Selected work

Representative Papers

KAN-AD: Time Series Anomaly Detection with Kolmogorov-Arnold Networks

Nov 01, 2024arXiv.org

To address the high false-alarm rate in time-series anomaly detection (TSAD) for cloud services and web systems—caused by overfitting to minor fluctuations—this paper proposes a robust smoothing-based modeling framework. It replaces B-splines with truncated Fourier expansions to construct more stable local normal-pattern representations, designs a lightweight global-aware learning mechanism to enable global-local co-optimization, and builds upon the Kolmogorov–Arnold network (KAN) architecture with a parameter-efficient training strategy. The resulting model employs fewer than 1,000 parameters and achieves a 50% inference speedup over the baseline KAN. Evaluated on four standard benchmarks, it attains an average 15% improvement in detection accuracy (peaking at 27%) and demonstrates significantly enhanced robustness against noise and practical applicability.

4 citationsRead paper

ConLA: Contrastive Latent Action Learning from Human Videos for Robotic Manipulation

Jan 31, 2026

This work addresses the challenge of learning transferable robotic manipulation policies from human demonstration videos without explicit action labels, while avoiding shortcut learning and representation entanglement caused by reconstructing visual appearance. The authors propose an unsupervised pretraining framework that leverages a contrastive disentanglement mechanism, integrating action category priors with temporal dynamics to effectively separate motion semantics from visual content. This approach yields clean, semantically consistent latent action representations. Notably, it is the first method to surpass the performance of models pretrained on real robot trajectories when using only human videos for pretraining. Extensive experiments demonstrate its strong generalization and practical utility across multiple robotic manipulation benchmarks, significantly enhancing both the disentanglement and transferability of learned action representations.

1 citationsRead paper

Learned Intelligent Recognizer with Adaptively Customized RIS Phases in Communication Systems

May 05, 2025

In RIS-aided communication systems, the strong coupling between RIS phase configuration and neural network parameters hinders joint optimization. Method: This paper proposes an integrated communication-and-sensing framework featuring an iterative sensing architecture that tightly fuses LSTM networks with a physics-based channel model, enabling end-to-end joint optimization of RIS phase responses and deep neural network parameters. The method dynamically customizes RIS phase profiles in real time based on scene, task, and objective characteristics—without requiring additional time-frequency resources. Contribution/Results: Experiments demonstrate that the proposed approach significantly outperforms state-of-the-art methods in target recognition accuracy while preserving nearly all communication throughput. It achieves, for the first time, zero-overhead, real-time, high-accuracy embedded sensing—fully integrating sensing functionality into the communication infrastructure without performance trade-offs.

1 citationsRead paper

Reconfigurable Intelligent Surface Aided Integrated Communication and Localization with a Single Access Point

May 05, 2025

This work addresses the limited indoor localization accuracy of single-access-point (AP) systems by proposing a multi-reconfigurable intelligent surface (RIS)-assisted integrated sensing and communication (ISAC) framework. Methodologically, it introduces a two-stage channel estimation procedure to extract multipath parameters and the angle-of-arrival (AoA) of RIS-reflected paths; innovatively integrates RIS phase scanning, Newtonized orthogonal matching pursuit (OMP), and angular-domain pseudospectrum analysis to establish a novel joint angular-spectral modeling paradigm; and achieves high-precision localization via linear least squares. The key contribution is the first demonstration of centimeter-level localization under a single-AP architecture through joint RIS phase control and angular-spectral estimation—eliminating reliance on multiple base stations. Experimental results confirm that pseudospectrum resolution, number of probing configurations, and reference point count yield significant positive gains in localization accuracy.

1 citationsRead paper

REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent

Aug 30, 2026

为解决大型语言模型在资源受限环境下的部署问题，REAL-Q通过动态梯度下降和细粒度修正方法优化了后训练量化过程，减少了端到端KL散度。

0 citationsRead paper

Recent publications

Latest Papers

REAL-Q: E2E LLM Quantization via Dynamic Gradient Descent

Aug 30, 2026

为解决大型语言模型在资源受限环境下的部署问题，REAL-Q通过动态梯度下降和细粒度修正方法优化了后训练量化过程，减少了端到端KL散度。

0 citationsRead paper

NebulaVLA: A Dual-Frequency Vision-Language-Action Model With Guide Action for Robotic Manipulation

Aug 17, 2026

This study addresses critical bottlenecks in Vision-Language-Action (VLA) model deployment, including the efficiency-performance trade-off, limited cross-embodiment generalization, and jerky motion generation. To overcome these challenges, we propose an asynchronous dual-frequency architecture that decouples semantic reasoning from action control. Furthermore, we introduce GESTURE-7, a unified action representation, alongside a mask-guided smoothing constraint algorithm. Experimental evaluations on the LIBERO-Plus benchmark demonstrate that our method achieves an average success rate of 85.5% and accelerates action generation by approximately 2.7× compared to synchronous baselines. These results confirm that the proposed framework effectively enables efficient, smooth, and generalizable robotic manipulation, significantly outperforming existing approaches in both computational efficiency and operational robustness.

0 citationsRead paper

ReQuant: Fixed-Grid Discrete Refinement for Post-Training Quantization

Aug 07, 2026

This work addresses a key limitation in existing post-training quantization (PTQ) methods, which typically treat quantized integer weights as final and lack mechanisms for further refinement. The authors propose a novel backpropagation-free, fixed-grid discrete optimization approach that iteratively redistributes the weights of an already quantized model to minimize mean squared reconstruction error, all while preserving the original quantization format. By explicitly treating the quantized model as an optimizable discrete solution, this method introduces an initialization-agnostic, plug-and-play post-processing stage that overcomes the one-shot nature of conventional PTQ. Extensive experiments demonstrate consistent performance gains across diverse model architectures, bit-widths, and downstream tasks, with particularly pronounced improvements under low-bit settings and simple PTQ initializations—often approaching or even surpassing the accuracy of GPTQ.

0 citationsRead paper

Movable Subarray-Aided ISAC in Hybrid Near-Far Field Channels

Aug 03, 2026

This work addresses the challenge of channel modeling and performance optimization in integrated sensing and communication (ISAC) systems operating in a hybrid near–far-field regime, where the overall array resides in the near field while individual subarrays experience far-field conditions. To tackle this, the paper proposes a novel architecture based on movable subarrays (MSA). A hybrid near–far-field channel model is developed, and transmit beamforming along with subarray positions are jointly optimized to minimize the Cramér–Rao bound (CRB) for angle and distance estimation, subject to constraints on communication SINR, transmit power, and subarray mobility. The main contributions include the first introduction of MSA for hybrid-field scenarios, derivation of the equivalent Fisher information matrix and CRB, and formulation of a joint optimization framework balancing sensing accuracy and communication quality. An efficient alternating optimization algorithm—combining iterative rank-one penalty semidefinite relaxation, projected finite-difference block coordinate descent, and backtracking—is designed to solve the resulting non-convex problem. Simulations confirm the proposed model’s close agreement with spherical wave models and demonstrate that MSA significantly reduces CRB, thereby enhancing sensing performance.

0 citationsRead paper

PACE: Adaptive Budget Allocation for Time-Efficient Embodied Planning

Aug 03, 2026

This work addresses the high latency of large language models in embodied planning, where sequential reasoning impedes real-time performance. To overcome this limitation, the authors propose PACE, a novel framework featuring an interleaved think-execute pipeline architecture coupled with a dynamic reasoning budget allocation mechanism. This design enables parallelization of cognitive reasoning and action execution, adaptively scheduling computational resources according to execution time windows. Evaluated on the Robotouille benchmark using the Qwen3-8B-AWQ model, PACE achieves a task success rate of 10%, representing a 67% improvement over the ReAct+Think baseline, while accelerating reasoning by 6.9×. Notably, 66.8% of the reasoning time is effectively hidden within action execution windows.

0 citationsRead paper