Search arXiv⌕ Search

arXiv subjects

Dohyun Kim

Publications and source records attributed to Dohyun Kim.

At least 19 recordsLinked to original sources
arXiv2026

Remap for ALE Hydrodynamics via Direct Interpolation and Constraint Optimization

This work presents an optimization-based remap approach for arbitrary Lagrangian-Eulerian (ALE) hydrodynamics that preserves sharp features and physical bounds while conserving global quantities such as volume, mass, momentum, and total energy. Achieving these properties is the main challenge in remap methods. The presented approach is well-suited for high-order finite element (FE) discretizations and ALE methods that allow for large, purely Lagrangian displacements between successive remeshing and remapping steps. Recent advances in parallel, GPU-capable interpolation routines enable the initial transfer of all fields via direct physical-space interpolation, avoiding computationally costly and diffusive pseudo-transient PDE-based advection. The interpolation produces bounded fields with minimal numerical diffusion. However, it may violate the desired conservation of global properties. Thus, a subsequent optimization step restores conservation while preserving physical bounds. The presented method provides minimally diffusive remapping of finite element fields of arbitrary order. Moreover, it directly remaps quantities defined at integration points, eliminating the need for projection onto finite element spaces. The approach also mitigates typical challenges associated with synchronizing primal variables and conserved quantities in advection-based remap methods. The proposed optimization-based remap method is evaluated using standard two- and three-dimensional remap benchmarks and full ALE hydrodynamics simulations.

math.NA↗
arXiv2026

Hamiltonian-Dependent Compressibility and Physical Preservation in Approximate Quantum Compilation using Tensor Networks

Approximate quantum compilation (AQC) using tensor networks provides a promising route to reducing the circuit resources required for quantum simulation. However, the factors governing the compressibility of many-body dynamics and the extent to which compressed circuits retain physically relevant information on quantum hardware remain poorly understood. We investigate these questions using AQCtensor for an 80-qubit one-dimensional anisotropic XYZ Heisenberg chain initialized in a N'eel state. By varying the exchange anisotropy and external magnetic field, we find that compression performance changes systematically with the anisotropic coupling configuration and field orientation. These variations are strongly associated with a nested-commutator measure related to second-order Trotter error, with corresponding directional trends also reflected in the target-state MPS bond dimension. For an 80-qubit quantum-hardware evaluation, AQCtensor compresses a 10-repetition target evolution to 5 repetitions while retaining a target-state fidelity of 0.995474, reducing the total circuit depth by 45.9% and the two-qubit gate count by 47.6%. The compressed circuit exhibits comparable behavior across magnetization, nearest-neighbor correlation, and Hamiltonian expectation measurements, while process-level diagnostics and local reduced-state tomography provide complementary characterization beyond these observables. Together, these results link AQCtensor compressibility to Hamiltonian commutator structure and target-state complexity and demonstrate substantial circuit reduction while retaining physically relevant features in an 80-qubit quantum-hardware experiment.

quant-ph↗
arXiv2026

Dimension-dependent symmetry breaking in a diffusive nonlocal interaction model

We study the equilibrium structure and long-time behavior of a nonlinear Fokker--Planck equation with diffusion and nonlocal interactions generated by squared simplex volumes. For $1\leq n<d$, the stationary states are isotropic Gaussian measures, and every solution converges to the equilibrium determined by its conserved mean. At $n=d$, the free energy fixes only the covariance determinant, allowing a continuum of anisotropic Gaussian equilibria when $d\geq2$. We establish global well-posedness for arbitrary initial probability measures with finite second moment, identify stationary states with the global minimizers of the free energy, and prove convergence in quadratic Wasserstein distance with explicit exponential rates.

math.AP↗
arXiv2026

DriftTTS: One-stage Single Step Text-to-Speech via Generative Drifting

Text-to-speech models built on score matching and flow matching produce natural speech, but only by calling the decoder many times, which makes synthesis slow. Existing ways to cut the number of calls either distill from a multi-step teacher or add a stage on top of a pretrained model, leaving single-step TTS without a pretrained TTS model unexplored. We question whether pretraining a TTS model is necessary, and propose DriftTTS, which adopts generative drifting, an objective that moves iterative sampling from inference into training. Drifting assumes samples of fixed shape, whereas speech has variable length; we address this issue with segment drifting. On LJSpeech, DriftTTS matches the multi-step quality of Matcha-TTS and RapFlow-TTS in a single decoder call, while being faster. To our knowledge, this is the first application of generative drifting to text-to-speech.

cs.SD↗
arXiv2026

Noise-Space Reward Alignment for Few-Step Generative Models

Few-step generative models produce high-quality samples in one or a few evaluations. However, steering them toward desired properties remains difficult, especially when only black-box reward evaluations are available. While one could apply approaches developed for iterative diffusion and flow-based models, such as direct RL fine-tuning or test-time scaling, we find that neither transfers well to the few-step regime: direct fine-tuning suffers from severe reward over-optimization, and test-time scaling sacrifices the efficiency advantage of few-step generation. We propose Noise Policy Optimization (NPO), which explicitly targets the alignment problem for few-step generative models. The key idea is to learn an amortized flow policy over the initial noise space of the generator. The trained policy transports Gaussian noise to a reward-tilted noise distribution, whose pushforward distribution through the frozen generator is exactly the reward-tilted output distribution. By keeping the generator frozen, we can effectively utilize the knowledge encoded in the pretrained model, yielding a better reward-quality trade-off than direct fine-tuning. The amortized policy moves the search cost into training, so inference adds only a lightweight noise sampling step on top of the pretrained model. To improve sample efficiency during training, we use ELBO-based likelihood estimation that requires only sampled endpoints rather than simulated trajectories. Across image, language, and protein structure generation, NPO consistently achieves a better reward-quality trade-off than direct fine-tuning, and matches test-time search methods with 16 to 549 times lower inference cost.

cs.LG↗
arXiv2026

Measuring the Depth of LLM Unlearning via Activation Patching

Large language model (LLM) unlearning has emerged as a crucial post-hoc mechanism for privacy protection and AI safety, yet auditing whether target knowledge is truly erased remains challenging. Existing output-level metrics fail to detect when this knowledge remains recoverable from internal representations. Recent white-box studies reveal such residual knowledge but often rely on auxiliary training or dataset-specific adaptations, leaving no generalizable metric. We close this gap with the Unlearning Depth Score (UDS), a metric that quantifies the mechanistic depth of unlearning via activation patching. UDS first identifies layers that encode the target knowledge using a retain model baseline, then measures how much of it is erased in the unlearned model on a 0-1 scale. In a meta-evaluation across 20 metrics on 150 unlearned models spanning 8 methods, UDS achieves the highest faithfulness and robustness, confirming our causal approach as the most reliable for unlearning evaluation. Case studies further show that UDS uncovers residual knowledge obscured from observational metrics by representational shifts, with erasure depth varying across prompt types. We provide guidelines for integrating UDS into existing benchmarking frameworks and streamlining the evaluation pipeline. Code and data are available at https://github.com/gnueaj/unlearning-depth-score.

cs.CL↗
arXiv2026

Dimension-dependent order parameter selection in higher-order Kuramoto dynamics on spheres: continuous versus quantized regimes

We study a high-dimensional Kuramoto model with attractive pairwise coupling and a repulsive higher-order effect. Although the pairwise interaction favors synchronization, the higher-order interaction prevents complete synchronization and selects an intermediate level of coherence corresponding to the balanced equilibria. For the unit sphere of dimension at least two, we provide an explicit basin of attraction for the balanced equilibria. We also classify nonzero-mean equilibria and show that all non-balanced equilibria are linearly unstable. This indicates that balanced states are the only natural stable candidates in the higher dimensions. In contrast, on the circle, the same model reduces to a higher-order Kuramoto-type model which exhibits a qualitatively different selection mechanism that depends fundamentally on the dimension. In this case, population imbalance at finite $N$ prevents exact stationary balance and instead generates a common angular drift producing phase-locked states. We construct basins of attraction for two-cluster locked states, in which the phases split into two groups, and identify the corresponding finite-$N$ selected values of the order parameters. This phenomenon is called continuous-versus-quantized order-parameter selection. We further demonstrate, through a root-selection mechanism, why such two-cluster locked states are typically observed in most simulations. Finally, we show that two-cluster locked states with a large population imbalance are linearly unstable, which explains why only certain locked states are dynamically robust.

nlin.AO↗
arXiv2026

PUBG Ally: A Conversational Embodied Agent as an AI Teammate

We introduce PUBG Ally, an embodied agent for PUBG: BATTLEGROUNDS that can reason, act autonomously, and play alongside players as a voice-enabled teammate. Building such a teammate requires combining two difficult capabilities: it must perceive and respond to a constantly changing game world under strict latency constraints while interacting naturally with players, keeping its speech synchronized with its actions. Ally therefore combines agentic tool use with real-time game control. A language-model agent uses a controlled interface to inspect game information, interpret player speech, maintain context, decide what to say, and issue high-level action choices that steer a faster control layer for movement, combat, and recovery. Because the player's and Ally's speech and actions continually shape each other and the course of the match, training requires data from actual gameplay. We therefore collect data across nearly 39k sessions in which real players play alongside Ally, recording gameplay, player speech, agent decisions, tool use, actions, and player feedback, and use these records for iterative training. To evaluate teammate quality, we use player feedback and preference comparisons to identify gaps between offline evaluations and player preferences, and iteratively refine the evaluation criteria. Deploying Ally in live service further requires low-latency on-device execution and safeguards for player-facing communication, which we address through model compression, context compaction, targeted safety training, runtime guardrails, and memory redaction. During the live service, we surveyed players in 141 countries. Among respondents whose play with Ally was confirmed in game records, positive responses exceeded negative responses by 25.1 percentage points when asked whether they would recommend Ally, with players describing Ally not only as a tool but also as a teammate or companion.

cs.AI↗
arXiv2026

Diffusion Drafts, AR Verifies: Accelerating Document OCR with Self-Speculative Decoding

Autoregressive OCR vision-language models accurately convert document images into text and structured markup, but require one sequential decoding step per output token, limiting inference speed. Unlike open-ended text generation, OCR outputs are strongly grounded in the input image, making diffusion-based parallel generation promising. However, when several tokens are predicted in one diffusion step, each is predicted before the others are known. Committing them directly can therefore introduce errors. We therefore introduce GravityOCR, a parameter-shared AR-block-diffusion model jointly trained for parallel drafting and causal AR verification. Verifying drafts before commitment lets the model commit multiple output tokens per round without a separate drafting network. The causal AR path also enables GRPO with sequence- and structure-level OCR rewards, avoiding diffusion-trajectory likelihood estimation while updating the shared drafter parameters. On OmniDocBench v1.6, AR-path GRPO improves the Overall score from 94.92 to 95.16 without reducing diffusion drafting efficiency, while the final model remains close to the original GLM-OCR score of 95.48. In an SGLang serving deployment, GravityOCR commits an average of 9.7 output tokens per forward pass and achieves a $3.94\times$ decode-only speedup on region crops and a $1.32\times$ end-to-end page-processing speedup over AR decoding.

cs.CL↗
arXiv2026

Compositional SVG Generation via VLM-Driven Hierarchical Semantic Parsing

While Vision-Language Models (VLMs) excel at visual reasoning, generating structured, editable Scalable Vector Graphics (SVG) remains a fundamental challenge. Existing pipelines predominantly yield flat, semantically agnostic collections of paths, where editing a single object requires manually identifying its constituent paths. To address this, we propose a VLM-driven agentic framework for semantic compositional SVG generation. Our pipeline recursively parses visual scenes into semantic and geometric hierarchies via top-down decomposition, visual grounding, and prompt-driven amodal occlusion recovery, ensuring each component is geometrically complete. Furthermore, we introduce the Semantic SVG Benchmark with human-annotated semantic groups and novel sub-component metrics (Semantic Recall/Precision, PERE) to explicitly evaluate structural compositionality and functional editability. Experiments show that our natively predicted structures surpass the upper bounds of existing flat-generation methods in both grouping quality and editability, while maintaining state-of-the-art visual fidelity.

cs.CV↗
arXiv2026

Aker: Density-Aware Approximate Caching for Vector Search (Extended Version)

Disk-based approximate nearest neighbor search (ANNS) incurs high I/O overhead due to frequent disk accesses during index traversal. Approximate caching, which reuses the results of past queries to serve future similar queries, offers a promising approach to bypass expensive disk searches. However, existing approaches suffer from two key limitations. First, their approximate hit predicates fail to simultaneously achieve high throughput and high accuracy, as they do not adapt to the varying local neighbor density in high-dimensional spaces. Second, they lack an effective refresh mechanism to maintain cache correctness under vector updates. We present Aker, an approximate cache for disk-based ANNS. Aker addresses these limitations through two core design choices. First, we introduce a per-query similarity threshold, where each cache entry maintains its own threshold that is dynamically adjusted based on observed cache hit patterns. This design enables Aker to adapt to neighborhood densities to preserve both efficiency and accuracy. Second, we propose del-consistency, a consistency model for ANNS caches that applies deletions eagerly and insertions lazily. Under this model, Aker implements a low-overhead refresh mechanism that bounds cache staleness and preserves high search accuracy. We integrate Aker into pgvector and evaluate it on representative workloads. Aker improves recall by up to 64 percentage points over prior solutions and increases QPS by up to 3.2x, while using 0.6x the memory of pgvector's shared buffers.

cs.DB↗
arXiv2026

Differentiable high-fidelity vectorial propagator

The numerical accuracy and computational efficiency of free-space field propagation directly determine what physics can be predicted and what devices can be designed, from nanophotonic metasurfaces to optically addressed quantum processors. Existing methods face a fundamental accuracy-speed trade-off, which is further compounded in the vectorial regime, where major formalisms remain fragmented with no unified computational framework. Here we address both challenges: a theory that unifies the vectorial diffraction framework, and an algorithm that resolves the accuracy-speed trade-off, each reinforcing the other. We prove that all major vectorial diffraction formalisms and vector-potential seeding methods arise from a single surface equivalence principle and reduce, without approximation, to a finite sequence of scalar angular-spectrum operations. Any advance in scalar propagation can therefore be exploited across the entire vectorial ecosystem. The error-compensating angular spectrum method (E-ASM) serves as a fast and accurate scalar solver, delivering peak signal-to-noise-ratio gains exceeding 60 dB over the state-of-the-art band-extended angular spectrum method at order-of-magnitude faster speed, with a freely configurable observation window. Their combination, implemented in a GPU-accelerated, fully differentiable form, turns rigorous vectorial field computation, previously prohibitive for higher-order beams over macroscopic distances, into a practical, inverse-design-ready tool. We demonstrate the framework on optically addressed local qubit gates and on vectorial Hermite-Gaussian and Laguerre-Gaussian beam propagation, resolving field structures inaccessible to scalar or approximate vectorial methods.

physics.optics↗
arXiv2026

Direct Diffusion Score Preference Optimization via Stepwise Contrastive Policy-Pair Supervision

Diffusion models have achieved impressive results in generative tasks such as text-to-image synthesis, yet they often struggle to fully align outputs with nuanced user intent and maintain consistent aesthetic quality. Existing preference-based training methods such as Diffusion Direct Preference Optimization help address these issues, but obtain their supervision targets from the forward process $q(\boldsymbol{x}_{t-1}\!\mid\!\boldsymbol{x}_t,\boldsymbol{x}_0)$ derived from terminal samples, which is not directly aligned with the model's actual backward denoising transitions at each step. In this work, we introduce Direct Diffusion Score Preference Optimization (DDSPO), which defines stepwise preference supervision directly over backward denoising transitions through a contrastive policy pair, rather than relying on forward-process approximations from terminal samples. We propose two practical instantiations of the contrastive policy pair: training separate winning and losing models on preference data, and inducing a contrastive policy pair without additional training by using a pretrained reference model conditioned on an original prompt and a semantically degraded variant, requiring neither reward modeling nor manual annotations. Empirical results show that contrastive-policy-pair supervision is more effective than forward-process-based supervision across text-image alignment and aesthetic-quality tasks. Our implementation is available at: https://dohyun-as.github.io/DDSPO

cs.CV↗
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

Finite-dimensional reduction of a Wasserstein gradient flow and sharp decay rates

We study the Wasserstein gradient flow generated by a family of extended generalized variance functionals, defined as the expected squared $n$-dimensional volume of a simplex, which includes the classical variance-type interaction and generalized variance as special cases. The key structural observation is that this functional depends only on the covariance matrix. Consequently, the Wasserstein gradient flow reduces to a finite-dimensional system for the eigenvalues of the covariance matrix, and the full measure-valued solution can be recovered through an explicit linear pushforward representation. Using this representation, we establish global well-posedness for arbitrary initial data in $\mathcal P_2(\mathbb R^d)$ without assuming compact support. We also study the long-time behavior of the flow. For every initial datum in $\mathcal P_2(\mathbb R^d)$, the solution converges to a limiting equilibrium measure whose covariance has rank strictly less than $n$. Moreover, we obtain sharp convergence rates in all spectral regimes: exponential in the non-degenerate case and algebraic with the optimal exponent in the degenerate case.

math.AP↗
arXiv2026

Collective Optimization on Riemannian Manifolds with Bounded Curvature

In this paper, we develop an intrinsic consensus-based optimization framework on Riemannian manifolds with bounded sectional curvature. In contrast to extrinsic approaches based on an ambient Euclidean embedding, our model is formulated directly in terms of the Riemannian structure, using logarithmic and exponential maps induced by the intrinsic geodesic distance. We prove the global well-posedness of the proposed particle system and its associated McKean--Vlasov dynamics. We also establish the global convergence of the mean-field equation toward a global minimizer of the objective function under suitable conditions. Numerical experiments on the sphere, hyperbolic space, and the special orthogonal group demonstrate the effectiveness of the intrinsic CBO dynamics for nonconvex optimization problems on manifolds.

math.OC↗
arXiv2026

Aligning Few-Step Generative Models by Amortizing Sample-based Variational Inference

Aligning a few-step generative model is challenging, since existing alignment frameworks typically rely on restrictive assumptions: a tractable likelihood, a specific ODE/SDE solver, or a particular model family. We introduce FAV, Few-step Generative Models Alignment via Sample-based Variational Inference, a general alignment framework that requires only sample access to the generator and the reference distribution. We cast alignment as sampling from a reward-tilted distribution anchored to a reference distribution. We leverage Stein Variational Gradient Descent as a sample-based variational inference scheme and amortize its particle updates into the generator parameters via fixed-point regression. We evaluate FAV on two domains: robotics manipulation and image generator alignment. On generative policy alignment for robotic manipulation, FAV outperforms prevailing policy extraction baselines across 56 offline and 30 offline-to-online RL tasks. For image generator alignment, FAV fine-tunes diverse few-step backbones, including GAN, drifting model, consistency models, and flow maps, scaling from ImageNet-$256$ to 1024$^2$ text-to-image synthesis. Code is available at https://github.com/Jaewoopudding/FAV.

cs.LG↗
arXiv2026

The SiMPL Method for Multi-Material Topology Optimization

We introduce an efficient and scalable method for density-based multi-material topology optimization, integrating classical mirror descent techniques with point-wise polytopal design constraints. Such constraints arise naturally in this class of problems, wherein the vertices of convex polytopes correspond to distinct design states, only one of which should be occupied at each point in space. The framework generates a descending sequence of iterates by penalizing the design space around the previous iterate with a generalized distance function tailored to the convex geometry of the $n$-dimensional polytope. This distance function, called a Bregman divergence, smooths the optimization landscape, ensuring that each iterate strictly satisfies the point-wise constraints. Subsequently, global constraints (e.g., bounds on the structural mass) can be enforced easily by solving a small, finite-dimensional dual problem. The resulting method is simple to implement and demonstrates robustness and efficiency when combined with an Armijo-type line search algorithm. We validate the method in structural design problems involving the optimal arrangement of both isotropic and anisotropic materials, as well as magnetic flux optimization in electric motors.

math.NA↗