arXiv subjects

Yu Cui

Publications and source records attributed to Yu Cui.

At least 37 records · Page 2Linked to original sources

arXiv2025

End-to-End Dexterous Arm-Hand VLA Policies via Shared Autonomy: VR Teleoperation Augmented by Autonomous Hand VLA Policy for Efficient Data Collection

Achieving human-like dexterous manipulation remains a major challenge for general-purpose robots. While Vision-Language-Action (VLA) models show potential in learning skills from demonstrations, their scalability is limited by scarce high-quality training data. Existing data collection methods face inherent constraints: manual teleoperation overloads human operators, while automated planning often produces unnatural motions. We propose a Shared Autonomy framework that divides control between macro and micro motions. A human operator guides the robot's arm pose through intuitive VR teleoperation, while an autonomous DexGrasp-VLA policy handles fine-grained hand control using real-time tactile and visual feedback. This division significantly reduces cognitive load and enables efficient collection of high-quality coordinated arm-hand demonstrations. Using this data, we train an end-to-end VLA policy enhanced with our novel Arm-Hand Feature Enhancement module, which captures both distinct and shared representations of macro and micro movements for more natural coordination. Our Corrective Teleoperation system enables continuous policy improvement through human-in-the-loop failure recovery. Experiments demonstrate that our framework generates high-quality data with minimal manpower and achieves a 90% success rate across diverse objects, including unseen instances. Comprehensive evaluations validate the system's effectiveness in developing dexterous manipulation capabilities.

cs.RO↗

arXiv2025

HatLLM: Hierarchical Attention Masking for Enhanced Collaborative Modeling in LLM-based Recommendation

Recent years have witnessed a surge of research on leveraging large language models (LLMs) for sequential recommendation. LLMs have demonstrated remarkable potential in inferring users' nuanced preferences through fine-grained semantic reasoning. However, they also exhibit a notable limitation in effectively modeling collaborative signals, i.e., behavioral correlations inherent in users' historical interactions. Our empirical analysis further reveals that the attention mechanisms in LLMs tend to disproportionately focus on tokens within the same item, thereby impeding the capture of cross-item correlations. To address this limitation, we propose a novel hierarchical attention masking strategy for LLM-based recommendation, termed HatLLM. Specifically, in shallow layers, HatLLM masks attention between tokens from different items, facilitating intra-item semantic understanding; in contrast, in deep layers, HatLLM masks attention within items, thereby compelling the model to capture cross-item correlations. This progressive, layer-wise approach enables LLMs to jointly model both token-level and item-level dependencies. Extensive experiments on three real-world datasets demonstrate that HatLLM achieves significant performance gains (9.13% on average) over existing LLM-based methods.

cs.IR↗

arXiv2025

VortexPIA: Indirect Prompt Injection Attack against LLMs for Efficient Extraction of User Privacy

Large language models (LLMs) have been widely deployed in Conversational AIs (CAIs), while exposing privacy and security threats. Recent research shows that LLM-based CAIs can be manipulated to extract private information from human users, posing serious security threats. However, the methods proposed in that study rely on a white-box setting that adversaries can directly modify the system prompt. This condition is unlikely to hold in real-world deployments. The limitation raises a critical question: can unprivileged attackers still induce such privacy risks in practical LLM-integrated applications? To address this question, we propose \textsc{VortexPIA}, a novel indirect prompt injection attack that induces privacy extraction in LLM-integrated applications under black-box settings. By injecting token-efficient data containing false memories, \textsc{VortexPIA} misleads LLMs to actively request private information in batches. Unlike prior methods, \textsc{VortexPIA} allows attackers to flexibly define multiple categories of sensitive data. We evaluate \textsc{VortexPIA} on six LLMs, covering both traditional and reasoning LLMs, across four benchmark datasets. The results show that \textsc{VortexPIA} significantly outperforms baselines and achieves state-of-the-art (SOTA) performance. It also demonstrates efficient privacy requests, reduced token consumption, and enhanced robustness against defense mechanisms. We further validate \textsc{VortexPIA} on multiple realistic open-source LLM-integrated applications, demonstrating its practical effectiveness.

cs.CR↗

arXiv2025

Free-MAD: Consensus-Free Multi-Agent Debate

Multi-agent debate (MAD) is an emerging approach to improving the reasoning capabilities of large language models (LLMs). Existing MAD methods rely on multiple rounds of interaction among agents to reach consensus, and the final output is selected by majority voting in the last round. However, this consensus-based design faces several limitations. First, multiple rounds of communication increases token overhead and limits scalability. Second, due to the inherent conformity of LLMs, agents that initially produce correct responses may be influenced by incorrect ones during the debate process, causing error propagation. Third, majority voting introduces randomness and unfairness in the decision-making phase, and can degrade the reasoning performance. To address these issues, we propose \textsc{Free-MAD}, a novel MAD framework that eliminates the need for consensus among agents. \textsc{Free-MAD} introduces a novel score-based decision mechanism that evaluates the entire debate trajectory rather than relying on the last round only. This mechanism tracks how each agent's reasoning evolves, enabling more accurate and fair outcomes. In addition, \textsc{Free-MAD} reconstructs the debate phase by introducing anti-conformity, a mechanism that enables agents to mitigate excessive influence from the majority. Experiments on eight benchmark datasets demonstrate that \textsc{Free-MAD} significantly improves reasoning performance while requiring only a single-round debate and thus reducing token costs. We also show that compared to existing MAD approaches, \textsc{Free-MAD} exhibits improved robustness in real-world attack scenarios.

cs.AI↗

arXiv2025

OptiProxy-NAS: Optimization Proxy based End-to-End Neural Architecture Search

Neural architecture search (NAS) is a hard computationally expensive optimization problem with a discrete, vast, and spiky search space. One of the key research efforts dedicated to this space focuses on accelerating NAS via certain proxy evaluations of neural architectures. Different from the prevalent predictor-based methods using surrogate models and differentiable architecture search via supernetworks, we propose an optimization proxy to streamline the NAS as an end-to-end optimization framework, named OptiProxy-NAS. In particular, using a proxy representation, the NAS space is reformulated to be continuous, differentiable, and smooth. Thereby, any differentiable optimization method can be applied to the gradient-based search of the relaxed architecture parameters. Our comprehensive experiments on $12$ NAS tasks of $4$ search spaces across three different domains including computer vision, natural language processing, and resource-constrained NAS fully demonstrate the superior search results and efficiency. Further experiments on low-fidelity scenarios verify the flexibility.

cs.LG↗

arXiv2025

MAD-Spear: A Conformity-Driven Prompt Injection Attack on Multi-Agent Debate Systems

Multi-agent debate (MAD) systems leverage collaborative interactions among large language models (LLMs) agents to improve reasoning capabilities. While recent studies have focused on increasing the accuracy and scalability of MAD systems, their security vulnerabilities have received limited attention. In this work, we introduce MAD-Spear, a targeted prompt injection attack that compromises a small subset of agents but significantly disrupts the overall MAD process. Manipulated agents produce multiple plausible yet incorrect responses, exploiting LLMs' conformity tendencies to propagate misinformation and degrade consensus quality. Furthermore, the attack can be composed with other strategies, such as communication attacks, to further amplify its impact by increasing the exposure of agents to incorrect responses. To assess MAD's resilience under attack, we propose a formal definition of MAD fault-tolerance and develop a comprehensive evaluation framework that jointly considers accuracy, consensus efficiency, and scalability. Extensive experiments on five benchmark datasets with varying difficulty levels demonstrate that MAD-Spear consistently outperforms the baseline attack in degrading system performance. Additionally, we observe that agent diversity substantially improves MAD performance in mathematical reasoning tasks, which challenges prior work suggesting that agent diversity has minimal impact on performance. These findings highlight the urgent need to improve the security in MAD design.

cs.CR↗

arXiv2025

RRTL: Red Teaming Reasoning Large Language Models in Tool Learning

While tool learning significantly enhances the capabilities of large language models (LLMs), it also introduces substantial security risks. Prior research has revealed various vulnerabilities in traditional LLMs during tool learning. However, the safety of newly emerging reasoning LLMs (RLLMs), such as DeepSeek-R1, in the context of tool learning remains underexplored. To bridge this gap, we propose RRTL, a red teaming approach specifically designed to evaluate RLLMs in tool learning. It integrates two novel strategies: (1) the identification of deceptive threats, which evaluates the model's behavior in concealing the usage of unsafe tools and their potential risks; and (2) the use of Chain-of-Thought (CoT) prompting to force tool invocation. Our approach also includes a benchmark for traditional LLMs. We conduct a comprehensive evaluation on seven mainstream RLLMs and uncover three key findings: (1) RLLMs generally achieve stronger safety performance than traditional LLMs, yet substantial safety disparities persist across models; (2) RLLMs can pose serious deceptive risks by frequently failing to disclose tool usage and to warn users of potential tool output risks; (3) CoT prompting reveals multi-lingual safety vulnerabilities in RLLMs. Our work provides important insights into enhancing the security of RLLMs in tool learning.

cs.CL↗

arXiv2025

Practical Reasoning Interruption Attacks on Reasoning Large Language Models

Reasoning large language models (RLLMs) have demonstrated outstanding performance across a variety of tasks, yet they also expose numerous security vulnerabilities. Most of these vulnerabilities have centered on the generation of unsafe content. However, recent work has identified a distinct "thinking-stopped" vulnerability in DeepSeek-R1: under adversarial prompts, the model's reasoning process ceases at the system level and produces an empty final answer. Building upon this vulnerability, researchers developed a novel prompt injection attack, termed reasoning interruption attack, and also offered an initial analysis of its root cause. Through extensive experiments, we verify the previous analyses, correct key errors based on three experimental findings, and present a more rigorous explanation of the fundamental causes driving the vulnerability. Moreover, existing attacks typically require over 2,000 tokens, impose significant overhead, reduce practicality, and are easily detected. To overcome these limitations, we propose the first practical reasoning interruption attack. It succeeds with just 109 tokens by exploiting our newly uncovered "reasoning token overflow" (RTO) effect to overwrite the model's final answer, forcing it to return an invalid response. Experimental results demonstrate that our proposed attack is highly effective. Furthermore, we discover that the method for triggering RTO differs between the official DeepSeek-R1 release and common unofficial deployments. As a broadened application of RTO, we also construct a novel jailbreak attack that enables the transfer of unsafe content within the reasoning tokens into final answer, thereby exposing it to the user. Our work carries significant implications for enhancing the security of RLLMs.

cs.CR↗

arXiv2025

Token-Efficient Prompt Injection Attack: Provoking Cessation in LLM Reasoning via Adaptive Token Compression

While reasoning large language models (LLMs) demonstrate remarkable performance across various tasks, they also contain notable security vulnerabilities. Recent research has uncovered a "thinking-stopped" vulnerability in DeepSeek-R1, where model-generated reasoning tokens can forcibly interrupt the inference process, resulting in empty responses that compromise LLM-integrated applications. However, existing methods triggering this vulnerability require complex mathematical word problems with long prompts--even exceeding 5,000 tokens. To reduce the token cost and formally define this vulnerability, we propose a novel prompt injection attack named "Reasoning Interruption Attack", based on adaptive token compression. We demonstrate that simple standalone arithmetic tasks can effectively trigger this vulnerability, and the prompts based on such tasks exhibit simpler logical structures than mathematical word problems. We develop a systematic approach to efficiently collect attack prompts and an adaptive token compression framework that utilizes LLMs to automatically compress these prompts. Experiments show our compression framework significantly reduces prompt length while maintaining effective attack capabilities. We further investigate the attack's performance via output prefix and analyze the underlying causes of the vulnerability, providing valuable insights for improving security in reasoning LLMs.

cs.CR↗

arXiv2025

Process or Result? Manipulated Ending Tokens Can Mislead Reasoning LLMs to Ignore the Correct Reasoning Steps

Recent reasoning large language models (LLMs) have demonstrated remarkable improvements in mathematical reasoning capabilities through long Chain-of-Thought. The reasoning tokens of these models enable self-correction within reasoning chains, enhancing robustness. This motivates our exploration: how vulnerable are reasoning LLMs to subtle errors in their input reasoning chains? We introduce "Compromising Thought" (CPT), a vulnerability where models presented with reasoning tokens containing manipulated calculation results tend to ignore correct reasoning steps and adopt incorrect results instead. Through systematic evaluation across multiple reasoning LLMs, we design three increasingly explicit prompting methods to measure CPT resistance, revealing that models struggle significantly to identify and correct these manipulations. Notably, contrary to existing research suggesting structural alterations affect model performance more than content modifications, we find that local ending token manipulations have greater impact on reasoning outcomes than structural changes. Moreover, we discover a security vulnerability in DeepSeek-R1 where tampered reasoning tokens can trigger complete reasoning cessation. Our work enhances understanding of reasoning robustness and highlights security considerations for reasoning-intensive applications.

cs.AI↗

arXiv2025

Ramp Up NTT in Record Time using GPU-Accelerated Algorithms and LLM-based Code Generation

Homomorphic encryption (HE) is a core building block in privacy-preserving machine learning (PPML), but HE is also widely known as its efficiency bottleneck. Therefore, many GPU-accelerated cryptographic schemes have been proposed to improve the performance of HE. However, these methods often require complex modifications tailored to specific algorithms and are tightly coupled with specific GPU and operating systems. It is interesting to ask how to generally offer more practical GPU-accelerated cryptographic algorithm implementations. Given the powerful code generation capabilities of large language models (LLMs), we aim to explore their potential to automatically generate practical GPU-friendly algorithm code using CPU-friendly code. In this paper, we focus on number theoretic transform (NTT) -- the core mechanism of HE. We first develop and optimize a GPU-friendly NTT (GNTT) family that exploits PyTorch's fast matrix computation and precomputation, achieving an approximately 62x speedup -- a significant boost over existing ones. Then we explore GPU-friendly code generation using various LLMs, including DeepSeek-R1, OpenAI o1 and o3-mini. We discover many interesting findings throughout the process. For instance, somewhat surprisingly, our experiments demonstrate that DeepSeek-R1 significantly outperforms OpenAI o3-mini and o1, but still cannot beat our optimized protocol. The findings provide valuable insights for turbocharging PPML and enhancing code generation capabilities of LLMs. Codes are available at: https://github.com/LMPC-Lab/GenGPUCrypto.

cs.CR↗

arXiv2024

Diamond of Thought: A Design Thinking-Based Framework for LLMs in Wearable Design

Wearable design is an interdisciplinary field that balances technological innovation, human factors, and human-computer interactions. Despite contributions from various disciplines, many projects lack stable interdisciplinary teams, which often leads to design failures. Large language models (LLMs) integrate diverse information and generate innovative solutions, making them a valuable tool for enhancing design processes. Thus, we have explored the use of LLMs in wearable design by combining design-thinking principles with LLM capabilities. We have developed the "Diamond of Thought" framework and analysed 1,603 prototypes and 1,129 products from a body-centric perspective to create a comprehensive database. We employed retrieval-augmented generation to input database details into the LLMs, ensuring applicability to wearable design challenges and integration of embodied cognition into the process. Our LLM-based methodology for wearables has been experimentally validated, demonstrating the potential of LLMs for the advancement of design practices. This study offers new tools and methods for future wearable designs.

cs.HC↗

arXiv2024

Curvature perturbations from kinetic preheating after $α$-attractor inflation

Preheating at the end of inflation is a violent nonlinear process that efficiently transfers the energy of the inflaton to a second field, the preheat field. When the preheat field is light during inflation and its background value modulates the preheating process, the superhorizon isocurvature perturbations of the preheat field may be converted to curvature perturbations that leave an imprint on the cosmic microwave background and the large-scale structure of the universe. We use high-precision lattice simulations to study kinetic preheating after $α$-attractor inflation, a case where the effective mass of the preheat field is naturally suppressed during inflation. By comparing the expansion e-folds between different Hubble patches, we find that the conversion from isocurvature perturbations to curvature perturbations is very inefficient and can hardly be detected by cosmological observations.

astro-ph.CO↗

arXiv2024

Distillation Matters: Empowering Sequential Recommenders to Match the Performance of Large Language Model

Owing to their powerful semantic reasoning capabilities, Large Language Models (LLMs) have been effectively utilized as recommenders, achieving impressive performance. However, the high inference latency of LLMs significantly restricts their practical deployment. To address this issue, this work investigates knowledge distillation from cumbersome LLM-based recommendation models to lightweight conventional sequential models. It encounters three challenges: 1) the teacher's knowledge may not always be reliable; 2) the capacity gap between the teacher and student makes it difficult for the student to assimilate the teacher's knowledge; 3) divergence in semantic space poses a challenge to distill the knowledge from embeddings. To tackle these challenges, this work proposes a novel distillation strategy, DLLM2Rec, specifically tailored for knowledge distillation from LLM-based recommendation models to conventional sequential models. DLLM2Rec comprises: 1) Importance-aware ranking distillation, which filters reliable and student-friendly knowledge by weighting instances according to teacher confidence and student-teacher consistency; 2) Collaborative embedding distillation integrates knowledge from teacher embeddings with collaborative signals mined from the data. Extensive experiments demonstrate the effectiveness of the proposed DLLM2Rec, boosting three typical sequential models with an average improvement of 47.97%, even enabling them to surpass LLM-based recommenders in some cases.

cs.IR↗

arXiv2024

Deep Mamba Multi-modal Learning

Inspired by the excellent performance of Mamba networks, we propose a novel Deep Mamba Multi-modal Learning (DMML). It can be used to achieve the fusion of multi-modal features. We apply DMML to the field of multimedia retrieval and propose an innovative Deep Mamba Multi-modal Hashing (DMMH) method. It combines the advantages of algorithm accuracy and inference speed. We validated the effectiveness of DMMH on three public datasets and achieved state-of-the-art results.

cs.MM↗

arXiv2024

Adaptive Confidence Multi-View Hashing for Multimedia Retrieval

The multi-view hash method converts heterogeneous data from multiple views into binary hash codes, which is one of the critical technologies in multimedia retrieval. However, the current methods mainly explore the complementarity among multiple views while lacking confidence learning and fusion. Moreover, in practical application scenarios, the single-view data contain redundant noise. To conduct the confidence learning and eliminate unnecessary noise, we propose a novel Adaptive Confidence Multi-View Hashing (ACMVH) method. First, a confidence network is developed to extract useful information from various single-view features and remove noise information. Furthermore, an adaptive confidence multi-view network is employed to measure the confidence of each view and then fuse multi-view features through a weighted summation. Lastly, a dilation network is designed to further enhance the feature representation of the fused features. To the best of our knowledge, we pioneer the application of confidence learning into the field of multimedia retrieval. Extensive experiments on two public datasets show that the proposed ACMVH performs better than state-of-the-art methods (maximum increase of 3.24%). The source code is available at https://github.com/HackerHyper/ACMVH.

cs.CV↗

arXiv2023

DexRepNet: Learning Dexterous Robotic Grasping Network with Geometric and Spatial Hand-Object Representations

Robotic dexterous grasping is a challenging problem due to the high degree of freedom (DoF) and complex contacts of multi-fingered robotic hands. Existing deep reinforcement learning (DRL) based methods leverage human demonstrations to reduce sample complexity due to the high dimensional action space with dexterous grasping. However, less attention has been paid to hand-object interaction representations for high-level generalization. In this paper, we propose a novel geometric and spatial hand-object interaction representation, named DexRep, to capture dynamic object shape features and the spatial relations between hands and objects during grasping. DexRep comprises Occupancy Feature for rough shapes within sensing range by moving hands, Surface Feature for changing hand-object surface distances, and Local-Geo Feature for local geometric surface features most related to potential contacts. Based on the new representation, we propose a dexterous deep reinforcement learning method to learn a generalizable grasping policy DexRepNet. Experimental results show that our method outperforms baselines using existing representations for robotic grasping dramatically both in grasp success rate and convergence speed. It achieves a 93% grasping success rate on seen objects and higher than 80% grasping success rates on diverse objects of unseen categories in both simulation and real-world experiments.

cs.RO↗

arXiv2023

The rotating excitons in two-dimensional materials: Valley Zeeman effect and chirality

We propose the rotational dynamics of the intralayer and interlayer excitons with their inherent momenta of inertia in the monolayer and bilayer transition metal dichalcogenides, respectively, where the new chirality of exciton is endowed by the rotational angular momentum, namely, the formations of left- and right-handed excitons at the +K and -K valleys, respectively. We find that angular momenta exchange between excitons and its surrounding phononic bath result in the large fluctuation of the effective g-factor and the asymmetry of valley Zeeman splitting observed in most recently experiments, both of which sensitively depend on the magnetic moments provided by the phononic environment. This rotating exciton model not only proposes a new controllable knob in valleytronics, but opens the door to explore the angular momentum exchange of the chiral quasiparticles with the many-body environment.

cond-mat.mes-hall↗