arXiv subjects

Jun Zheng

Publications and source records attributed to Jun Zheng.

At least 37 records · Page 2Linked to original sources

arXiv2025

GRPO-Guard: Mitigating Implicit Over-Optimization in Flow Matching via Regulated Clipping

Recently, GRPO-based reinforcement learning has shown remarkable progress in optimizing flow-matching models, effectively improving their alignment with task-specific rewards. Within these frameworks, the policy update relies on importance-ratio clipping to constrain overconfident positive and negative gradients. However, in practice, we observe a systematic shift in the importance-ratio distribution-its mean falls below 1 and its variance differs substantially across timesteps. This left-shifted and inconsistent distribution prevents positive-advantage samples from entering the clipped region, causing the mechanism to fail in constraining overconfident positive updates. As a result, the policy model inevitably enters an implicit over-optimization stage-while the proxy reward continues to increase, essential metrics such as image quality and text-prompt alignment deteriorate sharply, ultimately making the learned policy impractical for real-world use. To address this issue, we introduce GRPO-Guard, a simple yet effective enhancement to existing GRPO frameworks. Our method incorporates ratio normalization, which restores a balanced and step-consistent importance ratio, ensuring that PPO clipping properly constrains harmful updates across denoising timesteps. In addition, a gradient reweighting strategy equalizes policy gradients over noise conditions, preventing excessive updates from particular timestep regions. Together, these designs act as a regulated clipping mechanism, stabilizing optimization and substantially mitigating implicit over-optimization without relying on heavy KL regularization. Extensive experiments on multiple diffusion backbones (e.g., SD3.5M, Flux.1-dev) and diverse proxy tasks demonstrate that GRPO-Guard significantly reduces over-optimization while maintaining or even improving generation quality.

cs.CV↗

arXiv2025

Local integral input-to-state stability for non-autonomous infinite-dimensional systems

In this paper, we prove comparison principles for nonlinear differential equations with time-varying coefficients and develop Lyapunov analytical tools for the integral input-to-state stability (iISS) analysis of nonlinear non-autonomous infinite-dimensional systems, which involve nonlinearities satisfying a superlinear growth, {bringing} difficulties to the iISS {analysis.} Specifically, our approach starts by establishing several forms of comparison principles for a wide range of ordinary differential equations having time-varying coefficients and superlinear terms, paving the way to conduct iISS assessment for general nonlinear non-autonomous infinite-dimensional systems within the Lyapunov stability framework. Then, by using the comparison principles, we prove a local {iISS} {(LiISS)} Lyapunov theorem for the nonlinear non-autonomous infinite-dimensional systems in the framework of Banach spaces. {Furthermore,} we provide sufficient conditions of the existence of a local iISS Lyapunonv functional (LiISS-LF) and construct LiISS-LFs for the systems in the framework of Hilbert spaces. Finally, we preset two examples to illustrate the proposed {Lyapunov} method for the LiISS analysis: one is to show how to obtain the LiISS of a nonlinear finite-dimensional system with time-varying coefficients and superlinear terms under linear state feedback control law while another one is to show how to employ the interpolation inequalities to handle superliner terms and establish the LiISS-LF for a class of multi-dimensional parabolic equations with space-time-varying coefficients. To demonstrate the validity of the results, numerical experiments are also conducted to verify the LiISS of these two classes of systems.

math.OC↗

arXiv2025

A minimum problem associated with scalar Ginzburg-Landau equation and free boundary

Let $N>2$, $p\in \left(\frac{2N}{N+2},+\infty\right)$, and $\Omega$ be an open bounded domain in $\mathbb{R}^N$. We consider the minimum problem $$ \mathcal{J} (u) := \displaystyle\int_{\Omega } \left(\frac{1}{p}| \nabla u| ^p+\lambda_1\left(1-(u^+)^2\right)^2+\lambda_2u^+\right)\text{d}x\rightarrow \text{min} $$ over a certain class $\mathcal{K}$, where $\lambda_1\geq 0$ and $ \lambda_2\in \mathbb{R}$ are constants, and $u^+:=\max\{u,0\}$. The corresponding Euler-Lagrange equation is related to the Ginzburg-Landau equation and involves a subcritical exponent when $\lambda_1>0$. For $\lambda_1\geq 0$ and $ \lambda_2\in \mathbb{R}$, we prove the existence, non-negativity, and uniform boundedness of minimizers of $\mathcal{J} (u) $. Then, we show that any minimizer is locally $C^{1,\alpha}$-continuous with some $\alpha\in (0,1)$ and admits the optimal growth $\frac{p}{p-1}$ near the free boundary. Finally, under the additional assumption that $\lambda_2>0$, we establish non-degeneracy for minimizers near the free boundary and show that there exists at least one minimizer for which the corresponding free boundary has finite ($N-1$)-dimensional Hausdorff measure.

math.AP↗

arXiv2025

WebNovelBench: Placing LLM Novelists on the Web Novel Distribution

Robustly evaluating the long-form storytelling capabilities of Large Language Models (LLMs) remains a significant challenge, as existing benchmarks often lack the necessary scale, diversity, or objective measures. To address this, we introduce WebNovelBench, a novel benchmark specifically designed for evaluating long-form novel generation. WebNovelBench leverages a large-scale dataset of over 4,000 Chinese web novels, framing evaluation as a synopsis-to-story generation task. We propose a multi-faceted framework encompassing eight narrative quality dimensions, assessed automatically via an LLM-as-Judge approach. Scores are aggregated using Principal Component Analysis and mapped to a percentile rank against human-authored works. Our experiments demonstrate that WebNovelBench effectively differentiates between human-written masterpieces, popular web novels, and LLM-generated content. We provide a comprehensive analysis of 24 state-of-the-art LLMs, ranking their storytelling abilities and offering insights for future development. This benchmark provides a scalable, replicable, and data-driven methodology for assessing and advancing LLM-driven narrative generation.

cs.CL↗

arXiv2025

WISA: World Simulator Assistant for Physics-Aware Text-to-Video Generation

Recent rapid advancements in text-to-video (T2V) generation, such as SoRA and Kling, have shown great potential for building world simulators. However, current T2V models struggle to grasp abstract physical principles and generate videos that adhere to physical laws. This challenge arises primarily from a lack of clear guidance on physical information due to a significant gap between abstract physical principles and generation models. To this end, we introduce the World Simulator Assistant (WISA), an effective framework for decomposing and incorporating physical principles into T2V models. Specifically, WISA decomposes physical principles into textual physical descriptions, qualitative physical categories, and quantitative physical properties. To effectively embed these physical attributes into the generation process, WISA incorporates several key designs, including Mixture-of-Physical-Experts Attention (MoPA) and a Physical Classifier, enhancing the model's physics awareness. Furthermore, most existing datasets feature videos where physical phenomena are either weakly represented or entangled with multiple co-occurring processes, limiting their suitability as dedicated resources for learning explicit physical principles. We propose a novel video dataset, WISA-32K, collected based on qualitative physical categories. It consists of 32,000 videos, representing 17 physical laws across three domains of physics: dynamics, thermodynamics, and optics. Experimental results demonstrate that WISA can effectively enhance the compatibility of T2V models with real-world physical laws, achieving a considerable improvement on the VideoPhy benchmark. The visual exhibitions of WISA and WISA-32K are available in the https://360cvgroup.github.io/WISA/.

cs.CV↗

arXiv2025

CatV2TON: Taming Diffusion Transformers for Vision-Based Virtual Try-On with Temporal Concatenation

Virtual try-on (VTON) technology has gained attention due to its potential to transform online retail by enabling realistic clothing visualization of images and videos. However, most existing methods struggle to achieve high-quality results across image and video try-on tasks, especially in long video scenarios. In this work, we introduce CatV2TON, a simple and effective vision-based virtual try-on (V2TON) method that supports both image and video try-on tasks with a single diffusion transformer model. By temporally concatenating garment and person inputs and training on a mix of image and video datasets, CatV2TON achieves robust try-on performance across static and dynamic settings. For efficient long-video generation, we propose an overlapping clip-based inference strategy that uses sequential frame guidance and Adaptive Clip Normalization (AdaCN) to maintain temporal consistency with reduced resource demands. We also present ViViD-S, a refined video try-on dataset, achieved by filtering back-facing frames and applying 3D mask smoothing for enhanced temporal consistency. Comprehensive experiments demonstrate that CatV2TON outperforms existing methods in both image and video try-on tasks, offering a versatile and reliable solution for realistic virtual try-ons across diverse scenarios.

cs.CV↗

arXiv2025

Determination and evaluation of the critical liquid nitrogen for superconducting levitator based on a novel temperature-weight coupling measurement device

Liquid nitrogen (LN2) is the only cooling medium for the high-temperature superconducting (HTS) bulks in the superconducting levitator, which is the heart of the maglev train, to reach working state. The detection and determination of the critical LN2 content are crucial for reliable operation of the HTS maglev train. However, the related intelligent detection model and technology is lack in the combination filed of the cryogenic environment and maglev application, and there is no existing method to detect the LN2 content in superconducting levitator. This paper proposes to employ multisensor fusion framework to fuse and enhance the accuracy of critical LN2 content testing. Four temperature sensors were deployed inside superconducting levitator to measure the temperature change during the LN2 content changing from 100 % to 0. It was first obtained that the critical LN2 content in the superconducting levitator is 4%. To accurately monitor the critical LN2 content in the superconducting levitator, a matrix-weighted information fusion Kalman filter algorithm was used. Compared with the previous single sensor method, the testing accuracy of the multisensor fusion method can be improved by 5.6%. The work can provide a preliminary research foundation for the online monitoring and fault diagnosis of HTS maglev train.

physics.ins-det↗

arXiv2024

Dynamic Try-On: Taming Video Virtual Try-on with Dynamic Attention Mechanism

Video try-on stands as a promising area for its tremendous real-world potential. Previous research on video try-on has primarily focused on transferring product clothing images to videos with simple human poses, while performing poorly with complex movements. To better preserve clothing details, those approaches are armed with an additional garment encoder, resulting in higher computational resource consumption. The primary challenges in this domain are twofold: (1) leveraging the garment encoder's capabilities in video try-on while lowering computational requirements; (2) ensuring temporal consistency in the synthesis of human body parts, especially during rapid movements. To tackle these issues, we propose a novel video try-on framework based on Diffusion Transformer(DiT), named Dynamic Try-On. To reduce computational overhead, we adopt a straightforward approach by utilizing the DiT backbone itself as the garment encoder and employing a dynamic feature fusion module to store and integrate garment features. To ensure temporal consistency of human body parts, we introduce a limb-aware dynamic attention module that enforces the DiT backbone to focus on the regions of human limbs during the denoising process. Extensive experiments demonstrate the superiority of Dynamic Try-On in generating stable and smooth try-on results, even for videos featuring complicated human postures.

cs.CV↗

arXiv2024

Finite-time input-to-state stability for infinite-dimensional systems

In this paper, we extend the notion of finite-time input-to-state stability (FTISS) for finite-dimensional systems to infinite-dimensional systems. More specifically, we first prove an FTISS Lyapunov theorem for a class of infinite-dimensional systems, namely, the existence of an FTISS Lyapunov functional (FTISS-LF) implies the FTISS of the system, and then, provide a sufficient condition for ensuring the existence of an FTISS-LF for a class of abstract infinite-dimensional systems under the framework of compact semigroup theory and Hilbert spaces. As an application of the FTISS Lyapunov theorem, we verify the FTISS for a class of parabolic PDEs involving sublinear terms and distributed in-domain disturbances. Since the nonlinear terms of the corresponding abstract system are not Lipschitz continuous, the well-posedness is proved based on the application of compact semigroup theory and the FTISS is assessed by using the Lyapunov method with the aid of an interpolation inequality. Numerical simulations are conducted to confirm the theoretical results.

math.OC↗

arXiv2024

Input-to-state stabilization of $1$-D parabolic equations with Dirichlet boundary disturbances under boundary fixed-time control

This paper addresses the problem of stabilization of $1$-D parabolic equations with destabilizing terms and Dirichlet boundary disturbances. By using the method of backstepping and the technique of splitting, a boundary feedback controller is designed to ensure the input-to-state stability (ISS) of the closed-loop system with Dirichlet boundary disturbances, while preserving fixed-time stability (FTS) of the corresponding disturbance-free system, for which the fixed time is either determined by the Riemann zeta function or freely prescribed. To overcome the difficulty brought by Dirichlet boundary disturbances, the ISS and FTS properties of the involved systems are assessed by applying the generalized Lyapunov method. Numerical simulations are conducted to illustrate the effectiveness of the proposed scheme of control design.

math.OC↗

arXiv2024

Input-to-State Stabilization of 1-D Parabolic PDEs under Output Feedback Control

This paper addresses the problem of input-to-state stabilization for a class of parabolic equations with time-varying coefficients, as well as Dirichlet and Robin boundary disturbances. By using time-invariant kernel functions, which can reduce the complexity in control design and implementation, an observer-based output feedback controller is designed via backstepping. By using the generalized Lyapunov method, which can be used to handle Dirichlet boundary terms, the input-to-state stability of the closed-loop system under output feedback control, as well as the state estimation error system, is established in the spatial $L^\infty$-norm. Numerical simulations are conducted to confirm the theoretical results and to illustrate the effectiveness of the proposed control scheme.

math.OC↗

arXiv2024

VITON-DiT: Learning In-the-Wild Video Try-On from Human Dance Videos via Diffusion Transformers

Video try-on stands as a promising area for its tremendous real-world potential. Prior works are limited to transferring product clothing images onto person videos with simple poses and backgrounds, while underperforming on casually captured videos. Recently, Sora revealed the scalability of Diffusion Transformer (DiT) in generating lifelike videos featuring real-world scenarios. Inspired by this, we explore and propose the first DiT-based video try-on framework for practical in-the-wild applications, named VITON-DiT. Specifically, VITON-DiT consists of a garment extractor, a Spatial-Temporal denoising DiT, and an identity preservation ControlNet. To faithfully recover the clothing details, the extracted garment features are fused with the self-attention outputs of the denoising DiT and the ControlNet. We also introduce novel random selection strategies during training and an Interpolated Auto-Regressive (IAR) technique at inference to facilitate long video generation. Unlike existing attempts that require the laborious and restrictive construction of a paired training dataset, severely limiting their scalability, VITON-DiT alleviates this by relying solely on unpaired human dance videos and a carefully designed multi-stage training strategy. Furthermore, we curate a challenging benchmark dataset to evaluate the performance of casual video try-on. Extensive experiments demonstrate the superiority of VITON-DiT in generating spatio-temporal consistent try-on results for in-the-wild videos with complicated human poses.

cs.CV↗

arXiv2024

Electromagnetic chirality-induced negative refraction with the same amplitude and anti-phase of the two chirality coefficients

We suggest a scheme of electromagnetic chirality-induced negative refraction utilizing magneto-electric cross coupling in a four-level atomic system. The negative refraction can be achieved with the two chirality coefficients having the same amplitude but the opposite phase,and without requiring the simultaneous presence of an electric-dipole and a magnetic-dipole transition near the same transition frequency. The simultaneously negative electric permittivity and magnetic permeability does not require, either.

quant-ph↗

arXiv2023

Power Tracking Control of Heterogeneous Populations of TCLs with Partially Measured States

This paper presents a new aggregate power tracking control scheme for populations of thermostatically controlled loads (TCLs). The control design is performed in the framework of partial differential equations (PDEs) based on a late-lumping procedure without truncating the infinite-dimensional model describing the dynamics of the TCL population. An input-output linearization control scheme, which is independent of system parameters and uses only partial state measurement, is derived, and a sliding model-like control is applied to achieve finite-time input-to-state stability for tracking error dynamics. Such a control strategy can ensure robust performance in the presence of modeling uncertainties, while considerably reducing the communication burden in large scale distributed systems similar to that considered in the present work. A rigorous analysis of the closed-loop stability of the underlying PDE system was conducted, which guaranteed the validity of the developed control scheme. Simulation studies were performed while considering two TCL populations with a significant difference in their size, and the results show that the developed control scheme performs well in both cases, thereby confirming the effectiveness of the proposed solution.

eess.SY↗

arXiv2023

RIFL: A Reliable Link Layer Network Protocol for Data Center Communication

More and more latency-sensitive services and applications are being deployed into the data center. Performance can be limited by the high latency of the network interconnect. Because the conventional network stack is designed not only for LAN, but also for WAN, it carries a great amount of redundancy that is not required in a data center network. This paper introduces the concept of a three-layer protocol stack that can fulfill the exact demands of data center network communications. The detailed design and implementation of the first layer of the stack, which we call RIFL, is presented. A novel low latency in-band hop-by-hop re-transmission protocol is proposed and adopted in RIFL, which guarantees lossless transmission in a data center environment. Experimental results show that RIFL achieves 110 nanoseconds point-to-point latency on 10-meter Active Optical Cables, at a line rate of 112 Gbps. RIFL is a multi-lane protocol with scalable throughput up to multi-hundred gigabits per second. It can be the enabler of low latency, high throughput, flexible, scalable, and lossless data center networks.

cs.NI↗

arXiv2023

Power Tracking Control of Heterogeneous TCL Populations with Modeling Uncertainties and Communication Restrictions

This paper presents a new aggregate power tracking control scheme for populations of thermostatically controlled loads (TCLs). The control design is carried out in the framework of partial differential equations (PDEs) based on a late-lumping procedure without truncating the infinite-dimensional model describing the dynamics of the TCL population. An input-output linearization control scheme, which is independent of the system parameters and uses only partial state measurement, is derived, and a sliding model control is applied, which allows achieving a finite-time input-to-state stability for the tracking error dynamics. Such a control strategy can ensure a robust performance in the presence of modeling uncertainties while considerably reducing the communication burden in large scale distributed systems as the one considered in the present work. To guarantee the validity of the developed control scheme, a rigourous analysis on the solutions to the underlying PDE is conducted. Two implementations of the proposed control strategy, based on discrete-time approximation and fuzzy logic control, respectively, are validated through simulation studies.

math.OC↗

arXiv2023

Proof of Proposition 3.1 in the paper titled "Backstepping control of a class of space-time-varying linear parabolic PDEs via time invariant kernel functions''

We provide a detailed proof of Proposition 3.1 in the paper titled ``Backstepping control of a class of space-time-varying linear parabolic PDEs via time invariant kernel functions''. In the paper titled ``Backstepping control of a class of space-time-varying linear parabolic PDEs via time invariant kernel functions'', we addressed the problem of exponential stabilization and continuous dependence of solutions on initial data in different norms for a class of $1$-D linear parabolic PDEs with space-time-varying coefficients under backstepping boundary control. In order to stabilize the system without involving a Gevrey-like condition or the event-triggered scheme, a boundary feedback controller was designed via a time invariant kernel function. By using the approximative Lyapunov method, the exponential stability of the closed-loop system was established in the spatial $L^{p}$-norm and $W^{1,p}$-norm, respectively, whenever $p\in [1, +\infty]$. It was also shown that the solution to the considered system depends continuously on the spatial $L^{p}$-norm and $W^{1,p}$-norm, respectively, of the initial data.

math.AP↗

arXiv2022

ABODE-Net: An Attention-based Deep Learning Model for Non-intrusive Building Occupancy Detection Using Smart Meter Data

Occupancy information is useful for efficient energy management in the building sector. The massive high-resolution electrical power consumption data collected by smart meters in the advanced metering infrastructure (AMI) network make it possible to infer buildings' occupancy status in a non-intrusive way. In this paper, we propose a deep leaning model called ABODE-Net which employs a novel Parallel Attention (PA) block for building occupancy detection using smart meter data. The PA block combines the temporal, variable, and channel attention modules in a parallel way to signify important features for occupancy detection. We adopt two smart meter datasets widely used for building occupancy detection in our performance evaluation. A set of state-of-the-art shallow machine learning and deep learning models are included for performance comparison. The results show that ABODE-Net significantly outperforms other models in all experimental cases, which proves its validity as a solution for non-intrusive building occupancy detection.

cs.LG↗