arXiv subjects

Yong Xu

Publications and source records attributed to Yong Xu.

At least 37 records · Page 2Linked to original sources

arXiv2026

GMoT: Gated Motion-Aware Tokenization for Fine-Grained Micro-Gesture Video Reasoning with Multimodal LLMs

Micro-gesture recognition demands the detection of fleeting, spatially localized movements that are frequently overwhelmed by dominant static appearances and background noise. While Multimodal Large Language Models (MLLMs) excel at general video understanding, they inherently struggle with subtle kinematics and often rely on static posture priors. To this end, we propose GMoT, a Gated Motion-Aware Tokenization module that explicitly distills sparse kinematic evidence into a compact sequence prior to temporal modeling. GMoT dynamically spotlights action-relevant regions via spatially weighted pooling, extracts adjacent-frame temporal differencing to capture precise motion energy, and adaptively fuses these cues into the visual stream using a conservatively initialized semantic gate. To transition from simple classification to evidence-grounded reasoning, we further introduce a progressive reward-guided policy refinement paradigm, supported by a semi-supervised annotation pipeline that generates anatomically focused captions. Beyond achieving the best Top-1 accuracy among the compared methods on iMiGUE (67.32\%) and SMG (73.11\%), improving the Qwen3-VL-8B baseline by +6.80 and +3.11 points, our framework introduces Body-Region Grounding (BRG) Recall as an anatomical-grounding proxy conditioned on correct predictions, together with an overlapping-label cross-domain transfer protocol between iMiGUE and SMG. Extensive evaluations demonstrate that our GMoT-augmented model improves in-domain accuracy, retains clear gains under label-preserving corruptions, and improves accuracy-oriented cross-domain transfer under explicit small-split caveats while maintaining high anatomical grounding in its generated rationales.

cs.CV↗

arXiv2026

Large and moderate deviation for rough slow-fast system with level 3 geometric rough path

This work is to give the large deviation principle (LDP) and moderate deviation principle (MDP) for a slow-fast system driven by the mixed fractional Brownian motion (FBM) (1/4,1/3) via the level-3 geometric rough path (RP). Firstly, we provide a different way to lift the translation of mixed FBM in the Cameron-Martin direction to the geometric RP. The LDP turns to the weak convergence of the controlled system by averaging the fast one. Lacking the invariant measure for the controlled fast one, a "replaced process" whose limiting measure could be decoupled from the control term and owes the exponential ergodicity is constructed. Here, the stability under level-3 RPs is established under more elaborate estimates. Besides, different from LDP we give the MDP, that requires more intricate bounded estimates of the deviation component. The MDP for level-2 RP system is recovered as a special case.

math.PR↗

arXiv2026

Gravitational Wave from Graviton Bremsstrahlung during Reheating

We revisit graviton production via Bremsstrahlung from the decay of the inflaton during inflationary reheating. Using two complementary computational techniques, we first show that such 3-body differential decay rates differ from previously reported results in the literature. We then compute the stochastic gravitational wave (GW) background that forms during the period of reheating, when the inflaton perturbatively decays with the radiative emission of gravitons. By computing the number of relativistic degrees of freedom in terms of $ΔN_\text{eff}$, we constrain the resulting GW energy density from BBN and CMB. Finally, we project current and future GW detector sensitivities in probing such a stochastic GW background, which typically peaks in the GHz to THz ballpark, opening up the opportunity to be detected with microwave cavities and space-based GW detectors.

hep-ph↗

arXiv2026

Central limit theorem for slow-fast system under mixed fractional Brownian motion

This work considers a type of slow-fast system, where the slow component is driven by fractional Brownian motion (FBM) with $H > 1/2$ and the fast component is a Markovian stationary process. Our solution mapping is defined based on the Young-Wiener sense, which is constructed via the stochastic sewing lemma. Then, we aim to show the fluctuation from the averaging limit by applying the Poisson PDE method. Unlike the case of standard Brownian motion, the Poisson PDE method must be developed to a non-Markovian fractional setting. The first part addresses the central limit theorem problem for a slow-fast system under small FBM, which is fully coupled with the fast varying one. Firstly, a Wiener-Young-Ito formula is constructed for the Wiener-Young-Ito integral. The behavior of the deviation component is related to solutions of Poisson PDEs associated with the Laplace operator of the fast process. The fast one is assumed to satisfy "stricter" Holder conditions related to the time scale parameter. The tightness is then derived through the Holder semi-norm of the fast one, the properties of the Poisson PDE solution, and a Gronwall-type result for a linear Young differential equation. The weak limit shown includes an extra Gaussian process. The second part is dedicated to the problem under a general FBM. In contrast to the former one, here the FBM integral term will be more difficult to bound due to the coupling between the regularity of the Poisson PDE solution and dependence on the unbounded Holder seminorm of the fast one.

math.PR↗

arXiv2026

ModaFlow: Modality-Aware Flow Matching for High-Fidelity Virtual Try-On

Image-based virtual try-on has emerged as a compelling task in e-commerce and augmented reality, yet existing methods struggle to simultaneously preserve fine garment semantics and adapt to diverse person body geometries under large clothing-body deformations. We present ModaFlow, a modality-aware flow-matching based framework for high-fidelity virtual try-on that achieves precise alignment between textual descriptions and garment appearance. Unlike prior methods that treat multimodal conditions uniformly, ModaFlow introduces a modality-aware guidance scheme: visual garment embeddings extracted by a pretrained image prompt adapter provide deterministic, persistent structural guidance, while textual embeddings generated from garment descriptions are controlled via classifier-free guidance (CFG) with adaptive scaling and zero-initialized velocity. To further enhance flow field accuracy, we propose two regularization losses, cosine similarity and perceptual flow discrimination, that jointly improve directional consistency and perceptual realism of the velocity field. Additionally, a mask manipulation strategy stochastically samples among box, transparent, and relaxed masks during training, simulating diverse occlusion scenarios and enabling robust inference under unpaired settings where only a box mask is available. Experiments show that ModaFlow achieves state-of-the-art results in both qualitative and quantitative evaluations, reducing FID by approximately 30% on paired and 20% on unpaired benchmarks.

cs.CV↗

arXiv2026

RSICCLLM: A Multimodal Large Language Model for Remote Sensing Image Change Captioning

Remote Sensing Image Change Captioning (RSICC) aims to describe changes between bi-temporal remote sensing images and holds significant research and application value. However, most existing methods rely on conventional deep learning architectures, and the limited model capacity constrains performance. Although large-model post-training techniques have achieved great success in general domains, their direct transfer to RSICC remains challenging due to data scarcity and the need for fine-grained change understanding. To address this, we propose RSICCLLM, the first post-training framework for large vision-language models in RSICC. Specifically, we design a data generation paradigm, release the instruction dataset RSICI, and establish a task-specific RSICC benchmark. We further introduce Difference-aware Supervised Fine-tuning to explicitly extract change representations and guide the model in perceiving and understanding temporal differences. In addition, we propose Dual-Negative Preference Optimization (DNPO), which employs two complementary negative-sample construction strategies to construct the preference dataset RSICP and further refine model performance. Extensive experiments validate the superior capability of RSICCLLM, which achieves outstanding results with only 7B parameters, surpassing models of substantially larger scales. The code and dataset will be made publicly available at https://github.com/keaill/RSICCLLM.

cs.CV↗

arXiv2026

Towards Understanding The Calibration Benefits of Sharpness-Aware Minimization

Deep neural networks have been increasingly used in safety-critical applications such as medical diagnosis and autonomous driving. However, many studies suggest that they are prone to being poorly calibrated and have a propensity for overconfidence, which may have disastrous consequences. In this paper, unlike standard training such as stochastic gradient descent, we show that the recently proposed sharpness-aware minimization (SAM) counteracts this tendency towards overconfidence. The theoretical analysis suggests that SAM allows us to learn models that are already well-calibrated by implicitly maximizing the entropy of the predictive distribution. Inspired by this finding, we further propose a variant of SAM, coined as CSAM, to ameliorate model calibration. Extensive experiments on various datasets, including ImageNet-1K, demonstrate the benefits of SAM in reducing calibration error. Meanwhile, CSAM performs even better than SAM and consistently achieves lower calibration error than other approaches

cs.LG↗

arXiv2026

Navigating User Behavior toward Personalized Multimodal Generation

Modern AIGC pipelines deliver high-fidelity images and videos but presuppose a well-formed creation instruction, while end users rarely articulate visual details, leaving generators misaligned with user demand. We study personalized content generation, which turns a user's interaction history into an executable instruction for downstream synthesis, and identify two obstacles: behavior must be encoded in a form legible to language reasoning, and the model must acquire instruction-writing skill absent from both pretraining and behavior data. We propose NaviGen, which represents each item with a dual identifier coupling a collaborative code and a textual code as a behavioral substrate and a semantic bridge in one token stream. On this representation, a two-stage SFT+RL pipeline first distills preference reasoning and instruction writing from evolutionarily searched supervision, then aligns generation with user intent through hierarchical and self-consistent rewards. Experiments across product, game, and short-video domains show that NaviGen improves personalized image and video generation, strengthens next-item prediction, and yields more specific, relevant, and visually generatable instructions. Our code is released at: https://github.com/iLearn-Lab/NaviGen.

cs.AI↗

arXiv2026

AI-accelerated metallized $σ$-bonding screening for superconductor discovery

The computational discovery of phonon-mediated superconductors is hindered by the prohibitive cost of density functional perturbation theory (DFPT). Here, guided by the metallized $σ$-bonding picture, we introduce the $σ$-bonding density of states ($σ$DOS) as an efficient physical descriptor to identify high-transition-temperature ($T_{\mathrm{c}}$) superconductors from density functional theory (DFT)-level electronic structure without explicit DFPT calculations. The evaluation of $σ$DOS can be further accelerated by a deep-learning DFT Hamiltonian method, enabling efficient large-scale screening for superconductors. Screening 2 million materials, we identify B$_{13}$Se as an ambient-pressure superconductor candidate with predicted $T_{\mathrm{c}} > 40$~K, together with a family of high-$T_{\mathrm{c}}$ B$_{13}X$ candidates, supporting the effectiveness of this discovery strategy. By bridging physics priors with AI acceleration, this study delivers an efficient and generalizable route for computational materials discovery in the AI era.

physics.comp-ph↗

arXiv2026

A deep learning framework for jointly solving transient Fokker-Planck equations with arbitrary parameters and initial distributions

Efficiently solving the Fokker-Planck equation (FPE) is central to analyzing complex parameterized stochastic systems. However, current numerical methods lack parallel computation capabilities across varying conditions, severely limiting comprehensive parameter exploration and transient analysis. This paper introduces a deep learning-based pseudo-analytical probability solution (PAPS) that, via a single training process, simultaneously resolves transient FPE solutions for arbitrary multi-modal initial distributions, system parameters, and time points. The core idea is to unify initial, transient, and stationary distributions via Gaussian mixture distributions (GMDs) and develop a constraint-preserving autoencoder that bijectively maps constrained GMD parameters to unconstrained, low-dimensional latent representations. In this representation space, the panoramic transient dynamics across varying initial conditions and system parameters can be modeled by a single evolution network. Extensive experiments on paradigmatic systems demonstrate that the proposed PAPS maintains high accuracy while achieving inference speeds four orders of magnitude faster than GPU-accelerated Monte Carlo simulations. This efficiency leap enables previously intractable real-time parameter sweeps and systematic investigations of stochastic bifurcations. By decoupling representation learning from physics-informed transient dynamics, our work establishes a scalable paradigm for probabilistic modeling of multi-dimensional, parameterized stochastic systems.

physics.comp-ph↗

arXiv2026

Anomalous quantized nonlinear soliton pumping

It has recently been theoretically predicted and experimentally observed that a soliton resulting from nonlinearity can be pumped across an integer or fractional number of unit cells as a system parameter is slowly varied over a pump period. Nonlinear soliton pumping is now understood as the flow of instantaneous Wannier functions, ruling out the possibility of pumping a soliton across a nonzero number of unit cells over one cycle when a corresponding Wannier function does not exhibit any flow, i.e., when the corresponding Bloch band that the soliton bifurcates from is topologically trivial. Here we surprisingly find an anomalous nonlinear soliton pump where the displacement of a soliton over one cycle differs from the Chern number of the Bloch band from which the soliton comes. We show that this anomalous behavior arises from a transition of a soliton between different Wannier functions by passing through an intersite-soliton (or dipole-soliton) state. Furthermore, we find a nonlinearity-induced integer quantized pump of a soliton, allowing a soliton to travel across one unit cell during a pump period, even when the corresponding band is topologically trivial. Our results open the door to studying nonlinearity-induced pumping of solitons.

nlin.PS↗

arXiv2026

Graviton Production from Inflaton Condensate: Boltzmann vs Bogoliubov

We study graviton production from an oscillating inflaton condensate during reheating by systematically comparing Boltzmann and Bogoliubov descriptions for inflaton potentials of the form $V(ϕ)\proptoϕ^n$ around the minimum. The Bogoliubov framework provides a unified description of graviton production, capturing both perturbative and non-perturbative effects across short and long wavelengths, whereas the Boltzmann approach is restricted to perturbative production at short wavelengths. For the quadratic case ($n=2$), we find that the two approaches yield identical graviton spectra at short wavelengths, indicating that the Boltzmann treatments fully captures perturbative gravitational production in this regime. For steeper potentials ($n>2$), however, we identify a sizable contribution arising from the non-adiabatic transition between inflation and reheating. This component is naturally incorporated in the Bogoliubov formalism but absent in the Boltzmann description, and we show that it is important over a broad range of momenta. We derive analytic approximations within both frameworks that clarify the physical origin and scaling behavior of the spectrum. Our results delineate the regime of validity of Boltzmann approaches and show that, for steeper inflaton potentials, graviton production is governed by non-adiabatic transition dynamics for which the Bogoliubov formalism provides the most appropriate description.

hep-ph↗

arXiv2026

Sterile Neutrino Dark Matter as a Probe of Inflationary Reheating

Sterile neutrinos offer a minimal and testable explanation for dark matter (DM), with their radiative decay actively searched for in X-ray observations. We show that cold sterile neutrino DM can be efficiently produced during reheating from inflaton decays with a small branching ratio, ${\rm BR}\lesssim 10^{-4}$. This production mechanism opens regions of parameter space where the active-sterile mixing is small enough to evade current X-ray constraints while reproducing the observed DM abundance. We systematically map the viable parameter space in terms of the sterile neutrino mass, mixing angle, inflaton mass, reheating temperature, and branching ratio. We further demonstrate that sterile neutrino DM can serve as a probe of inflationary reheating, with future X-ray observations capable of yielding information on the inflaton mass and the reheating-temperature ratio $m_ϕ/T_{\rm rh}$. For a given inflationary model in which the inflaton mass is known, this leads to a lower bound on the reheating temperature that is several orders of magnitude stronger than the existing bound from Big Bang Nucleosynthesis.

hep-ph↗

arXiv2026

Saturated and Anisotropic Magnetostriction in an Altermagnet

Magnetostriction, a fundamental phenomenon bridging magnetism and mechanics, has enabled a broad spectrum of applications. For almost two centuries, it has been mainly investigated for ferromagnets. Regarding the magnetostriction of antiferromagnets (AFMs), limitedly known examples for both conventional collinear AFMs and noncollinear AFMs predominantly exhibit non-saturating magnetic-field dependence. Herein, we report an easily saturated magnetostriction effect in a prototypical altermagnet - MnTe, which is an emerging class of collinear AFMs with special crystal symmetries. For high-quality MnTe single crystals, the magnetostriction saturates under a moderate field of ~0.7 T with an intriguing two-fold-symmetry anisotropy. First-principles calculations reveal that the saturated and anisotropic magnetostriction originates from symmetry-allowed coupling between elastic strain and its Néel order parameter. These findings break the traditional wisdom on antiferromagnetic magnetostriction.

cond-mat.mtrl-sci↗

arXiv2026

Memory-Dependent FPK Equations for Nonlinear SDOF Oscillators Under Fractional Gaussian Noise Excitation

This paper investigates the transient probabilistic responses of nonlinear single-degree-of-freedom oscillators subjected to external fractional Gaussian noise (FGN) excitation. Owing to the inherent long-range correlations and memory characteristics of FGN, the resulting response process exhibits non-Markovian properties, rendering the classical Fokker-Planck-Kolmogorov (FPK) equation method inapplicable in its direct form. To overcome this critical challenge, a memory-dependent FPK (memFPK) equation is formulated for two-dimensional nonlinear stochastic systems within the fractional Wick-Itô-Skorohod integral framework. The derived memFPK equation incorporates mixed second-order derivative terms, as well as time-dependent and state-dependent diffusion coefficients, which inherently capture the long-range correlations and memory effects induced by FGN excitation. For the numerical solution of the memFPK equation, a discretized local mean treatment is developed to estimate the memory-dependent diffusion coefficients involving conditional expectations. The proposed approach integrates local statistical averaging and smoothing techniques to enhance the stability of coefficient estimation. Subsequently, the memFPK equation is numerically solved using a finite difference scheme. The accuracy and effectiveness of the proposed framework are validated through linear and nonlinear numerical examples. Comparative results demonstrate the excellent agreement with analytical solutions or Monte Carlo simulations in terms of transient joint probability density functions (PDFs), marginal PDFs, low-probability tail regions, and statistical moments. These findings confirm that the proposed memFPK equation method serves as a robust and effective tool for analyzing the transient non-Markovian probabilistic responses of nonlinear SDOF systems under FGN excitation.

math.PR↗

arXiv2026

An EFT Map of Axion Dark Radiation from Reheating

Light, weakly coupled sectors can retain information about the cosmological background in which they are produced. We study light axions produced during reheating and their contribution to dark radiation, $ΔN_{\rm eff}$. We develop a shift-symmetric EFT in which an inflaton-dependent axion kinetic term systematically organizes the leading production channels. The same kinetic function generates both direct inflaton decay and inflaton annihilation from the oscillating inflaton background. Direct decay is described by an invisible inflaton branching fraction, while annihilation is a genuinely reheating-sensitive source controlled by a coherent combination of Wilson coefficients. We derive the contribution to $ΔN_{\rm eff}$ from both channels and show that they scale oppositely with the reheating temperature: the decay contribution falls as $T_{\rm rh}^{-2}$, whereas the annihilation contribution grows approximately as $T_{\rm rh}^{4/3}$. Their crossing is missed by treatments that keep only one production channel. We translate current and projected $ΔN_{\rm eff}$ sensitivities into constraints on the Wilson coefficients of the kinetic function, obtaining a two-dimensional EFT map of axion dark radiation from reheating. This map can imply both lower and upper bounds on the reheating temperature, showing that light axion relics can turn dark radiation measurements into constraints on reheating.

hep-ph↗

arXiv2026

Irreducible Graviton Floor from Reheating

Inflaton decay inevitably emits gravitons through bremsstrahlung during reheating. We show that the soft part of this emission amplitude, fixed by Weinberg's soft-graviton theorem, becomes an irreducible stochastic gravitational-wave (GW) background after accounting for cosmological evolution. The theorem fixes the infrared branch of the spectrum, $Ω_{\rm GW}\propto f$, independently of the microscopic operator responsible for inflaton decay, while the normalization is controlled by the hard inflaton decay rate and by a phase-space factor. We carry this out for inflaton $n$-body decays, including the phase-space integrals, finding that the maximum of the spectrum scales as $2/n$ relative to the $n=2$ case. The signal can reach $Ω_{\rm GW}h^2\sim \mathcal O(10^{-17})$ at frequencies above the GHz scale. This predicts a stochastic graviton floor from perturbative reheating: a larger signal would require either other processes beyond perturbative bremsstrahlung or inflationary scenarios beyond conventional single-field slow roll.

hep-ph↗

arXiv2026

Exciton-roton mode in moiré fractional Chern insulators

Moiré fractional Chern insulators (FCIs) are a novel class of quantum matter that realizes fractional quantum Hall (FQH) physics in zero magnetic field and provides a platform for exploring unconventional collective excitations. Here we show that hybridization between the magneto-roton and moiré interband excitations gives rise to an exciton-roton mode absent in continuum FQH systems in the long-wavelength limit. Using exact diagonalization and a variational Bethe-Salpeter equation for twisted MoTe$_2$, we demonstrate that this hybridization is controlled by the quantum geometry and yields a mode that combines excitonic optical response with the characteristic FCI roton minimum. The resulting exciton-roton remains low-lying, with excitation energy below the interband transition, and acquires optical activity, leading to a double-peak spectroscopic signature. These results identify optical spectroscopy as a direct probe of collective excitations in moiré FCIs.

cond-mat.str-el↗