Searcharxiv⌕ Search

arXiv subjects

Jonghyun Lee

Publications and source records attributed to Jonghyun Lee.

At least 19 recordsLinked to original sources
arXiv2026

SEAL: Semantic-aware Single-image Sticker Personalization with a Large-scale Sticker-tag Dataset

Synthesizing a target concept from a single reference image is challenging in diffusion-based personalized text-to-image generation, particularly for sticker personalization where prompts often require explicit attribute edits. With only one reference, test-time fine-tuning (TTF) methods tend to overfit, producing \textit{visual entanglement}, where background artifacts are absorbed into the learned concept, and \textit{structural rigidity}, where the model memorizes reference-specific spatial configurations and loses contextual controllability. To address these issues, we introduce \textbf{SE}mantic-aware single-image sticker person\textbf{AL}ization (\textbf{SEAL}), a plug-and-play, architecture-agnostic adaptation module that integrates into existing personalization pipelines without modifying their U-Net-based diffusion backbones. SEAL applies three components during embedding adaptation: (1) a Semantic-guided Spatial Attention Loss, (2) a Split-merge Token Strategy, and (3) Structure-aware Layer Restriction. To support sticker-domain personalization with attribute-level control, we present StickerBench, a large-scale sticker image dataset with structured tags under a six-attribute schema (Appearance, Emotion, Action, Camera Composition, Style, Background). These annotations provide a consistent interface for varying context while keeping target identity fixed, enabling systematic evaluation of identity disentanglement and contextual controllability. Experiments show that SEAL consistently improves identity preservation while maintaining contextual controllability, highlighting the importance of explicit spatial and structural constraints during test-time adaptation. The code, StickerBench, and project page will be publicly released.

cs.CV↗
arXiv2026

KVoiceBench, KOpenAudioBench, and KMMAU: Agent-Driven Korean Speech Benchmarks for Evaluating SpeechLMs

Speech language models (SpeechLMs) have achieved substantial progress by extending large language models (LLMs) to the speech modality. However, SpeechLM evaluation remains heavily centered on English, limiting reliable assessment of multilingual speech capabilities. Straightforward benchmark transfer through ASR, translation, normalization, and TTS can corrupt language-specific instructions, answer constraints, and spoken forms; for audio understanding, transferring source-language audio also fails to preserve target-language speaker attributes, accents, and paralinguistic properties. To address these limitations, we propose two human-agent benchmark-construction frameworks: one transfers source-language SpokenQA benchmarks into target-language SpokenQA benchmarks, and the other converts target-language ASR corpora into audio understanding benchmarks using transcriptions and speaker metadata. Using these frameworks, we construct and publicly release three Korean speech benchmarks: KVoiceBench and KOpenAudioBench for Korean SpokenQA, and KMMAU for Korean audio understanding, comprising 12,345 samples in total. We evaluate eight recent SpeechLMs and find that English-Korean performance gaps vary substantially across models and task families, and that SpokenQA and audio understanding rankings diverge, revealing complementary weaknesses invisible to English-only evaluation.

cs.CL↗
arXiv2026

OTRO: Oblivious Tokenization Path with Square-Root ORAM

The CPU-side large language model (LLM) tokenizer is a critical security gap in LLM serving through a confidential computing stack with CPU and GPU trusted execution environments (TEEs). Tokenizers converts the prompts through table-driven lookups, and the resulting memory access patterns are a powerful source of side-channel leakage. Recent work demonstrates end-to-end recovery of user prompts from tokenizer access pattern on production Intel TDX. However, a drop-in use of the popular tree-based Oblivious RAMs (e.g., PathORAM) to prevent access-pattern leakage introduces $\sim$13$\times$ tokenizer slowdown, resulting in 10-58% higher time-to-first-token (TTFT). In this paper, we present OTRO, an efficient, oblivious tokenization path tailored to latency-critical LLM serving. OTRO relies on square-root ORAM for fast single-access lookups, but avoids its prohibitive $O(N\log^2N$) rebuild cost every $\sqrt{N}$ accesses through three key innovations. First, OTRO provides a pool of replicated square-root ORAM instances that utilize the read-only nature of tokenizer table. Second, an epoch-based rotation policy decouples accesses from rebuilds and pads each epoch with dummy accesses to its boundaries, minimizing observable information. Lastly, chunked KV-cache-aware tokenization further overlaps rebuilds with GPU prefill and minimizes the instance count. Implemented as modules in HuggingFace Tokenizers and nano-vLLM, running within a TDX-enabled CVM with an NVIDIA H100 GPU, OTRO limits TTFT overhead to at most 4.5%, keeps tokenizer-induced latency under 10\% of total TTFT, and adds less than 0.5 GB of memory overhead while reducing the tokenizer's observable leakage across various model families and sizes.

cs.CR↗
arXiv2026

Looped Diffusion Language Models

Masked diffusion models (MDMs) have emerged as a promising alternative to autoregressive models for language modeling, yet the effective design of transformer architectures for MDMs remains underexplored. In this paper, we show that selectively looping the early-middle transformer layers significantly improves both training efficiency and model performance in MDMs. We call this approach LoopMDM(Looped Masked Diffusion Model), which brings two key benefits: looping layers at training-time yields a depth-scaling effect without adding parameters, while varying the number of loops at inference-time enables flexible compute scaling. Despite the simplicity, the results are striking: across multiple pre-training corpora, LoopMDM matches the performance of same-size MDMs with up to 3.3 fewer training FLOPs, while its final performance outperforms them on various reasoning benchmarks, including up to 8.5 points on GSM8K. It even surpasses deeper non-looped MDMs trained with comparable per-step compute, indicating that selective looping is more effective than naive depth scaling. Furthermore, LoopMDM can scale inference-time compute by increasing the number of loops. Adaptively adjusting the number of loops throughout the sampling process further yields additional gains in compute efficiency while maintaining performance. Lastly, with attention analysis, we provide evidence that looping is effective in MDMs by promoting interactions among masked positions. Our code and weights will be publicly released.

cs.LG↗
arXiv2026

MMTB: Evaluating Terminal Agents on Multimedia-File Tasks

Terminals provide a powerful interface for AI agents by exposing diverse tools for automating complex workflows, yet existing terminal-agent benchmarks largely focus on tasks grounded in text, code, and structured files. However, many real-world workflows require practitioners to work directly with audio and video files. Working with such multimedia files calls for terminal agents not only to understand multimedia content, but also to convert auditory and visual evidence across related files into appropriate actions. To evaluate terminal agents on multimedia-file tasks, we introduce MultiMedia-TerminalBench (MMTB), a benchmark of 105 tasks across 5 meta-categories where terminal agents directly operate with audio and video files. Alongside MMTB, we propose Terminus-MM, a multimedia harness that extends Terminus-KIRA with audio and video perception for terminal agents. Together, MMTB and Terminus-MM support a controlled study of multimedia terminal agents, revealing how different forms of multimedia access shape task outcomes and determine which evidence agents rely on to construct executable terminal workflows. MMTB media and metadata are released at https://huggingface.co/datasets/mm-tbench/mmtb-media

cs.MM↗
arXiv2026

Orak: A Foundational Benchmark for Training and Evaluating LLM Agents on Diverse Video Games

Large Language Model (LLM) agents are reshaping the game industry, by enabling more intelligent and human-preferable characters. Yet, current game benchmarks fall short of practical needs: they lack evaluations of diverse LLM capabilities across various game genres, studies of agentic modules crucial for complex gameplay, and fine-tuning datasets to adapt pre-trained LLMs into gaming agents. To fill these gaps, we present Orak, a benchmark for training and evaluating LLM agents across 12 popular video games spanning all major genres. Using a plug-and-play interface built on Model Context Protocol (MCP), Orak supports systematic and reproducible studies of agentic modules in varied game scenarios. We further release a fine-tuning dataset of expert LLM gameplay trajectories covering multiple genres, turning general LLMs into effective game agents. Orak offers a united evaluation framework, including game leaderboards, LLM battle arenas, and \fix{ablation studies} of input modality, agentic strategies, and fine-tuning effects, establishing a foundation towards versatile gaming agents. Code and datasets are available at https://github.com/krafton-ai/Orak and https://huggingface.co/datasets/KRAFTON/Orak.

cs.AI↗
arXiv2026

Raon-Speech Technical Report

We present Raon-Speech, a top-performing 9B-parameter speech language model (SpeechLM) for English and Korean speech understanding, answering, and generation, and Raon-SpeechChat, a high-performing full-duplex extension for natural real-time conversation. Raon-Speech successfully transforms a pre-trained LLM into a SpeechLM that both understands and generates speech while preserving strong text capabilities. It trains on 1.38M hours of highly curated English and Korean speech and text datasets with the following training stages: (1) speech modules alignment, (2) end-to-end SpeechLM pre-training with knowledge distillation, and (3) multi-task preference optimization-based post-training. Across 42 English and Korean speech and text benchmarks, Raon-Speech establishes the strongest overall profile on speech-centric tasks in our comparison against eight similarly sized recent audio foundation models, including Qwen2.5-Omni and Fun-Audio-Chat, while preserving strong text question answering performance. Building upon it, Raon-SpeechChat enables natural full-duplex conversation by continual training on 119K hours of time-aligned real and synthetic dialogue data. It proceeds through three complementary training stages: (1) causal encoder adaptation, (2) full-duplex pre-training, (3) full-duplex fine-tuning for voice and role-control. On multiple full-duplex benchmarks, Raon-SpeechChat shows its clearest strengths on the turn-taking and interruption-sensitive behaviors covered by FDB v1.0, and remains competitive across the broader full-duplex evaluation suite. We open-source all model checkpoints, the training and inference pipeline, and an interactive demo.

cs.CL↗
arXiv2026

See and Fix the Flaws: Enabling VLMs and Diffusion Models to Comprehend Visual Artifacts via Agentic Data Synthesis

Despite recent advances in diffusion models, AI generated images still often contain visual artifacts that compromise realism. Although more thorough pre-training and bigger models might reduce artifacts, there is no assurance that they can be completely eliminated, which makes artifact mitigation a highly crucial area of study. Previous artifact-aware methodologies depend on human-labeled artifact datasets, which are costly and difficult to scale, underscoring the need for an automated approach to reliably acquire artifact-annotated datasets. In this paper, we propose ArtiAgent, which efficiently creates pairs of real and artifact-injected images. It comprises three agents: a perception agent that recognizes and grounds entities and subentities from real images, a synthesis agent that introduces artifacts via artifact injection tools through novel patch-wise embedding manipulation within a diffusion transformer, and a curation agent that filters the synthesized artifacts and generates both local and global explanations for each instance. Using ArtiAgent, we synthesize 100K images with rich artifact annotations and demonstrate both efficacy and versatility across diverse applications. Code is available at link.

cs.CV↗
arXiv2026

STAG: Structural Test-time Alignment of Gradients for Online Adaptation

Test-Time Adaptation (TTA) adapts pre-trained models using only unlabeled test streams, requiring real-time inference and update without access to source data. We propose StructuralTest-time Alignment of Gradients (STAG), a lightweight plug-in enhancer that exploits an always-available structural signal: the classifier's intrinsic geometry. STAG derives class-wise structural anchors from classifier weights via self-structural entropy, and during adaptation analytically computes the predicted-class entropy gradient from forward-pass quantities, aligning it to the corresponding anchor with a cosine-similarity loss. This closed-form design incurs near-zero memory and latency overhead and requires no additional backpropagation beyond the underlying baseline. Across corrupted image classification and continual semantic segmentation, STAG provides broadly applicable performance gains for strong TTA baselines on both CNN and Transformer architectures regardless of the underlying normalization scheme, with particularly large gains under challenging online regimes such as imbalanced label shifts, single-sample adaptation, mixed corruption streams and long-horizon continual TTA.

cs.CV↗
arXiv2026

A Neural Operator Emulator for Coastal and Riverine Shallow Water Dynamics

Coastal regions and river floodplains are particularly vulnerable to the impacts of extreme weather events. Accurate real-time forecasting of hydrodynamic processes in these areas is essential for infrastructure planning and climate adaptation. Yet high-fidelity numerical models are often too computationally expensive for real-time use, and lower-cost approaches, such as traditional model order reduction algorithms or conventional neural networks, typically struggle to generalize to out-of-distribution conditions. In this study, we present the Multiple-Input Temporal Operator Network (MITONet), a novel autoregressive neural emulator that employs latent-space operator learning to efficiently approximate high-dimensional numerical solvers for complex, nonlinear problems that are governed by time-dependent, parameterized partial differential equations. We showcase MITONet's predictive capabilities by forecasting regional tide-driven dynamics in the Shinnecock Inlet in New York and riverine flow in a section of the Red River in Louisiana, both described by the two-dimensional shallow-water equations (2D SWE), while incorporating initial conditions, time-varying boundary conditions, and domain parameters such as the bottom friction coefficient. Despite the distinct flow regimes, the complex geometries and meshes, and the wide range of bottom friction coefficients studied, MITONet displays consistently high predictive skill, with anomaly correlation coefficients above 0.9, a maximum normalized root mean square error of 0.011, and computational speedups between 100x-1,250x, even for 175 days of autoregressive rollout forecast from random initial conditions and with unseen parameter values.

cs.CE↗
arXiv2025

Exploring Multimodal Perception in Large Language Models Through Perceptual Strength Ratings

This study investigated whether multimodal large language models can achieve human-like sensory grounding by examining their ability to capture perceptual strength ratings across sensory modalities. We explored how model characteristics (size, multimodal capabilities, architectural generation) influence grounding performance, distributional factor dependencies (word frequency, embeddings, feature distances), and human-model processing differences. We evaluated 21 models from four families (GPT, Gemini, LLaMA, Qwen) using 3,611 words from the Lancaster Sensorimotor Norms through correlation, distance metrics, and qualitative analysis. Results showed that larger (6 out of 8 comparisons), multimodal (5 of 7), and newer models (5 of 8) generally outperformed their smaller, text-based, and older counterparts. Top models achieved 85-90% accuracy and 0.58-0.65 correlations with human ratings, demonstrating substantial similarity. Moreover, distributional factors showed minimal impact, not exceeding human dependency levels. However, despite strong alignment, models were not identical to humans, as even top performers showed differences in distance and correlation measures, with qualitative analysis revealing processing patterns related to absent sensory grounding. Additionally, it remains questionable whether introducing multimodality resolves this grounding deficit. Although multimodality improved performance, it seems to provide similar information as massive text rather than qualitatively different data, as benefits occurred across unrelated sensory dimensions and massive text-only models achieved comparable results. Our findings demonstrate that while advanced LLMs can approximate human sensory-linguistic associations through statistical learning, they still differ from human embodied cognition in processing mechanisms, even with multimodal integration.

cs.CL↗
arXiv2025

Characterization of GPU TEE Overheads in Distributed Data Parallel ML Training

Confidential computing (CC) or trusted execution enclaves (TEEs) is now the most common approach to enable secure computing in the cloud. The recent introduction of GPU TEEs by NVIDIA enables machine learning (ML) models to be trained without leaking model weights or data to the cloud provider. However, the potential performance implications of using GPU TEEs for ML training are not well characterized. In this work, we present an in-depth characterization study on performance overhead associated with running distributed data parallel (DDP) ML training with GPU Trusted Execution Environments (TEE). Our study reveals the performance challenges in DDP training within GPU TEEs. DDP uses ring-all-reduce, a well-known approach, to aggregate gradients from multiple devices. Ring all-reduce consists of multiple scatter-reduce and all-gather operations. In GPU TEEs only the GPU package (GPU and HBM memory) is trusted. Hence, any data communicated outside the GPU packages must be encrypted and authenticated for confidentiality and integrity verification. Hence, each phase of the ring-all-reduce requires encryption and message authentication code (MAC) generation from the sender, and decryption and MAC authentication on the receiver. As the number of GPUs participating in DDP increases, the overhead of secure inter-GPU communication during ring-all-reduce grows proportionally. Additionally, larger models lead to more asynchronous all-reduce operations, exacerbating the communication cost. Our results show that with four GPU TEEs, depending on the model that is being trained, the runtime per training iteration increases by an average of 8x and up to a maximum of 41.6x compared to DDP training without TEE.

cs.CR↗
arXiv2025

Improved methodology for deep aquifer characterization using hydrogeological, self-potential, and magnetotellurics data

Estimating subsurface properties like hydraulic conductivity using hydrogeological data alone is challenging in field sites with sparse wells. Geophysical data, including Self-potential (SP) and Magnetotelluric (MT), can improve understanding of hydrogeological structures and interpolate data between wells. However, determining hydraulic conductivity requires a proper petrophysical relationship between hydraulic conductivity and inferred geophysical properties, which may not exist or be unique. In this work, we propose a joint-inversion approach without assuming petrophysical relationships, using self-potential data to connect groundwater flow velocity to electrical potential differences, and MT data to estimate hydraulic conductivity and electrical conductivity. A spectral method is employed for the self-potential forward problem. To accelerate joint data inversion, a dimension reduction technique through the Principal Component Geostatistical Approach is used. The applicability and robustness of the joint-inversion method are demonstrated through inversion tests using hydrogeophysical data sets generated from subsurface models with and without petrophysical relationships. The joint hydraulic head-SP-MT data inversion can reasonably estimate hydraulic conductivity and electrical resistivity, even without knowledge of a one-to-one petrophysical relationship. On average, joint inversion yields a 25% improvement in hydraulic conductivity estimates compared to single data-type inversion. Our proposed joint inversion approach, with SP-data compensating for the absence of a known petrophysical relationship, provided close agreement with joint inversion of head and MT data using a known petrophysical relationship. Successful inversion tests demonstrate the usefulness of SP data in connecting hydrogeological properties and geophysical data without requiring a petrophysical relationship.

physics.geo-ph↗
arXiv2024

Multiplicative Thom-Sebastiani for Bernstein-Sato polynomials

We show that if $f\in \mathcal{O}_X(X)$ and $g\in \mathcal{O}_Y(Y)$ are nonzero regular functions on smooth complex algebraic varieties $X$ and $Y$, then the Bernstein-Sato polynomial of the product function $fg \in \mathcal{O}_{X\times Y}(X \times Y)$ is given by $b_{fg}(s)=b_f(s)b_g(s)$, answering a question of Budur and Popa.

math.AG↗
arXiv2024

Fastrack: Fast IO for Secure ML using GPU TEEs

As cloud-based ML expands, ensuring data security during training and inference is critical. GPU-based Trusted Execution Environments (TEEs) offer secure, high-performance solutions, with CPU TEEs managing data movement and GPU TEEs handling authentication and computation. However, CPU-to-GPU communication overheads significantly hinder performance, as data must be encrypted, authenticated, decrypted, and verified, increasing costs by 12.69 to 33.53 times. This results in GPU TEE inference becoming 54.12% to 903.9% slower and training 10% to 455% slower than non-TEE systems, undermining GPU TEE advantages in latency-sensitive applications. This paper analyzes Nvidia H100 TEE protocols and identifies three key overheads: 1) redundant CPU re-encryption, 2) limited authentication parallelism, and 3) unnecessary operation serialization. We propose Fastrack, optimizing with 1) direct GPU TEE communication, 2) parallelized authentication, and 3) overlapping decryption with PCI-e transmission. These optimizations cut communication costs and reduce inference/training runtime by up to 84.6%, with minimal overhead compared to non-TEE systems.

cs.CR↗
arXiv2024

Efficient Diffusion-Driven Corruption Editor for Test-Time Adaptation

Test-time adaptation (TTA) addresses the unforeseen distribution shifts occurring during test time. In TTA, performance, memory consumption, and time consumption are crucial considerations. A recent diffusion-based TTA approach for restoring corrupted images involves image-level updates. However, using pixel space diffusion significantly increases resource requirements compared to conventional model updating TTA approaches, revealing limitations as a TTA method. To address this, we propose a novel TTA method that leverages an image editing model based on a latent diffusion model (LDM) and fine-tunes it using our newly introduced corruption modeling scheme. This scheme enhances the robustness of the diffusion model against distribution shifts by creating (clean, corrupted) image pairs and fine-tuning the model to edit corrupted images into clean ones. Moreover, we introduce a distilled variant to accelerate the model for corruption editing using only 4 network function evaluations (NFEs). We extensively validated our method across various architectures and datasets including image and video domains. Our model achieves the best performance with a 100 times faster runtime than that of a diffusion-based baseline. Furthermore, it is three times faster than the previous model updating TTA method that utilizes data augmentation, making an image-level updating approach more feasible.

cs.CV↗