arXiv subjects

Zihan Cheng

Publications and source records attributed to Zihan Cheng.

At least 19 recordsLinked to original sources

arXiv2026

HarmTrace: Anchor-Calibrated Decoupled Optimization for Fine-Grained Target Identification in Harmful Memes

Multimodal harmful meme detection is typically formulated as image--text harmfulness classification. A model may correctly predict harmfulness while misidentifying the attacked target or its supporting evidence. We therefore extend harmful meme detection with fine-grained target identification, asking what type of target is attacked, who is targeted, and where the target appears in the meme. The model predicts harmfulness for every meme and, for harmful memes, outputs the target category, target entity, textual mention, and visual region. To support this task, we introduce Meme3W, which unifies multiple public harmful meme datasets and provides human-verified annotations for harmful instances. We further introduce Joint Record Accuracy (JRA), a strict record-level metric requiring the harmfulness label and all target-identification fields to be jointly correct. Experiments with representative multimodal large language models reveal a substantial gap between harmfulness accuracy and JRA. To narrow this gap, we propose HarmTrace, an anchor-calibrated decoupled optimization framework. HarmTrace strengthens target-entity supervision through entity-aware supervised fine-tuning. It then applies Conditional Target-identification Policy Optimization (CTPO) to decouple harmfulness and target-identification advantages, restricting target-identification optimization to label-correct responses for harmful examples. CTPO uses a Virtual Positive Anchor (VPA) as a fully correct reference for target-identification advantage normalization. HarmTrace improves both JRA and harmfulness accuracy across the evaluated backbones, with JRA on the Qwen3-VL-8B backbone increasing from 17.58\% to 52.51\%. Our code is publicly available at https://github.com/llly1234/HarmTrace-for-Harmful-Memes.

cs.CV↗

arXiv2026

UniDFKD: A Unified Semantic Prior Framework for Architecture-Agnostic Data-Free Knowledge Distillation

Data-Free Knowledge Distillation (DFKD) transfers knowledge from a pretrained teacher model to a compact student model by synthesizing semantically informative data, eliminating the need for access to the original training dataset. Existing DFKD methods rely heavily on architecture-specific statistical priors (e.g., Batch Normalization statistics) to guide data synthesis, however, such architecture-dependent priors are often absent in modern architectures such as Vision Transformers (ViTs), resulting in degraded semantic quality of the synthesized data and consequently catastrophic performance degradation. In this paper, we propose \emph{UniDFKD}, a unified data-free knowledge distillation framework that replaces architecture-specific statistics with explicit, architecture-agnostic semantic priors. \emph{UniDFKD} governs the entire synthesis-distillation pipeline along three dimensions: (1) Categorical Semantic Conditioning (CSC) defines \emph{what} to synthesize by persistently modulating the generator with language-derived embeddings to capture semantic diversity; (2) Spatial Semantic Anchoring (SSA) dictates \emph{where} evidence belongs by anchoring the teacher's spatial attributions to a Gaussian prior; and (3) Spatial Semantic Distillation (SSD) controls \emph{how} knowledge is transferred by explicitly aligning teacher-student spatial evidence alongside predictions. Extensive experiments across CNNs and ViTs demonstrate that UniDFKD establishes a new state-of-the-art, outperforming existing methods by an average absolute margin of over 20\% in both homogeneous and heterogeneous settings.

cs.CV↗

arXiv2026

Quantum resource localizability transitions in deep thermalization

We investigate how quantum resource constraints affect deep thermalization, the emergence of universal local wavefunction distributions from partial measurements of a quantum many-body state. Quantum resources, such as non-stabilizerness (magic), coherence, asymmetry, imaginarity, and non-Gaussianity, are essential for quantum information processing, and constraints on their global abundance can reshape these emergent distributions. To address this question, we develop a unified framework for deep thermalization within general quantum resource theories (QRTs). Our central result is that QRTs fall into two classes: ``smoothly localizable'' (SL) QRTs, where the resource content of local post-measurement states changes continuously with the global resource density, set by the initial state and measurement basis, yielding continuously tunable wavefunction distributions; and ``threshold localizable'' (TL) QRTs, where the local resource content jumps discontinuously from minimal to near-maximal past a critical global resource threshold, producing a sharp transition between a resourceless, ``deep-ergodicity breaking'' distribution and a resourceful, maximally random one. We trace this SL-TL dichotomy to an information-theoretic mechanism, block sharpening: by viewing each QRT as coherence between blocks in Hilbert space, we show that the local resource content depends on the measurement's ability to collapse an initial superposition into a single resourceless block. Our theory is analytically tractable and quantitatively predicts the phase boundaries across all studied QRTs, which we validate with extensive numerical simulations. Finally, we highlight two consequences: a novel magic transition in zero-rate quantum error-correcting codes--previously believed to occur only at finite rates--and new implications for quantum resource certification protocols based on post-measurement state ensembles.

quant-ph↗

arXiv2026

Rethinking SAR ATR: A Target-Aware Frequency-Spatial Enhancement Framework with Noise-Resilient Knowledge Guidance

Synthetic aperture radar automatic target recognition (SAR ATR) is of considerable importance in marine navigation and disaster monitoring. However, the coherent speckle noise inherent in SAR imagery often obscures salient target features, leading to degraded recognition accuracy and limited model generalization. To address this issue, this paper proposes a target-aware frequency-spatial enhancement framework with noise-resilient knowledge guidance (FSCE) for SAR target recognition. The proposed framework incorporates a frequency-spatial shallow feature adaptive enhancement (DSAF) module, which processes shallow features through spatial multi-scale convolution and frequency-domain wavelet convolution. In addition, a teacher-student learning paradigm combined with an online knowledge distillation method (KD) is employed to guide the student network to focus more effectively on target regions, thereby enhancing its robustness to high-noise backgrounds. Through the collaborative optimization of attention transfer and noise-resilient representation learning, the proposed approach significantly improves the stability of target recognition under noisy conditions. Based on the FSCE framework, two network architectures with different performance emphases are developed: lightweight DSAFNet-M and high-precision DSAFNet-L. Extensive experiments are conducted on the MSTAR, FUSARShip and OpenSARShip datasets. The results show that DSAFNet-L achieves competitive or superior performance compared with various methods on three datasets; DSAFNet-M significantly reduces the model complexity while maintaining comparable accuracy. These results indicate that the proposed FSCE framework exhibits strong cross-model generalization.

cs.CV↗

arXiv2025

MAGIC: Achieving Superior Model Merging via Magnitude Calibration

The proliferation of pre-trained models has given rise to a wide array of specialised, fine-tuned models. Model merging aims to merge the distinct capabilities of these specialised models into a unified model, requiring minimal or even no additional training. A core objective of model merging is to ensure the merged model retains the behavioural characteristics of the specialised models, typically achieved through feature alignment. We identify that features consist of two critical components: direction and magnitude. Prior research has predominantly focused on directional alignment, while the influence of magnitude remains largely neglected, despite its pronounced vulnerability to perturbations introduced by common merging operations (e.g., parameter fusion and sparsification). Such perturbations to magnitude inevitably lead to feature deviations in the merged model from the specialised models, resulting in subsequent performance degradation. To address this, we propose MAGnItude Calibration (MAGIC), a plug-and-play framework that rectifies layer-wise magnitudes in feature and weight spaces, with three variants. Specifically, our Feature Space Calibration (FSC) realigns the merged model's features using a small set of unlabelled data, while Weight Space Calibration (WSC) extends this calibration to the weight space without requiring additional data. Combining these yields Dual Space Calibration (DSC). Comprehensive experiments demonstrate that MAGIC consistently boosts performance across diverse Computer Vision tasks (+4.3% on eight datasets) and NLP tasks (+8.0% on Llama) without additional training. Our code is available at: https://github.com/lyymuwu/MAGIC

cs.LG↗

arXiv2025

PRISM: Precision-Recall Informed Data-Free Knowledge Distillation via Generative Diffusion

Data-free knowledge distillation (DFKD) transfers knowledge from a teacher to a student without access to the real in-distribution (ID) data. While existing methods perform well on small-scale images, they suffer from mode collapse when synthesizing large-scale images, resulting in limited knowledge transfer. Recently, leveraging advanced generative models to synthesize photorealistic images has emerged as a promising alternative. Nevertheless, directly using off-the-shelf diffusion to generate datasets faces the precision-recall challenges: 1) ensuring synthetic data aligns with the real distribution, and 2) ensuring coverage of the real ID manifold. In response, we propose PRISM, a precision-recall informed synthesis method. Specifically, we introduce Energy-guided Distribution Alignment to avoid the generation of out-of-distribution samples, and design the Diversified Prompt Engineering to enhance coverage of the real ID manifold. Extensive experiments on various large-scale image datasets demonstrate the superiority of PRISM. Moreover, we demonstrate that models trained with PRISM exhibit strong domain generalization.

cs.CV↗

arXiv2025

Simulation of bilayer Hamiltonians based on monitored quantum trajectories

In the study of open quantum systems it is often useful to treat mixed states as pure states of a fictitious doubled system. In this work we explore the opposite approach: mapping isolated bilayer systems to open monolayer systems. Specifically, we show that arbitrary bilayer Hamiltonians possessing an antiunitary layer exchange symmetry, and subject to a constraint on the sign of interlayer couplings, can be mapped to Lindbladians on a monolayer system with some of the jump operators postselected on a fixed outcome ("monitored"). Low-energy states of the bilayer Hamiltonian then correspond to late-time states of the monolayer dynamics. Simulating the latter by quantum trajectory methods has the potential of substantially reducing the computational cost of estimating low-energy observables in the bilayer Hamiltonian by effectively halving the system size. The overhead due to sampling quantum trajectories can be controlled by a suitable importance sampling scheme. We show that, when the quantum trajectories exhibit free fermion dynamics, our approach reduces to the auxiliary field quantum Monte Carlo (AFQMC) method. This provides a physically transparent interpretation of the AFQMC sign-free criteria in terms of properties of quantum dynamics. Finally, we benchmark our approach on the 1D quantum Ashkin-Teller model.

quant-ph↗

arXiv2025

Diffusion Once and Done: Degradation-Aware LoRA for Efficient All-in-One Image Restoration

Diffusion models have revealed powerful potential in all-in-one image restoration (AiOIR), which is talented in generating abundant texture details. The existing AiOIR methods either retrain a diffusion model or fine-tune the pretrained diffusion model with extra conditional guidance. However, they often suffer from high inference costs and limited adaptability to diverse degradation types. In this paper, we propose an efficient AiOIR method, Diffusion Once and Done (DOD), which aims to achieve superior restoration performance with only one-step sampling of Stable Diffusion (SD) models. Specifically, multi-degradation feature modulation is first introduced to capture different degradation prompts with a pretrained diffusion model. Then, parameter-efficient conditional low-rank adaptation integrates the prompts to enable the fine-tuning of the SD model for adapting to different degradation types. Besides, a high-fidelity detail enhancement module is integrated into the decoder of SD to improve structural and textural details. Experiments demonstrate that our method outperforms existing diffusion-based restoration approaches in both visual quality and inference efficiency.

cs.CV↗

arXiv2025

A Survey on Agent Workflow -- Status and Future

In the age of large language models (LLMs), autonomous agents have emerged as a powerful paradigm for achieving general intelligence. These agents dynamically leverage tools, memory, and reasoning capabilities to accomplish user-defined goals. As agent systems grow in complexity, agent workflows-structured orchestration frameworks-have become central to enabling scalable, controllable, and secure AI behaviors. This survey provides a comprehensive review of agent workflow systems, spanning academic frameworks and industrial implementations. We classify existing systems along two key dimensions: functional capabilities (e.g., planning, multi-agent collaboration, external API integration) and architectural features (e.g., agent roles, orchestration flows, specification languages). By comparing over 20 representative systems, we highlight common patterns, potential technical challenges, and emerging trends. We further address concerns related to workflow optimization strategies and security. Finally, we outline open problems such as standardization and multimodal integration, offering insights for future research at the intersection of agent design, workflow infrastructure, and safe automation.

cs.AI↗

arXiv2025

UniLDiff: Unlocking the Power of Diffusion Priors for All-in-One Image Restoration

All-in-One Image Restoration (AiOIR) has emerged as a promising yet challenging research direction. To address the core challenges of diverse degradation modeling and detail preservation, we propose UniLDiff, a unified framework enhanced with degradation- and detail-aware mechanisms, unlocking the power of diffusion priors for robust image restoration. Specifically, we introduce a Degradation-Aware Feature Fusion (DAFF) to dynamically inject low-quality features into each denoising step via decoupled fusion and adaptive modulation, enabling implicit modeling of diverse and compound degradations. Furthermore, we design a Detail-Aware Expert Module (DAEM) in the decoder to enhance texture and fine-structure recovery through expert routing. Extensive experiments across multi-task and mixed degradation settings demonstrate that our method consistently achieves state-of-the-art performance, highlighting the practical potential of diffusion priors for unified image restoration. Our code will be released.

cs.CV↗

arXiv2025

Mamba-Sea: A Mamba-based Framework with Global-to-Local Sequence Augmentation for Generalizable Medical Image Segmentation

To segment medical images with distribution shifts, domain generalization (DG) has emerged as a promising setting to train models on source domains that can generalize to unseen target domains. Existing DG methods are mainly based on CNN or ViT architectures. Recently, advanced state space models, represented by Mamba, have shown promising results in various supervised medical image segmentation. The success of Mamba is primarily owing to its ability to capture long-range dependencies while keeping linear complexity with input sequence length, making it a promising alternative to CNNs and ViTs. Inspired by the success, in the paper, we explore the potential of the Mamba architecture to address distribution shifts in DG for medical image segmentation. Specifically, we propose a novel Mamba-based framework, Mamba-Sea, incorporating global-to-local sequence augmentation to improve the model's generalizability under domain shift issues. Our Mamba-Sea introduces a global augmentation mechanism designed to simulate potential variations in appearance across different sites, aiming to suppress the model's learning of domain-specific information. At the local level, we propose a sequence-wise augmentation along input sequences, which perturbs the style of tokens within random continuous sub-sequences by modeling and resampling style statistics associated with domain shifts. To our best knowledge, Mamba-Sea is the first work to explore the generalization of Mamba for medical image segmentation, providing an advanced and promising Mamba-based architecture with strong robustness to domain shifts. Remarkably, our proposed method is the first to surpass a Dice coefficient of 90% on the Prostate dataset, which exceeds previous SOTA of 88.61%. The code is available at https://github.com/orange-czh/Mamba-Sea.

cs.CV↗

arXiv2025

Hardness of observing strong-to-weak symmetry breaking

Spontaneous symmetry breaking (SSB) is the cornerstone of our understanding of quantum phases of matter. Recent works have generalized this concept to the domain of mixed states in open quantum systems, where symmetries can be realized in two distinct ways dubbed strong and weak. Novel intrinsically mixed phases of quantum matter can then be defined by the spontaneous breaking of strong symmetry down to weak symmetry. However, proposed order parameters for strong-to-weak SSB (based on mixed-state fidelities or purities) seem to require exponentially many copies of the state, raising the question: is it possible to efficiently detect strong-to-weak SSB in general? Here we answer this question negatively in the paradigmatic cases of $Z_2$ and $U(1)$ symmetries. We construct ensembles of pseudorandom mixed states that do not break the strong symmetry, yet are computationally indistinguishable from states that do. This rules out the existence of efficient state-agnostic protocols to detect strong-to-weak SSB.

quant-ph↗

arXiv2024

Emergent unitary designs for encoded qubits from coherent errors and syndrome measurements

Unitary $k$-designs are distributions of unitary gates that match the Haar distribution up to its $k$-th statistical moment. They are a crucial resource for randomized quantum protocols. However, their implementation on encoded logical qubits is nontrivial due to the need for magic gates, which can require a large resource overhead. In this work, we propose an efficient approach to generate unitary designs for encoded qubits in surface codes by applying local unitary rotations ("coherent errors") on the physical qubits followed by syndrome measurement and error correction. We prove that under some conditions on the coherent errors (notably including all single-qubit unitaries) and on the error correcting code, this process induces a unitary transformation of the logical subspace. We numerically show that the ensemble of logical unitaries (indexed by the random syndrome outcomes) converges to a unitary design in the thermodynamic limit, provided the density or strength of coherent errors is above a finite threshold. This "unitary design" phase transition coincides with the code's coherent error threshold under optimal decoding. Furthermore, we propose a classical algorithm to simulate the protocol based on a "staircase" implementation of the surface code encoder and decoder circuits. This enables a mapping to a 1+1D monitored circuit, where we observe an entanglement phase transition (and thus a classical complexity phase transition of the decoding algorithm) coinciding with the aforementioned unitary design phase transition. Our results provide a practical way to realize unitary designs on encoded qubits, with applications including quantum state tomography and benchmarking in error correcting codes.

quant-ph↗

arXiv2024

Pseudoentanglement from tensor networks

Pseudoentangled states are defined by their ability to hide their entanglement structure: they are indistinguishable from random states to any observer with polynomial resources, yet can have much less entanglement than random states. Existing constructions of pseudoentanglement based on phase- and/or subset-states are limited in the entanglement structures they can hide: e.g., the states may have low entanglement on a single cut, on all cuts at once, or on local cuts in one dimension. Here we introduce new constructions of pseudoentangled states based on (pseudo)random tensor networks that affords much more flexibility in the achievable entanglement structures. We illustrate our construction with the simplest example of a matrix product state, realizable as a staircase circuit of pseudorandom unitary gates, which exhibits pseudo-area-law scaling of entanglement in one dimension. We then generalize our construction to arbitrary tensor network structures that admit an isometric realization. A notable application of this result is the construction of pseudoentangled `holographic' states whose entanglement entropy obeys a Ryu-Takayanagi `minimum-cut' formula, answering a question posed in [Aaronson et al., arXiv:2211.00747].

quant-ph↗

arXiv2023

Universal structure of measurement-induced information in many-body ground states

Unlike unitary dynamics, measurements of a subsystem can induce long-range entanglement via quantum teleportation. The amount of measurement-induced entanglement or mutual information depends jointly on the measurement basis and the entanglement structure of the state (before measurement), and has operational significance for whether the state is a resource for measurement-based quantum computing, as well as for the computational complexity of simulating the state using quantum or classical computers. In this work, we examine entropic measures of measurement-induced entanglement (MIE) and information (MII) for the ground-states of quantum many-body systems in one- and two- spatial dimensions. From numerical and analytic analysis of a variety of models encompassing critical points, quantum Hall states, string-net topological orders, and Fermi liquids, we identify universal features of the long-distance structure of MIE and MII that depend only on the underlying phase or critical universality class of the state. We argue that, whereas in $1d$ the leading contributions to long-range MIE and MII are universal, in $2d$, the existence of a teleportation transition for finite-depth circuits implies that trivial $2d$ states can exhibit long-range MIE, and the universal features lie in sub-leading corrections. We introduce modified MIE measures that directly extract these universal contributions. As a corollary, we show that the leading contributions to strange-correlators, used to numerically identify topological phases, are in fact non-universal in two or more dimensions, and explain how our modified constructions enable one to isolate universal components. We discuss the implications of these results for classical- and quantum- computational simulation of quantum materials.

quant-ph↗

arXiv2023

Efficient sampling of noisy shallow circuits via monitored unraveling

We introduce a classical algorithm for sampling the output of shallow, noisy random circuits on two-dimensional qubit arrays. The algorithm builds on the recently-proposed "space-evolving block decimation" (SEBD) and extends it to the case of noisy circuits. SEBD is based on a mapping of 2D unitary circuits to 1D {\it monitored} ones, which feature measurements alongside unitary gates; it exploits the presence of a measurement-induced entanglement phase transition to achieve efficient (approximate) sampling below a finite critical depth $T_c$. Our noisy-SEBD algorithm unravels the action of noise into measurements, further lowering entanglement and enabling efficient classical sampling up to larger circuit depths. We analyze a class of physically-relevant noise models (unital qubit channels) within a two-replica statistical mechanics treatment, finding weak measurements to be the optimal (i.e. most disentangling) unraveling. We then locate the noisy-SEBD complexity transition as a function of circuit depth and noise strength in realistic circuit models. As an illustrative example, we show that circuits on heavy-hexagon qubit arrays with noise rates of $\approx 2\%$ per CNOT, based on IBM Quantum processors, can be efficiently sampled up to a depth of 5 iSWAP (or 10 CNOT) gate layers. Our results help sharpen the requirements for practical hardness of simulation of noisy hardware.

quant-ph↗

arXiv2022

Characterizing a non-equilibrium phase transition on a quantum computer

At transitions between phases of matter, physical systems can exhibit universal behavior independent of their microscopic details. Probing such behavior in quantum many-body systems is a challenging and practically important problem that can be solved by quantum computers, potentially exponentially faster than by classical computers. In this work, we use the Quantinuum H1-1 quantum computer to realize a quantum extension of a simple classical disease spreading process that is known to exhibit a non-equilibrium phase transition between an active and absorbing state. Using techniques such as qubit-reuse and error avoidance based on real-time conditional logic (utilized extensively in quantum error correction), we are able to implement large instances of the model with $73$ sites and up to $72$ circuit layers, and quantitatively determine the model's critical properties. This work demonstrates how quantum computers capable of mid-circuit resets, measurements, and conditional logic enable the study of difficult problems in quantum many-body physics: the simulation of open quantum system dynamics and non-equilibrium phase transitions.

quant-ph↗

arXiv2022

A matrix product operator approach to non-equilibrium Floquet steady states

We present a numerical method to simulate non-equilibrium Floquet steady states of one-dimensional periodically-driven (Floquet) many-body systems coupled to a dissipative bath, called open-system Floquet DMRG (OFDMRG). This method is based on a matrix product operator ansatz for the Floquet density matrix in frequency-space, and enables access to large systems beyond the reach of exact master-equation or quantum trajectory simulations, while retaining information about the periodic micro-motion in Floquet steady states. An excited-state extension of this technique also allows computation of the dynamical approach to the steady state on asymptotically long timescales. We benchmark the OFDMRG approach with a driven-dissipative Ising model, and apply it to study the possibility of dissipatively stabilizing pre-thermal discrete time-crystalline order by coupling to a cold bath.

quant-ph↗