arXiv subjects

Jinwook Jung

Publications and source records attributed to Jinwook Jung.

At least 19 recordsLinked to original sources

arXiv2026

Regularity theory and low Mach number limit for the fractional Euler-alignment system

We study the compressible Euler-alignment system with pressure under a singular pressure scaling, where the hypersingular communication weight induces a fractional alignment operator of order $2α$, $0<α<1$. The scaling corresponds to a large-time and small-velocity regime and leads to a low Mach number problem in which the density is forced to remain close to a constant state. Our main result is a uniform regularity theory for this scaled pressure system. We establish uniform estimates with respect to the scaling parameter and construct global strong solutions near the constant state. A key feature of the analysis is that, in the low-order fractional regime $0<α\le\frac12$, the estimates close under the lower Sobolev condition $s>\frac d2+1-2α$, gaining $α$ derivatives over the threshold $s>\frac d2 + 1 - α$ arising from a direct use of the fractional alignment dissipation. This is achieved by combining refined commutator estimates for the singular alignment operator with the density dissipation induced by the pressure scaling. As an application of the uniform estimates, we justify the low Mach number limit toward the incompressible Navier--Stokes system with fractional dissipation. For general small, possibly ill-prepared initial data, a Helmholtz decomposition combined with dispersive estimates for the acoustic component yields subsequential strong convergence locally in space-time to a distributional solution of the limiting system. For well-prepared initial data, a relative-energy argument further identifies the limit with a prescribed sufficiently regular solution and yields global strong convergence in the fractional dissipation norm.

math.AP↗

arXiv2026

Learning to Recover Task Experts from a Multi-Task Merged Model

Multi-task model merging aims to consolidate several task-specific experts into a unified model, yet static merging consistently suffers from parameter interference. While dynamic merging models aim to bridge this gap, many works rely on the costly storage and loading of redundant expert components at inference. In this work, from the perspective of task expert, we view parameter interference as parameter perturbation introduced to each expert during merging process. We show that such parameter perturbations can be modeled as affine transformation, which can be approximated as additive offsets. Motivated by these, we propose Recover Task eXpert (ReTeX), a framework that predicts those offsets, in order to undo parameter interference and recover task-expert performance from a single merged checkpoint. To recover the appropriate expert when task identity is unknown, we introduce a router-free task identifier based on SVD subspace signatures computed offline before inference. At inference, the identifier selects the task whose subspace yields the smallest projection residual for a given input. As a result, ReTeX recovers over 95% of individual-expert performance in both vision and NLP domains, while significantly improving generalization to unseen tasks. Crucially, we also show that the parameter offset prediction leads to emergent adaptive interpolation of expert knowledge for out-of-distribution (OOD) tasks. ReTeX adaptively interpolates seen expert knowledge to handle unseen tasks. Our code is available at https://github.com/BAIKLAB/ReTeX

cs.AI↗

arXiv2026

Training-free Task Classification for Multi-Task Model Merging

Ever since the advent of foundation models and the pre-training-finetuning paradigm, there have been numerous efforts to merge multiple task-specific experts into a single multi-task model. Prior work largely focuses on finding a single merged model, but it often underperforms individual experts due to parameter interference. To resolve this, dynamic model merging employs routing to activate task-relevant parameters per input. However, existing routers typically require either additional training with abundant labeled datasets or assume the access to task IDs of each input at inference time. In this work, we aim to close the gap to expert performance without additional training or task-ID-access assumption. To this end, we formulate routing as training-free task classification for each test input. Using singular value decomposition (SVD)-based low-rank manifold approximations for each task, SiM scores tasks by the projection residual of the test input feature onto each task manifold and routes accordingly. The task manifolds are pre-computable offline from a pretrained backbone using a small per-task support set (e.g., 32 examples per task) prior to merging process, requiring no router training and no data during the merging process. Moreover, SiM integrates seamlessly with subspace-/mask-based merging that represents task-expert via lightweight compressed task vectors, avoiding the need to store full expert parameters. Experiments across computer vision and natural language processing benchmarks under task-unknown inference demonstrate that SiM substantially improves merged-model performance and consistently narrows the gap to individual task experts. Our code is available at https://github.com/BAIKLAB/SiM

cs.LG↗

arXiv2026

Optimal control of diffusive mean-field models for swarming particles on the sphere

We study a mean-field optimal control problem for a consensus (high-dimensional Kuramoto-type) dynamics with diffusion on the unit sphere. The control acts through a prescribed drift field and an interaction gain, and the cost functional is given to track a given target density while penalizing the control effort. At the microscopic level, we formulate the corresponding controlled $N$-particle Liouville problem and establish the existence of optimal controls. For fixed controls, we obtain a quantitative stochastic mean-field limit showing that the one-particle marginal converges to the mean-field solution with the convergence rate $\mathcal O(1/\sqrt{N})$. Finally, we show that microscopic optimal controls approximate a mean-field optimal control: any weak limit of particle-level minimizers is optimal for the mean-field problem.

math.OC↗

arXiv2026

Coupled Vlasov and non-Newtonian fluid dynamics: existence and large-time behavior

We study a coupled kinetic-non-Newtonian fluid system on the periodic domain ${\mathbb T}^3$, where particles evolve by a Vlasov equation and interact with an incompressible power-law fluid through a drag force. We prove the global existence of weak solutions for all $p > \frac{8}{5}$, where $p > 1$ denotes the power-law exponent of the fluid's stress-strain relation. Under an additional uniform boundedness assumption on the particle density, we also establish large-time decay of a modulated energy functional measuring deviation from velocity alignment. The decay rate is algebraic when $p > 2$ and exponential when $\frac{6}{5} \le p \le 2$, reflecting the role of fluid dissipation in the large-time dynamics.

math.AP↗

arXiv2026

Quantitative light-particle limit for the Vlasov-Fokker-Planck-Navier-Stokes system

We investigate the hydrodynamic limit of the Vlasov-Fokker-Planck-Navier-Stokes system in the light particle regime, where the particle relaxation takes place on a singularly fast time scale. Using a relative entropy method adapted to this scaling, we develop the first quantitative convergence theory for the light particle limit. Our analysis yields explicit rates for the convergence of both the kinetic distribution and the fluid velocity, extending the qualitative compactness-based result of Goudon, Jabin, and Vasseur [Indiana Univ. Math. J., 53, (2004), 1495-1515]. Moreover, we derive refined convergence estimates for the macroscopic density and fluid velocity in negative Sobolev spaces, consistent with the formally optimal rates predicted by the Hilbert expansion. The results apply to both the torus and the whole space, providing a unified quantitative description of the light particle hydrodynamic limit.

math.AP↗

arXiv2026

Global large smooth solutions and overdamped limits for the damped isothermal Euler-Poisson system

We consider the isothermal Euler-Poisson system with linear damping on a periodic domain in the large damping regime. For arbitrarily large smooth initial data with density bounded away from vacuum, we prove the global-in-time existence of smooth solutions. The argument is based on a large-damping bootstrap scheme, a modified density estimate revealing hidden parabolic dissipation, top-order weighted cancellations, and a comparison with an auxiliary drift-diffusion--Poisson system. We further prove exponential relaxation to the homogeneous equilibrium and establish a large-data overdamped limit as the damping coefficient tends to infinity. In the slow time scale, the density converges quantitatively to the large smooth solution of the drift-diffusion--Poisson system with the same initial density. After subtracting a fast initial layer from the rescaled flux, the flux converges quantitatively to the corresponding drift-diffusion flux.

math.AP↗

arXiv2026

Eventual regularity and asymptotic behavior of Leray-Hopf weak solutions for the Hall-MHD system

In this paper, we study the incompressible, viscous and resistive Hall-magnetohydrodynamic (Hall-MHD) system. We first prove that every two-dimensional Leray-Hopf weak solution becomes smooth after a finite time. In three dimensions, where eventual smoothness for arbitrary Leray-Hopf weak solutions is not known, we construct Leray-Hopf weak solutions for which the magneto-vorticity field $B+\nabla\times u$ eventually gains additional regularity. Finally, under suitable low-frequency pseudomeasure assumptions on initial data, we establish decoupled algebraic decay rates for the velocity and magnetic fields by combining a generalized Fourier splitting method with the eventual smoothness in two dimensions and strong regularity in three dimensions.

math.AP↗

arXiv2026

Bilinear Coordinate Alignment for Training-Free Task-Vector Transfer

Fine-tuning large-scale pre-trained models is a recent prevalent paradigm for adapting general representations to specialized tasks. However, when a new version of a pre-trained model becomes available, expertise acquired through fine-tuning cannot be directly reused because it is tied to the parameterization of the original model, requiring another costly fine-tuning. To address this inefficiency, recent work uses task vectors, defined as the parameter difference between a fine-tuned model and its base model, to transfer expertise across models. While existing methods bridge disparate models by matching activations or gradients, a significant performance gap remains relative to direct fine-tuning, suggesting that these partial correspondences are insufficient. In this work, instead of viewing a task vector merely as a parameter offset, we revisit the formation of task vectors and show that they can be derived as accumulated bilinear interactions between input-side activations and output-side gradients. Motivated by this observation, we formulate task-vector transfer as a dual-space alignment problem and propose BiCo, a training-free framework for transferring task vectors through Bilinear Coordinate alignment. BiCo estimates orthogonal Procrustes mappings in both spaces using a single forward-backward pass on a small calibration set, without any parameter update. Across extensive computer vision and natural language processing benchmarks, BiCo consistently outperforms existing transfer methods across models that differ in width, depth, and pre-training configuration.

cs.LG↗

arXiv2026

A unified relative entropy framework for macroscopic limits of Vlasov--Fokker--Planck equations

We develop a unified relative entropy framework for macroscopic limits of kinetic equations with Riesz-type interactions and Fokker-Planck relaxation. Our analysis covers three prototypical singular regimes: the diffusive limit leading to a drift-diffusion equation, the high-field limit yielding the aggregation equation in the repulsive regime, and the strong magnetic field limit producing a generalized surface quasi-geostrophic equation. The method combines entropy dissipation, Fisher-information control, and modulated interaction energies into a robust stability theory yielding both strong and weak convergence results. For the strong convergence, we establish quantitative relative entropy estimates toward macroscopic limits under well-prepared initial data, extending the scope of the method to settings where nonlocal forces and singular scalings play a decisive role. For the weak convergence, our approach captures three complementary phenomena: in the diffusive regime, it yields sharper quantitative estimates in weak topologies consistent with the formally optimal scaling; in the high-field regime, it propagates bounded Lipschitz stability for a class of mildly prepared initial data, even when the relative entropy diverges with respect to the singular scaling parameter; and in the strong magnetic field regime, it provides quantitative weak estimates, including bounded Lipschitz control of the rescaled momentum and negative Sobolev control of the density. This broader perspective shows that relative entropy provides not only a tool for strong convergence, but also a mechanism for treating low-regularity and mildly prepared regimes. The analysis highlights the unifying role of relative entropy in connecting microscopic dissipation with both strong and weak macroscopic convergence.

math.AP↗

arXiv2025

On the temporal estimates for the incompressible Navier-Stokes equations and the Hall-magnetohydrodynamic equations

In this paper, we derive decay rates for solutions to the incompressible Navier-Stokes equations and Hall-magnetohydrodynamic equations. We first improve the decay rate of weak solutions to these equations by refining the Fourier splitting method with initial data in the space of pseudo-measures. Additionally, we investigate these equations with initial data in the Lei-Lin spaces and establish decay rates for those solutions.

math.AP↗

arXiv2025

AIS-LLM: A Unified Framework for Maritime Trajectory Prediction, Anomaly Detection, and Collision Risk Assessment with Explainable Forecasting

With the increase in maritime traffic and the mandatory implementation of the Automatic Identification System (AIS), the importance and diversity of maritime traffic analysis tasks based on AIS data, such as vessel trajectory prediction, anomaly detection, and collision risk assessment, is rapidly growing. However, existing approaches tend to address these tasks individually, making it difficult to holistically consider complex maritime situations. To address this limitation, we propose a novel framework, AIS-LLM, which integrates time-series AIS data with a large language model (LLM). AIS-LLM consists of a Time-Series Encoder for processing AIS sequences, an LLM-based Prompt Encoder, a Cross-Modality Alignment Module for semantic alignment between time-series data and textual prompts, and an LLM-based Multi-Task Decoder. This architecture enables the simultaneous execution of three key tasks: trajectory prediction, anomaly detection, and risk assessment of vessel collisions within a single end-to-end system. Experimental results demonstrate that AIS-LLM outperforms existing methods across individual tasks, validating its effectiveness. Furthermore, by integratively analyzing task outputs to generate situation summaries and briefings, AIS-LLM presents the potential for more intelligent and efficient maritime traffic management.

cs.LG↗

arXiv2025

On the mean-field limit of Vlasov-Poisson-Fokker-Planck equations

The derivation of effective descriptions for interacting many-body systems is an important branch of applied mathematics. We prove a propagation of chaos result for a system of $N$ particles subject to Newtonian time evolution with or without additional white noise influencing the velocities of the particles. We assume that the particles interact according to a regularized Coulomb-interaction with a regularization parameter that vanishes in the $N\to\infty$ limit. The respective effective description is the so called Vlasov-Poisson-Fokker-Planck (VPFP), respectively the Vlasov-Poisson (VP) equation in the case of no or sub-dominant white noise. To obtain our result we combine the relative entropy method from \cite{jabinWang2016} with the control on the difference between the trajectories of the true and the effective description provided in \cite{HLP20} for the VPFP case respectively in \cite{LP} for the VP case. This allows us to prove strong convergence of the marginals, i.e. convergence in $L^1$.

math-ph↗

arXiv2025

Mitigating Parameter Interference in Model Merging via Sharpness-Aware Fine-Tuning

Large-scale deep learning models with a pretraining-finetuning paradigm have led to a surge of numerous task-specific models fine-tuned from a common pre-trained model. Recently, several research efforts have been made on merging these large models into a single multi-task model, particularly with simple arithmetic on parameters. Such merging methodology faces a central challenge: interference between model parameters fine-tuned on different tasks. Few recent works have focused on designing a new fine-tuning scheme that can lead to small parameter interference, however at the cost of the performance of each task-specific fine-tuned model and thereby limiting that of a merged model. To improve the performance of a merged model, we note that a fine-tuning scheme should aim for (1) smaller parameter interference and (2) better performance of each fine-tuned model on the corresponding task. In this work, we aim to design a new fine-tuning objective function to work towards these two goals. In the course of this process, we find such objective function to be strikingly similar to sharpness-aware minimization (SAM) objective function, which aims to achieve generalization by finding flat minima. Drawing upon our observation, we propose to fine-tune pre-trained models via sharpness-aware minimization. The experimental and theoretical results showcase the effectiveness and orthogonality of our proposed approach, improving performance upon various merging and fine-tuning methods. Our code is available at https://github.com/baiklab/SAFT-Merge.

cs.LG↗

arXiv2024

Global smooth solutions to the irrotational Euler-Riesz system in three dimensions

This paper investigates the global dynamics of the Euler--Riesz system in three dimensions, focusing on the well-posedness and large-time behavior of solutions near equilibrium. The system generalizes classical interactions by incorporating the Riesz interactions $\nabla (-Δ)^{-σ/2}(ρ- 1)$. We show that the system admits a global smooth solution for small irrotational initial perturbations. Specifically, we establish that if the initial data is sufficiently small, the solution remains regular globally in time and decays over time at a rate dependent on $σ$.

math.AP↗

arXiv2024

PBVS 2024 Solution: Self-Supervised Learning and Sampling Strategies for SAR Classification in Extreme Long-Tail Distribution

The Multimodal Learning Workshop (PBVS 2024) aims to improve the performance of automatic target recognition (ATR) systems by leveraging both Synthetic Aperture Radar (SAR) data, which is difficult to interpret but remains unaffected by weather conditions and visible light, and Electro-Optical (EO) data for simultaneous learning. The subtask, known as the Multi-modal Aerial View Imagery Challenge - Classification, focuses on predicting the class label of a low-resolution aerial image based on a set of SAR-EO image pairs and their respective class labels. The provided dataset consists of SAR-EO pairs, characterized by a severe long-tail distribution with over a 1000-fold difference between the largest and smallest classes, making typical long-tail methods difficult to apply. Additionally, the domain disparity between the SAR and EO datasets complicates the effectiveness of standard multimodal methods. To address these significant challenges, we propose a two-stage learning approach that utilizes self-supervised techniques, combined with multimodal learning and inference through SAR-to-EO translation for effective EO utilization. In the final testing phase of the PBVS 2024 Multi-modal Aerial View Image Challenge - Classification (SAR Classification) task, our model achieved an accuracy of 21.45%, an AUC of 0.56, and a total score of 0.30, placing us 9th in the competition.

cs.CV↗

arXiv2024

On well/ill-posedness for the generalized surface quasi-geostrophic equations in Hölder spaces

We establish the well/ill-posedness theories for the inviscid $α$-surface quasi-geostrophic ($α$-SQG) equations in Hölder spaces, where $α= 0$ and $α= 1$ correspond to the two-dimensional Euler equation in the vorticity formulation and SQG equation of geophysical significance, respectively. We first prove the local-in-time well-posedness of $α$-SQG equations in $C([0,T);C^{0,β}(\mathbb{R}^2))$ with $β\in (α,1)$ for some $T>0$. We then analyze the strong ill-posedness in $C^{0,α}(\mathbb{R}^2)$ constructing smooth solutions to the $α$-SQG equations that exhibit $C^{0,α}$--norm growth in a short time. In particular, we develop the nonexistence theory for $α$-SQG equations in $C^{0,α}(\mathbb{R}^2)$.

math.AP↗

arXiv2024

The global Cauchy problem for the Euler-Riesz equations

We completely resolve the global Cauchy problem for the multi-dimensional Euler-Riesz equations, where the interaction forcing is given by $\nabla (-Δ)^{-σ/2}ρ$ for some $σ\in (0,2)$. We construct the global-in-time unique solution to the Euler-Riesz system in a $H^s$ Sobolev space under a smallness assumption on the initial density and a dispersive spectral condition on the initial velocity. Moreover, we investigate the algebraic time decay of convergences for the constructed solutions. Our results cover the both attractive and repulsive cases as well as the whole regime $σ\in (0,2)$.

math.AP↗