepage of a billion-scale online food platform for several months and has yielded considerable profits in our business.

cs.IR↗

arXiv2025

The First Indoor Pathloss Radio Map Prediction Challenge

To encourage further research and to facilitate fair comparisons in the development of deep learning-based radio propagation models, in the less explored case of directional radio signal emissions in indoor propagation environments, we have launched the ICASSP 2025 First Indoor Pathloss Radio Map Prediction Challenge. This overview paper describes the indoor path loss prediction problem, the datasets used, the Challenge tasks, and the evaluation methodology. Finally, the results of the Challenge and a summary of the submitted methods are presented.

eess.SP↗