arXiv subjects

Xu Zhang

Publications and source records attributed to Xu Zhang.

At least 55 records · Page 3Linked to original sources

arXiv2026

Understanding In-Context Multimodal Jailbreaks via Posterior Reweighting

In-context learning (ICL) jailbreaks reveal a critical vulnerability in multimodal large language models (MLLMs): harmful demonstrations in the prompt can induce unsafe outputs without modifying model parameters. Despite extensive empirical evidence, existing work lacks a principled understanding of why such jailbreaks reliably succeed or how their effectiveness scales with context composition. We propose a posterior reweighting framework that models a safety-aligned MLLM as implicitly operating over competing behavioral modes, and interprets in-context demonstrations as inference-time evidence that dynamically shifts the model's posterior preference between safe and harmful behaviors. This view formalizes jailbreak as a process of evidence accumulation, yielding predictive scaling laws with respect to demonstration count, harmful ratio, adversarial strength, and semantic diversity. Guided by this framework, we introduce a posterior-aware inference-time defense that adaptively injects benign counter-evidence based on estimated risk, effectively suppressing harmful posterior drift while preserving model utility. Compared to existing in-context defenses, our method achieves a significantly improved robustness-utility trade-off under a fixed intervention budget. Together, our results establish posterior reweighting as a unifying and predictive framework for understanding and mitigating ICL jailbreak in MLLMs.

cs.CR↗

arXiv2026

ASTRA: Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment

Dynamic 3D scene reconstruction has made significant progress with multi-camera systems, often relying on temporally aligned observations across views. However, in real-world scenarios, temporal asynchrony among capturing devices remains a common limitation, leading to severe motion blur and geometric artifacts. Existing asynchronous reconstruction methods typically estimate temporal offsets through photometric supervision, but appearance matching provides weak temporal cues under large offsets and complex motions. We attribute this limitation to two critical issues: texture-induced collapse, where low-textured regions provide nearly vanishing alignment signals, and deformation-induced entanglement, where temporal errors are absorbed into distorted geometry or motion rather than being explicitly corrected. To address these issues, we propose ASTRA (Asynchronous Spatio-Temporal Reconstruction via Trajectory Alignment), a framework that introduces 2D motion trajectories as explicit, texture-robust supervision for asynchronous dynamic reconstruction. Instead of synchronizing cameras solely through rendered color residuals, ASTRA jointly optimizes temporal offsets and dynamic 3D representations by aligning the projected motion of reconstructed 3D points with observed 2D trajectories, while using dynamic and certainty masking to suppress unreliable trajectory constraints. Extensive experiments on different dynamic Gaussian Splatting backbones show that ASTRA preserves high-frequency spatial details and sustains strong robustness even under severe asynchrony with up to 25-frame offsets, achieving approximately 1.4 dB PSNR improvement, reducing temporal-offset MAE by 54.0%, and nearly quadrupling the synchronization success rate.

cs.CV↗

arXiv2026

Communication-Efficient Personalized Federated Learning via Layer-Wise Multi-Threshold Random Sketching

Personalized federated learning (PFL) is a promising paradigm for collaborative learning over distributed devices, where edge nodes collaboratively train personalized models without sharing raw data. Although PFL addresses data heterogeneity by learning client-specific models, it still suffers from substantial uplink and downlink communication costs when exchanging high-dimensional parameters in bandwidth-constrained systems. Recent one-bit methods achieve extreme compression, but they usually rely on a single thresholding rule applied to the whole model. This design has two limitations. First, it overlooks layer-wise differences in parameter distributions and quantization sensitivities. Second, a single threshold provides only coarse binary information and cannot capture fine-grained variations in parameter distributions. To address these issues, we propose a communication-efficient PFL framework via layer-wise multi-threshold random sketching. In the proposed method, each layer is assigned its own set of quantization thresholds, so that the compressed representation can adapt to layer-specific statistics while using multiple intervals to provide a finer low-bit description of sketched parameters. The proposed method supports bidirectional communication using compact low-bit sketches and improves the communication-accuracy tradeoff compared with existing one-bit compression approaches.

cs.LG↗

arXiv2026

Local spectroscopy of loop current order with individual magnetic atoms

Hidden ordered states--characterized by order parameters that elude conventional probes--pose a fundamental challenge for their identification in quantum materials. Recent experiments report evidence for time-reversal symmetry breaking orbital magnetic order and anomalous transport signatures in the $2a\times2a$ charge density wave state of the kagome metal CsV$_3$Sb$_5$ at a temperature $T<30\,$K. Theoretical analyses propose that a time-reversal symmetry breaking loop-current order could exist as the ground state of this charge density wave. However, this microscopic interpretation remains debated and experimentally unverified. In this work, we employ individual magnetic atoms as local quantum sensors to examine the quasiparticle excitations of the charge density wave in CsV$_3$Sb$_5$ with the scanning tunneling microscope. Our spectroscopic measurements show that the magnetic moment of Co induces a spatially localized $dI/dV$ peak inside the spectral gap of the charge density wave near the Fermi energy. Conducting temperature-dependent spectroscopy, we find that this spectral feature emerges at $T<30\,$K. By comparing our experimental observations with results of quantum many-body simulations and realistic tight-binding model calculations, we show that this spectroscopic signature can be naturally interpreted as a local flux defect in a loop current ordered state, arising from the Kondo coupling of the magnetic moment of Co with the loop current electrons. The excellent agreement between our experimental and theoretical results suggests the presence of loop-current order in the $2a\times2a$ charge density wave of CsV$_3$Sb$_5$ at $T<30\,$K. Our results provide a microscopic picture to the observation of time-reversal symmetry breaking orbital magnetism and anomalous transport signatures detected in measurements of the macroscopic material properties.

cond-mat.str-el↗

arXiv2026

Amplitude structure of $3\to 3$ scattering in a Mandelstam variable representation

We construct a dispersive representation of the relativistic $3\to 3$ scattering amplitude for three identical spinless particles in the $S$-wave. The two-particle subenergy, instead of the total energy, is used as the dispersive variable. This choice keeps the physical dispersive contour free of the kinematical cuts that complicate total-energy dispersion relations. By separating discontinuities across the two-particle subenergy cuts from that across the three-body cut, we derive a linear integral equation with one-particle exchange as the driving term. We further show that the solution satisfies three-body unitarity as a consequence of two-body unitarity, analyticity, and crossing symmetry. For pair-wise interactions, this representation can be rewritten into the form used for isobar-spectator scattering. Finally, we give prescriptions for contour deformation and the subtraction of poles in the two-body subsystem amplitudes, which continue the amplitude onto adjacent unphysical Riemann sheets and thus provide direct access to the analytic structure relevant for three-body resonance poles.

hep-ph↗

arXiv2026

Analytic discs and compactness of the $\bar\partial$-Neumann operator

We construct a bounded pseudoconvex complete Reinhardt domain $Ω$ with smooth boundary in $\mathbb{C}^3$ such that the $\bar\partial$-Neumann operator $N_1$ is compact although $bΩ$ contains an analytic disc and thus also fails Catlin's Property $(P)$ and McNeal's Property $(\tilde P)$. This example solves an open problem on compactness of the $\bar\partial$-Neumann operator in the negative.

math.CV↗

arXiv2026

Bilinear correlations in Fluctuating Gaussian States with anti-unitary symmetries

We study bilinear order-parameter correlations in a general class of sign problem-free systems that are described by what we call Fluctuating Gaussian States (FGS) with anti-unitary symmetries, where the weight of each Gaussian measurement in the space-time path integral is positive-definite. Supported by Monte Carlo simulations on fluctuating Gaussian fermionic and bosonic examples, we argue that such systems generally exhibit two types of broken-symmetry phases: a \emph{saddle point phase} and a \emph{non-local contraction phase}, with the competition between the two determined by the stability of the fluctuation saddle point of FGS. Due to the Fermi liquid instability in fermionic FGS with anti-unitary symmetries, we also discuss the possibility of representing fermionic FGS with a particular Projected Entangled-Pair State (PEPS) ansatz, which we argue a certain construction provides an efficient description for the saddle point phase, but not for the non-local contraction phase. Finally, we discuss the potential implication of our result in non-equilibrium systems and stabilizing a target bilinear order.

quant-ph↗

arXiv2026

Strange metal and Fermi arcs from disordering spin stripes

We revisit the effective theory for fluctuating spin stripes coupled to a Fermi surface, and consider the parameter regime where a spin nematic phase intervenes between the spin density wave state and the symmetric state. It is shown that adding potential disorder to this theory, which acts as an unconventional type of random-field disorder, naturally gives rise to a phase diagram containing a quantum critical point that is described by the universal theory of strange metals with spatial disorder in both the magnitude and sign of the electron-boson coupling term [A.A. Patel, H. Guo, I. Esterlis and S. Sachdev, Science 381, 790 (2023)]. One difference compared to the original theory, however, is that at non-zero temperatures the disordered spin-stripe model automatically self-averages over the sign of the coupling. We also study the effects of thermal fluctuations in a phenomenological model for the disordered spin density wave state, and find from Monte Carlo simulations that a short anti-ferromagnetic correlation length (order 4-5 lattice constants) already leads to pronounced Fermi arcs in the electronic spectral weight.

cond-mat.str-el↗

arXiv2026

Infinitely Many Binomial Coefficients of Deficiency One

A binomial coefficient $\binom{n}{k}$ is good if all its prime divisors exceed $k$, and its deficiency is the number of $k$-smooth integers in $(n-k,n]$. A short constructive proof of the existence of infinitely many good binomial coefficients with deficiency exactly one is given, which provides an answer to a problem of Erdős, Lacampagne, and Selfridge.

math.NT↗

arXiv2026

Rigidity for local holomorphic isometric maps from the complex unit ball to reducible bounded symmetric domains

Let $U$ be a connected open subset of the complex unit ball $B^n$, $n\geq2$, and let $F_i: U \to Ω_i$ be non-constant holomorphic maps into irreducible bounded symmetric domains. We prove that if $F=(F_1, \cdots, F_m)$ is a holomorphic isometric map from $U$ to the product of $Ω_i$, then every $F_i$ extends to a proper holomorphic isometric embedding up to a positive constant from $B^n$ into $Ω_i$.

math.CV↗

arXiv2026

Robust CurveMoE: Multi-Norm Adversarial Defense for Mixture-of-Experts Models via Mode Connectivity

Multi-norm adversarial defense aims to protect neural networks against perturbations defined by different norm constraints, but existing methods typically optimize competing robustness objectives within a single parameter configuration, leading to substantial training cost and unfavorable robustness trade-offs. We propose Robust CurveMoE, an efficient mixture-of-experts framework that connects models specialized for different perturbation norms through a low-loss path and exploits the complementary robustness profiles of models along this path. Robust CurveMoE derives clean and norm-specialized experts from robustness-constrained curve locations and selectively expertizes only influential layers, while sharing the remaining parameters across routing paths. To further reduce curve-construction cost, we introduce contribution-guided partial updating, which selects influential curve parameters using initialization-based gradient scores. We also theoretically bound the objective gap between partial and full curve optimization. Experiments on CIFAR-100 and ImageNet-100 with WideResNet and Vision Transformer architectures show that Robust CurveMoE consistently improves clean, norm-specific, and Union accuracy over MSD and ERMC. In particular, it improves Union accuracy by 2.37 and 2.13 percentage points over the strongest baseline on CIFAR-100 and ImageNet-100, respectively. Extensive ablations further validate the effectiveness of partial updating, selective expertization, and robustness-constrained expert selection.

cs.LG↗

arXiv2026

MACD: Multi-Agent Clinical Diagnosis with Self-Learned Knowledge for LLM

Large language models (LLMs) have shown promise in supporting medical diagnosis, with prompting-based methods offering a flexible and deployable means of capability enhancement. However, existing prompt engineering and multi-agent approaches often focus on optimizing single inferences, paying less attention to the accumulation of reusable experience from clinical practice, constraining their real-world applicability. To address this, this study proposes a novel Multi-Agent Clinical Diagnosis (MACD) framework, which allows LLMs to self-learn clinical knowledge via a multi-agent pipeline that summarizes, refines, and applies diagnostic insights, mirroring the professional development of human physicians. We further extend it to a MACD-human collaborative workflow, where multiple LLM-based diagnostician agents engage in iterative consultations, supported by a judge agent and human oversight for cases where agreement is not reached. The MIMIC-MACD cohort comprising 4,390 real-world patient cases across seven diseases is constructed, including 1,314 cases for knowledge learning and 3,076 held-out cases for evaluation. Across diverse open-weight LLMs, MACD significantly improves primary diagnostic accuracy, achieving an average improvement of 11.6 percentage points over established authoritative knowledge, while narrowing the performance gap between open-weight models and state-of-the-art LLMs. Furthermore, the MACD-human workflow yields an 18.3-percentage-point improvement over physician-only diagnosis on text-only vignettes, demonstrating the synergistic potential of human-AI collaboration. This work thus presents a scalable self-learning paradigm that bridges the gap between the intrinsic knowledge of LLMs and the demands of real-world clinical practice, advancing towards a reliable, interpretable, and deployable AI-assisted diagnosis.

cs.AI↗

arXiv2026

From Association to Causation: Improving Retrieval Precision of Retrieval-Augmented Generation via Causal Relations and an Attention Mechanism

Retrieval-Augmented Generation (RAG) grounds LLM generation on retrieved documents, but the standard terminal retrieval stage--dense-vector similarity, optionally followed by reranking--often returns documents that share keywords with the query without containing the needed information, a failure mode that grows with the knowledge base. We trace it to a conceptual gap: similarity captures only associational relations, whereas the documents that matter are linked to the query causally. We model the terminal retrieval stage with a causal graph grounded in Reichenbach's common cause principle: the keywords shared by the query and a retrieved document form a latent common cause A, and the document's residual keywords form a latent set B linking the document to the ideal output. Since a retrieved document is a collider (A -> d <- B), retrieval itself opens an associational path between the query and B, which licenses a training-free, attention-style re-scoring rule: the cosine similarity between the query embedding and the weighted centroid embedding of B. Unlike causality-enhanced RAG variants that model causal relations inside the knowledge content, our graph models the causal structure of the retrieval process itself. On a real 471-document enterprise knowledge base, the method promotes a relevant guideline from rank 6 to the top 3; on a controlled diagnostic corpus reproducing the keyword-stuffing regime, it improves the mean target rank from 2.88 to 1.25, while a trained cross-encoder reranker barely helps (2.63). Conversely, on three BEIR benchmarks the score underperforms the similarity baseline, delineating the applicability boundary: the method guards the keyword-stuffing regime of growing proprietary knowledge bases and complements neural rerankers; a corpus-level calibration gate selects the correct regime with >= 95% reliability. A fully local testbed demonstrates deployability.

cs.AI↗

arXiv2026

Broadband Chiral Primordial Gravitational Waves from Constant-roll Inflation in Parity-violating Symmetric Teleparallel Gravity

We investigate primordial gravitational waves (GWs) generated during constant-roll inflation in a parity-violating extension of symmetric teleparallel gravity. The parity-violating interactions leave the background evolution and linear scalar perturbations unchanged, while inducing velocity birefringence in the tensor sector. Consequently, one of the two circular polarization states undergoes tachyonic amplification, producing a strongly blue and nearly fully chiral tensor spectrum from the cosmic microwave background (CMB) to interferometer scales. We identify viable constant-roll parameter regions consistent with current CMB constraints and determine the largest coupling strength compatible with both CMB B-mode measurements and the LIGO-Virgo-KAGRA (LVK) O1--O4a bound on the stochastic GW background. The predicted CMB B-mode spectra may be detectable by LiteBIRD, while the enhanced high-frequency signal could be accessible to the LISA--Taiji network and LVK O5 run. The model also predicts nonvanishing TB and EB correlations. These results highlight the potential of combining CMB and multi-band GW observations to test parity-violating gravity during inflation.

astro-ph.CO↗

arXiv2026

Many Representations as Sums of Three Prime Cubes

Let $F_k(n)$ be the number of unordered representations \[ n=p_1^k+p_2^k+\cdots +p_k^k \] by primes, with repetitions allowed. Erdős stated that $\limsup_{n\to\infty} F_3(n)=\infty$, but his proof appears not to have been published. A complete unconditional proof is given. The principal input is the classical Hecke equidistribution theorem for the CM Fermat cubic; the rest of the argument uses standard estimates for primes in arithmetic progressions and elementary counting. The argument used for $k=3$ does not extend to the case $k=4$. Nevertheless, by applying the Green--Tao--Ziegler theorem to the linear forms arising from an admissible binary quartic identity, $\limsup_{n\to\infty}F_4(n)\ge2$ is shown.

math.NT↗

arXiv2026

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

Real-world deployment of Vision-Language-Action (VLA) models is often bottlenecked by efficiency-performance trade-offs, cross-embodiment generalization, and execution smoothness. We present NebulaVLA, an asynchronous dual-frequency architecture that decouples high-level semantic reasoning from low-level action control, optimizing computational resources and modularity. To bridge semantic gaps across heterogeneous robots, we introduce GESTURE-7, a unified language-grounded action representation. Furthermore, our Guide Action algorithm enforces kinematic continuity via mask-based smoothness constraints. Comprehensive evaluations demonstrate that NebulaVLA significantly outperforms synchronous baselines, achieving an 85.5\% average success rate on LIBERO-Plus and accelerating action generation by \textasciitilde 2.7$\times$. This asynchronous design enables highly efficient and responsive control for practical robotics.

cs.RO↗

arXiv2026

Adaptive Heterogeneous Compression for Resource-Efficient Federated Knowledge Distillation

Federated learning (FL) enables privacy-preserving distributed model training but faces challenges from heterogeneous model architectures and limited communication resources at the network edge. Federated knowledge distillation (FedKD) alleviates model heterogeneity by combining prototype-wise parameter aggregation and knowledge transfer across heterogeneous models. However, transmitting gradients still introduces considerable communication overhead, while existing compression approaches typically apply a uniform strategy across clients and ignore their diverse model characteristics and resource capacities. To address this issue, we propose a heterogeneous compression framework for FedKD that enables each client to select a compression strategy from a candidate strategy set. We formulate the compression strategy selection problem as a non-stationary stochastic multi-armed bandit (MAB), where each arm corresponds to a compression strategy. An efficiency-aware reward is designed by jointly considering local optimization improvement, global knowledge alignment, and execution time. Based on this formulation, we develop an Adaptive heterogeneouS Compression algorithm for fEderated kNowledge Distillation (ASCEND), which employs an exponential moving average (EMA)-enhanced $ε$-greedy policy to balance exploration and exploitation. Experimental results on multiple datasets demonstrate that ASCEND effectively adapts to heterogeneous model and resource settings, reducing communication overhead and training time while maintaining competitive model accuracy.

cs.DC↗

arXiv2026

Measures on General Codimensional Surfaces in Infinite Dimensions and Stokes-Type Theorems

In this paper, we give an explicit construction of surface measures on a class of surfaces with arbitrary, possibly infinite, codimension in $\ell^2$. These measures are constructed from local representations associated with a fixed Gaussian product measure. We then establish local and global Gauss--Green-type formulas, introduce a notion of top-degree differential form, and derive an associated Stokes-type identity. We also determine the orientation of the boundary induced by the orientation of the surface. Moreover, therelationship between $\mathcal F$-continuity and Borel measurability is examined,which reveals a phenomenon specific to the infinite-dimensional setting.

math.FA↗