arXiv subjects

Junyoung Park

Publications and source records attributed to Junyoung Park.

At least 19 recordsLinked to original sources

arXiv2026

Replay Pruning and Few-Bit Integer Inference for Hardware-Efficient Speed-Sign Recognition

Speed-limit sign recognition on an automotive camera system-on-chip (SoC) must fit a detector and a classifier (the reader) into the memory, arithmetic, and energy budgets of the device. In our two-stage recognizer, the detector, a YOLO-family network with ReLU and hard-sigmoid operators, is compressed by channel replay: an iterative pruning cascade identifies the channels to remove, the recorded deletions are replayed on the original task-trained network, the parent, and one continuous knowledge-distillation recovery follows. On a YOLOv5n, replay removes 60.3% of the parameters at the accuracy of the parent and, in validation mAP50-95 over five matched seeds, outperforms iterative pruning and training from scratch, and its per-step variant (per-step replay) is the most accurate pruning method. On a YOLOv8n, replay and per-step replay outperform every other method and the unpruned parent. The reader, a five-layer CNN with 18k parameters, and the post-processing of the YOLOv5n and YOLOv8n heads use integer arithmetic only. With 4-bit weights and 8-bit activations (W4A8), the reader matches FP32 accuracy in 12.1 KB of on-chip memory, and an executable streaming model reproduces the INT8 class predictions of its runtime on all 9,850 test crops. On a Jetson AGX Orin, INT8 is faster than FP16 only in the implicit formulation of TensorRT, and the replay detector runs fastest on the deep-learning accelerator (DLA). Trained on public data only, the combined system of 0.72 M parameters recognizes 95.2% of the signs of a held-out automotive camera at 90% precision and 94 frames per second.

cs.CV↗

arXiv2026

From Simulated Citizens to Simulated Deliberation: Challenges in Representation and Interaction

Can LLM agents simulate how a real public would deliberate? For such simulations to be meaningful, agents must reflect the population's opinions, and their views must be shaped by what others say during discussion. We evaluate these requirements across six LLMs using national surveys, controlled deliberation experiments, and a real deliberative poll. Comparing census-grounded Korean personas against national surveys, we find that agents often produce overly concentrated opinions and fail to reliably reproduce demographic differences. During simulated deliberation, agents exchange arguments, respond to one another, and frequently change positions. Yet similar changes occur even when agents cannot see one another's contributions, while simply restating their starting positions sharply reduces movement. We further test whether these limitations persist in a real-world setting, using agents constructed from actual respondents in the 2017 Shin-Kori deliberative poll. Even with real respondent profiles, agents fail to reproduce the panel's initial opinion distribution or how its opinions change across survey waves. Our findings show that generating deliberative discussions, representing public opinion, and reproducing interaction-driven opinion change are distinct capabilities. These results suggest a different potential role for LLM deliberation: surfacing arguments and perspectives, rather than simulating a real public.

cs.AI↗

arXiv2026

ARMED: Adaptive Retrieval-Augmented Medication Recommendation

Medication recommendation requires jointly considering diagnoses, comorbidities, longitudinal patient history, and treatment interactions. Although recent retrieval-augmented generation approaches using language models have shown promise for clinical recommendation, most existing systems rely on fixed retrieval pipelines that cannot adapt to diverse clinical retrieval scenarios and may introduce noisy or irrelevant evidence. We propose ARMED, an adaptive clinical RAG framework that dynamically expands retrieval according to retrieval sufficiency and current-admission relevance. ARMED first observes the initial retrieval outcome and selectively performs supplementary or fallback multi-query retrieval when additional evidence is needed. The framework also incorporates a relevance gate that activates only clinically relevant same-patient history and organizes retrieved evidence into structured query-level clinical evidence. Experiments on MIMIC-IV demonstrate that ARMED consistently outperforms previous static retrieval baselines and improves overall medication recommendation quality. By combining locally runnable small language models with adaptive retrieval, ARMED supports medication recommendation without relying on external language model APIs or task-specific fine-tuning.

cs.LG↗

arXiv2026

ValueDiff: Value-Geometric KV Cache Eviction for Sink-Suppressed LLMs

Modern LLMs with QK-normalization, gated attention, learned attention sinks, or logit softcapping exhibit weaker persistent attention sinks, on which existing KV cache eviction methods primarily rely. We observe that across these models, weaker sinks co-occur with greater value-vector dispersion relative to key-vector dispersion. Motivated by this value-side dispersion, we present ValueDiff, a value-geometric eviction that ranks tokens by the L2 deviation of their value vectors from the cache mean. The same score arises as the minimal-disturbance eviction under a max-entropy assumption about future attention. We evaluate under fixed cache budgets, with eviction at every block boundary during prefill and at every decoding step during generation. On RULER at a tight 2k token budget, ValueDiff retains 88-99% of dense across seven sink-suppressed models (best on 6 out of 7). On LongBench at the 4k budget, ValueDiff averages 92% retention across sink-suppressed models versus 83% for the strongest prior baseline. On MATH-500, ValueDiff is the strongest non-dense method on every sink-suppressed model tested at the 25% cache budget, outperforming prior methods by up to ~20 points on gated-attention models. Across all three benchmarks, value geometry emerges as the more reliable query-invariant eviction signal for sink-suppressed models.

cs.LG↗

arXiv2026

Semantic Watermarking for Malicious Image Manipulation Detection

The proliferation of high-fidelity generative editing models has made it possible to inject violent or sexual content into otherwise ordinary images while preserving visual plausibility, with concrete consequences for public discourse and vulnerable populations. We propose a robust semantic watermarking framework that reframes the watermark as a recoverable semantic reference rather than an opaque identifier. Our framework combines a $β$-VAE-based binary watermark (CLIP-VAE) with explicit channel-aware training---random bit-flip noise is injected during training so that the decoder learns graceful degradation under the noisy watermarking channel. As a downstream application, a lightweight module SDA-Net uses the recovered semantic embedding to expose not only whether but in which semantic direction an image has been altered. In a 5-way comparison against representative binary hashing baselines (SimHash, ITQ, HashNet, and their robust-MLP variants), CLIP-VAE achieves the highest reconstruction cosine similarity to the original CLIP embedding under realistic InstructPix2Pix bit-error rates, and uniquely supports direction-of-drift detection---a forensic complement to existing content-moderation pipelines.

cs.CV↗

arXiv2026

Same Bit Width, Different Outcomes: Post-Training Quantization of Text-to-Speech Across Architectures

Post-training quantization (PTQ) reduces the cost of on-device text-to-speech (TTS), but published evaluations cover one system or method. We evaluate PTQ across TTS architectures under one protocol with three core models, weight and activation ablations of eight more, and two held-out models quantized blind. Four-bit per-channel weights reduce UTMOS, a predicted mean opinion score, by 2.8 on Supertonic and 0.07 on Kokoro, and per-tensor scaling can cause severe degradation even at 8 bits. The same bit width yields different outcomes, because the sensitive component is model-specific and not reliably predicted from the model class. A staged ablation procedure identifies it, and per-layer GPTQ can restore it to within 0.1 UTMOS. Real int8 and int4 kernels reproduce the simulated ordering at hardware-dependent cost. On a Mac mini, a 4-bit weight kernel runs Supertonic at 0.60x the fp32 latency while int8 is slower, so each configuration requires validation on the target runtime.

eess.AS↗

arXiv2026

Personalized Korean Lipreading as Visual Speech Recognition: Transfer, Census and Adaptation on OLKAVS

We present a personalized Korean visual speech recognition (VSR) system and quantify, on the nine-camera OLKAVS corpus, the gap between the population-level benchmark score and an individual user's error. A video-only Conformer initialized from English-trained weights attains 9.95 - 12.19% character error rate (CER) under the corpus protocol against the published 26.64, and 19.00 - 21.52 on unseen wording. Per speaker, CER spans 1.0 to 52.2%, with seen wording lowering CER by 7.0 - 9.0 points and professional delivery and spontaneous speech raising it by 8.5 - 10.5 and 12.7 points. A low-rank adapter with 4.6% of the parameters, trained on 4 to 29 minutes of the user's frontal video, lowers the CER of twelve high-error speakers by 2.13 to 3.58 points, transfers to every camera without loss, and keeps 85% of the full fine-tuning gain at 12% of its cost to other speakers. Cameras above the mouth plane add about six CER points as a constant offset that training on all views keeps small.

eess.AS↗

arXiv2026

Ancient mean curvature flow asymptotic to a minimal quadratic cone

In this paper, we prove that a smooth, properly embedded ancient mean curvature flow that is asymptotic to an $O(n)\times O(m)$ symmetric minimal quadratic cone for $n +m \geq 10$, and lies on one side of the cone has to have unique asymptotics in the parabolic region. When additionally assuming mean convexity, we upgrade unique asymptotics to full uniqueness, and show that such flow has to be a stationary flow given by one of the leaves of the Hardt-Simon foliation. This is the first rigidity/unique asymptotics theory for ancient mean curvature flow with a $\textbf{singular minimal cone as the asymptotic model}$.

math.DG↗

arXiv2026

Fast distance computation of multivariate distributions via nonparanormal transport

With the increasing availability of data objects in the form of probability distributions, there is a growing need for statistical methods tailored to distributional data. Distance measures, especially the pairwise distance matrix between data objects, provide the foundation for a wide range of modern data analysis methods, such as clustering, multidimensional scaling, and distance-based regression, among others. The Wasserstein distance is commonly used with distributional data due to its compelling optimal transport property. However, while the Wasserstein distance can be efficiently computed for univariate distributions, its application to multivariate distributions is limited due to high computational costs. To address these scalability issues, we introduce the Nonparanormal Transport (NPT) metric, a closed-form distance based on the flexible nonparanormal distribution family for modeling skewed and non-Gaussian multivariate data. Simulation studies demonstrate that NPT maintains a high level of agreement with the Wasserstein distance, while being at least 1000 times faster than its efficient variants when computing a 100-distribution pairwise distance matrix in both 2 and 5 dimensions. We illustrate the utility of NPT through a multidimensional scaling analysis of bivariate oxygen desaturation distributions of 723 individuals with sleep apnea in the Sleep Heart Health Study.

stat.ME↗

arXiv2026

Foliation by bi-rotational self shrinkers in Euclidean space

In this note, we prove that there exist two families of bi-rotational asymptotically conical self shrinkers with boundary called `trumpets', and two families of bi-rotational self shrinkers with boundary which close up called `disks', so that these families together with two generalized self shrinking cylinders, and a minimal cone foliate the whole space except some compact set containing the origin.

math.DG↗

arXiv2026

A.X K2 Technical Report

We introduce A.X K2, a 688B-parameter Mixture-of-Experts (MoE) language model trained from scratch as a high-performance foundation for \emph{agentic} applications. Trained on approximately 8.5T tokens---fewer than its predecessor, A.X K1---on a smaller but higher-quality mixture with substantially expanded agentic and software-engineering data, it nonetheless improves over A.X K1 across the board, by over 30 percentage points on some benchmarks, reflecting large gains in token efficiency. To support long contexts efficiently, we introduce Sparse Gated Attention (SGA), which combines sparse attention with gated attention, and adopt Gated Norm (GN) to stabilize large-scale training. SGA is trained natively at 128K through a \emph{sparse} indexer warmup that optimizes the indexer against its own sparse top-$k$ selection rather than the dense attention distribution, making adaptation markedly cheaper: each query reads only 2,048 positions, yet long-context quality is unchanged and A.X K2 scores 94.6 on RULER out to 256K. The outlier suppression of GN in turn keeps 4-bit NVFP4 serving within one point of FP8 accuracy. A simple yet effective Think-Fusion recipe further lets users switch between thinking and non-thinking modes within a single unified model. Extensive evaluations show that A.X K2 performs competitively against strong open-weight baselines, matching or exceeding them on math and Korean-language benchmarks.

cs.AI↗

arXiv2026

LoViF 2026 The First Challenge on Unified Removal of Raindrops and Reflections: Methods and Results

This workshop paper comprehensively reviews the First Challenge on Unified Removal of Raindrops and Reflections. The challenge aims to address a frequently encountered practical problem in the field of autonomous driving, i.e., raindrop-reflection composite degradation on rainy days. This competition attracted 149 registered participants and received 12 valid final submissions with corresponding fact sheets, significantly contributing to the progress of unified removal of raindrops and reflections. All the methods are developed and evaluated on our real-shot RainDrop and ReFlection (RDRF) dataset. A detailed analysis of the submitted methods and corresponding results is provided in this report, which highlights effective approaches and provides interesting insights for future research.

cs.CV↗

arXiv2026

Fréchet regression of multivariate distributions with nonparanormal transport

Regression with distribution-valued responses and Euclidean predictors has gained increasing scientific relevance. While methodology for univariate distributional data has advanced rapidly in recent years, multivariate distributions, which additionally encode dependence across univariate marginals, have received less attention and pose computational and statistical challenges. In this work, we address these challenges with a new regression approach for multivariate distributional responses, in which distributions are modeled within the semiparametric nonparanormal family. By incorporating the nonparanormal transport (NPT) metric -- an efficient closed-form surrogate for the Wasserstein distance -- into the Fréchet regression framework, our approach decomposes the problem into separate regressions of marginal distributions and their dependence structure, facilitating both efficient estimation and granular interpretation of predictor effects. We provide theoretical justification for NPT, establishing its bi-Lipschitz equivalence to the Wasserstein distance and proving that it mitigates the curse of dimensionality. We further prove uniform convergence guarantees for regression estimators, both when distributional responses are fully observed and when they are estimated from empirical samples, attaining fast convergence rates comparable to the univariate case. The utility of our method is demonstrated via simulations and an application to continuous glucose monitoring data.

stat.ME↗

arXiv2026

Beyond fixed thresholds: optimizing summaries of wearable device data via piecewise linearization of quantile functions

Wearable devices, such as actigraphy monitors and continuous glucose monitors (CGMs), capture high-frequency data, which are often summarized by the percentages of time spent within fixed thresholds. For example, actigraphy data are categorized into sedentary, light, and moderate-to-vigorous activity, while CGM data are divided into hypoglycemia, normoglycemia, and hyperglycemia based on a standard glucose range of $70\unicode{x2013}180$ mg/dL. Although scientific and clinical guidelines inform the choice of thresholds, it remains unclear whether this choice is optimal and whether the same thresholds should be applied across different populations. In this work, we define threshold optimality with loss functions that quantify discrepancies between the full empirical distributions of wearable device measurements and their discretizations based on specific thresholds. We introduce two loss functions: one that aims to accurately reconstruct the original distributions and another that preserves the pairwise sample distances. Using the Wasserstein distance as the base measure, we reformulate the loss minimization as optimal piecewise linearization of quantile functions. We solve this optimization via stepwise algorithms and differential evolution. We also formulate semi-supervised approaches where some thresholds are predefined based on scientific rationale. Applications to CGM data from two distinct populations$\unicode{x2014}$individuals with type 1 diabetes and those with normal glycemic control$\unicode{x2014}$demonstrate that data-driven thresholds vary by population and improve discriminative power over fixed thresholds.

stat.ME↗

arXiv2026

MJ: Multi-turn LLM Jailbreaking via Decomposed Credit Assignment

Modern large language models (LLMs) operate in interactive multi-turn settings, making multi-turn jailbreaking a realistic threat model and an important setting for automated red teaming. A core challenge in learning multi-turn jailbreak attackers is credit assignment: different turns contribute differently to the final outcome, yet existing learning signals are often too coarse to identify their individual contributions. We propose decomposed credit GRPO (DC-GRPO), a unified turn-level credit assignment framework for Group Relative Policy Optimization in multi-turn jailbreak learning. DC-GRPO assigns a separate group-relative learning signal to each turn by combining immediate and future credit, avoiding the credit misassignment induced by broadcasting a single trajectory-level score across the dialogue. We instantiate this framework with static and dynamic weighting rules that differ in how the two credit sources are balanced while sharing the same turn-level structure. Across multiple victim LLMs and benchmarks, the dynamic- and static-weighted variants achieve average ASR5@3 scores of 98.26% and 97.88%, respectively, substantially outperforming the state-of-the-art methods, including SEMA (86.58%) and TROJail (86.23%). Their consistently strong performance indicates that the central empirical benefit comes from turn-level group-relative credit assignment rather than a particular weighting rule. Warning: This paper contains examples of harmful content.

cs.CL↗

arXiv2026

REAN: Reconstruction-aware ECG Anonymization Based on Privacy--Utility Orthogonality

A shared electrocardiogram (ECG) is itself a biometric fingerprint that can re-identify a patient and reveal personal information. Recent ECG anonymizers transform the signal before sharing to reduce privacy leakage. However, existing methods still face a privacy--utility trade-off, in which preserving privacy often compromises utility while preserving utility reveals personal information. We propose \emph{REAN} (\emph{RE}construction-aware ECG \emph{AN}onymizer), a raw ECG signal anonymizer, to address this privacy--utility trade-off. REAN reconstructs the signal using a 1-D U-Net trained with losses from frozen privacy and utility classifiers to reduce privacy leakage while preserving utility. The privacy and utility gradients are near-orthogonal ($\approx$93.8$^\circ$), so reducing privacy leakage leaves utility almost unchanged. On four public PhysioNet databases, REAN achieves the strongest privacy--utility balance among raw ECG signal baselines. It drives re-identification to chance (0.96$\to$0.00), keeps arrhythmia macro-AUROC at the clean level (Clean 0.9982 vs.\ REAN 0.9991), and maintains re-identification protection under unseen privacy-classifier architectures.

cs.CR↗

arXiv2026

Geometry-preserving and interpretable dimension reduction for compositional data

High-dimensional compositional data pose unique statistical challenges due to the simplex constraint and excess zeros. While dimension reduction is indispensable for analyzing such data, conventional approaches often rely on log-ratio transformations that compromise interpretability and distort the data through ad hoc zero replacements. To address these issues, we introduce a geometry-preserving framework for dimension reduction of compositional data, mapping high-dimensional compositions directly to a lower-dimensional simplex. This framework is interpretable as a softened amalgamation of compositions and enables dual visualization -- showing both projected data and how variables contribute to reduced components -- for at-a-glance interpretation. Within this geometry, we define a new sufficient dimension reduction (SDR) approach for compositional predictors, whose identifiable object, termed the central compositional subspace, differs from the classical central subspace in Euclidean SDR. For estimation, we propose a kernel-based method that yields sparse solutions and comes with an intrinsic predictive model for direct downstream analyses. We prove consistency through a new subspace-comparison argument that allows the estimated and target subspaces to have different dimensions. Applications to real microbiome datasets demonstrate that our approach provides a powerful graphical exploration tool for uncovering meaningful biological patterns in high-dimensional compositional data.

stat.ME↗

arXiv2026

Persona Attack: Incremental Memory Injection Jailbreak Attack against Large Language Models

As Large Language Models evolve for user convenience, vulnerability to jailbreak attacks continues to be reported despite ongoing efforts in safety training. Traditional jailbreak techniques typically focus on a single prompt injection, neglecting the models' ability to remember the flow of conversation and the user's instructions. In this paper, we propose Persona Attack, a memory injection based jailbreak method that manipulates the model's context window through a step by step approach. Experimental results from applying Persona Attack to several widely used LLMs reveal that, as injections accumulate in memory, models increasingly prioritize these instructions over their internal safety alignment mechanisms. Furthermore, our experiments empirically demonstrate that the attack success rate varies not only according to the memory implementation of the model, but also combinations of instructions and can reach 95% under specific instruction configurations.

cs.CR↗