Youngjun Jun (전영준)

I am a first-year PhD student at KAIST AI advised by Jong Chul Ye. Previously, I completed my M.S. in AI at Yonsei University advised by Seong Jae Hwang, and B.S. in Mathematics at Yonsei University.

My research goal is to build machines with agency in the physical world. By enabling machines to understand the world, anticipate how it will change, and act intelligently within it, I aim to develop systems that can autonomously pursue goals and adapt to the environments they encounter. Toward this goal, I explore generative models as a foundation for understanding, predicting, and acting in the world, with a current focus on action-generative AI for robotic manipulation.

News

• [Sep, 2026] Three papers were released on arXiv
• [Aug, 2026] I started PhD program at KAIST AI
• [Jul, 2026] I attended ICML 2026 (Seoul, South Korea)
• [Jun, 2026] My CVPR 2026 paper won the Best Paper Award (1st Prize) at KCC 2026 Workshop on XAI!
• [Jun, 2026] I attended KCC 2026 (Jeju, South Korea)
• [Jun, 2026] I attended CVPR 2026 (Denver, Colorado)
• [May, 2026] One paper was accepted to ICML 2026 Workshop on Foundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning
• [Apr, 2026] One paper was accepted to ICML 2026
• [Mar, 2026] One paper was accepted to ICLR 2026 Workshop on Real‑World Constrained and Preference‑Aligned Flow‑ and Diffusion‑Based Models
• [Mar, 2026] I joined BISPL Group as a researcher
• [Feb, 2026] I received my Master's degree
• [Feb, 2026] My second paper at Yonsei University was accepted to CVPR 2026!
• [Dec, 2025] One patent was granted
• [Aug, 2025] I started a collaboration with Min-Hung Chen at NVIDIA Research
• [Nov, 2025] One paper was accepted to as an oral presentation at AAAI 2026 Workshop on Artificial Intelligence with Biased or Scarce Data
• [Oct, 2025] One patent was granted
• [Aug, 2025] I started a collaboration with Fu-En Yang at NVIDIA Research
• [Jul, 2025] Two paper were accepted to MICCAI 2025
• [Jun, 2025] One paper was accepted to ICCV 2025
• [Feb, 2025] I attended IPIU 2025 (Jeju, South Korea)
• [Jan, 2025] My first domestic paper was accepted to IPIU 2025
• [Jun, 2025] I attended CVPR 2025 (Nashville, Tennessee)
• [Feb, 2025] I left NEWCURE M after my contract expired
• [Feb, 2025] Two patents were filed
• [Oct, 2024] My first paper at Yonsei University was accepted to WACV 2025!
• [Oct, 2024] I joined NEWCURE M as a AI researcher
• [Aug, 2024] I attended KCCV 2024 (Busan, South Korea)
• [Jul, 2024] One paper was accepted to MICCAI 2024
• [Jun, 2024] I attended OHBM Annual Meeting 2024 (Seoul, South Korea)
• [Jan, 2024] One paper was accepted to IPIU 2024
• [Sep, 2023] I started MS program in AI at Yonsei University
• [Sep, 2023] I joined MICV Group as a research intern

Selected Publications (show all)

2026

DriftOPD: Sequence-Level Reverse-KL Distillation for One-Step VLA Policies

Youngjun Jun, Kyumin Choi, Youngmin Kim, Seonghyun Jin, Sunwoo Park, Jangho Park, Jong Chul Ye

Preprint

Abstract arXiv Project

Vision-Language-Action (VLA) models increasingly rely on action experts that generate short action chunks under receding-horizon control. While chunk-level training is convenient across robot embodiments, it optimizes local action likelihood without explicitly accounting for long-horizon task success. Sequence-level reinforcement learning can address this limitation, but typically requires policy rollouts and closed-loop interaction, which are costly for real-robot manipulation. We introduce DriftOPD, a teacher-free, rollout-free framework for sequence-level on-policy distillation of continuous VLA action experts. We show that the sequence-level reverse Kullback-Leibler (KL) divergence decomposes into a chunk-level reverse-KL term and a future-potential term that captures the long-horizon effect of the current action. DriftOPD optimizes these two terms using a one-step drifting objective and a Q-function critic learned from offline demonstrations, respectively, enabling sequence-level optimization with only offline data and one-step action generation. Across multiple VLA architectures in simulation and real-world manipulation, DriftOPD generally outperforms existing one-step distillation baselines while achieving task success performance comparable to multi-step teacher policies. These results demonstrate that long-horizon behavior can be effectively distilled into one-step VLA action experts without online interaction or a separate teacher.

Adjoint Guidance Flow: Amortized Critic Guidance for VLA Policies

, Jeongsol Kim, Youngjun Jun, Kyumin Choi, Youngmin Kim, Seonghyun Jin, Sunwoo Park, Jangho Park, Kwanyoung Kim, Jong Chul Ye

Preprint

Abstract arXiv Project

Flow-based Vision-Language-Action (VLA) policies are typically trained by behavior cloning and thus do not explicitly optimize long-term task return. Critic guidance steers generation toward higher-value actions, but existing methods differentiate the critic through a one-step surrogate of the sampler and back-propagate a critic ensemble at every flow step. In contrast, here we propose Adjoint Guidance Flow (AGF), which amortizes trajectory-aware critic guidance into a lightweight guidance network while preserving the pretrained VLA policy. Specifically, we formulate critic-guided flow generation as a deterministic optimal control problem, whose optimal guidance is a costate that carries the terminal critic gradient back through the remaining flow, and regress the guidance network onto this costate while keeping both the VLA and critic frozen. This design provides favorable memory and throughput scaling during training, and inference needs one guidance-network forward pass per step, without the critic ensemble, back-propagation, or adjoint computation. Across LIBERO, RoboCasa, and LIBERO-Pro, AGF consistently improves pretrained VLAs, remains competitive with critic-guidance and policy-fine-tuning baselines, and is the most robust method when a single guidance strength is deployed across tasks. Compared with QGF, AGF runs 3.6× faster per guidance step with 7.0× fewer parameters, with comparable and even better performance, showing that critic guidance can be trajectory-aware and lightweight.

Don't Throw Away the Tail: Action Upcycling for Policy Acceleration

Taesung Kwon, Jangho Park, Sunwoo Park, Youngmin Kim, Seonghyun Jin, Youngjun Jun, Kyumin Choi, Jong Chul Ye

Preprint

Abstract arXiv Project

Modern robot policies predict a chunk of future actions from a single observation, execute only a prefix, and discard the rest before replanning. Choosing the length of this prefix, the execution horizon, poses a trade-off between reactivity and efficiency. A short horizon keeps the policy reactive to the environment, but requires frequent policy calls. Recent test-time methods adaptively select the horizon for each chunk, but they either read model internals, where the signal must be chosen for each architecture, or draw extra samples, which adds cost. We propose Action Upcycling, a training-free algorithm that reuses actions the policy would otherwise discard, without accessing model internals or drawing extra samples. We find that discarded actions stay close to their replanned versions as long as the action velocity remains smooth. Action Upcycling therefore extends the execution horizon up to the point where the velocity begins to fluctuate. Extensive experiments on simulated and real-world manipulation tasks show that Action Upcycling reduces policy calls by 1.2-1.7x with no loss in success rate, across multiple Vision-Language-Action Models (VLAs) and even a World Action Model (WAM). It applies to any chunked policy at negligible cost and is orthogonal to other policy acceleration methods such as few-step sampling and streaming action decoding, opening a new axis for policy acceleration.

TBA

Clinically Validated Deep Learning-Based CT Segmentation of White Matter Hyperintensities from Amyloid PET/CT: A Single-Modality Approach to Vascular Burden Assessment in Alzheimer's Disease

In submission

Physics in 2-Steps: Locking Motion Priors Before Visual Refinement Erases Them

Woojung Han, Seil Kang, Youngjun Jun, Min-Hung Chen, Fu-En Yang, Seong Jae Hwang

ICML 2026 & ICML 2026 Workshop (FoGen)

Abstract OpenReview arXiv Project Code

Video diffusion models can generate visually stunning content, yet frequently produce motion that violates physical laws, objects accelerate implausibly or vanish mid-trajectory. We reveal a surprising finding: a 2-step generation often exhibits better physical consistency than a 50-step output from the same model. Through spectral analysis, we trace this to phase erosion during denoising, motion dynamics are 8.5× more sensitive to phase corruption than magnitude, yet the refinement process progressively destroys this critical component. Building on this insight, we propose PhaseLock, a training-free framework that locks motion dynamics to fast inference priors. Rather than requiring 50 steps to establish physics, PhaseLock extracts a motion prior from just 2 steps and enforces it onto high-fidelity generation via Latent Delta Guidance. This decouples physical consistency from visual refinement, ensuring the final output remains grounded in valid trajectories. PhaseLock achieves strong physical consistency with negligible overhead (1.06× time, 1.02× memory), eliminating the need for expensive external guidance methods (∼5× time).

Interpretable Motion-Attentive Maps: Spatio-Temporally Localizing Concepts in Video Diffusion Transformers

Youngjun Jun, Seil Kang, Woojung Han, Seong Jae Hwang

CVPR 2026 Highlight KCC 2026 Workshop (XAI) Best Paper Award (1st Prize)

Abstract arXiv Project Code

Video Diffusion Transformers (DiTs) have been synthesizing high-quality video with high fidelity from given text descriptions involving motion. However, understanding how Video DiTs convert motion words into video remains insufficient. Furthermore, while prior studies on interpretable saliency maps primarily target objects, motion-related behavior in Video DiTs remains largely unexplored. In this paper, we investigate concrete motion features that specify when and which object moves for a given motion concept. First, to spatially localize, we introduce GramCol, which adaptively produces per-frame saliency maps for any text concept, including both motion and non-motion. Second, we propose a motion-feature selection algorithm to obtain an Interpretable Motion-Attentive Map (IMAP) that localizes motion spatially and temporally. Our method discovers concept saliency maps without the need for any gradient calculation or parameter update. Experimentally, our method shows outstanding localization capability on the motion localization task and zero-shot video semantic segmentation, providing interpretable and clearer saliency maps for both motion and non-motion concepts.

Inference-Time CLIP Embedding Manipulation for Compositional Text-to-Image Alignment

Sujung Hong, Tae Eun Choi, Youngjun Jun, Chanyong Yoon, Seong Jae Hwang

ICLR 2026 Workshop (ReALM-GEN)

Abstract OpenReview

Text-to-image (T2I) diffusion models have advanced notably, but still fail to satisfy prompt conditions, resulting in attribute mismatches and missing objects. These errors normally fall into two categories: concept loss, when an object or attribute (i.e., concept) goes missing, and concept confusion, when an attribute is assigned to the wrong object or multiple objects blend together. Recent studies suggest that these challenges are caused by the limited capacity of CLIP text encoder to capture fine-grained semantic details. Although several methods have been proposed, concept loss and concept confusion persist in recent T2I models when handling multi-object, multi-attribute prompts. In this paper, we conduct an embedding-level analysis to develop an inference time, model independent solution, addressing concept loss and concept confusion. We find that (1) concepts mentioned later exhibit higher embedding entropy, indicating higher uncertainty and making them more vulnerable to concept loss, and (2) the first CLIP attention layer captures the strength of binding between each object and its attribute. Guided by our findings, we introduce TIE, a method that improves semantic alignment through a single text-embedding update. TIE addresses concept loss via entropy-aware singular value amplification and resolves concept confusion through interpolation–extrapolation binding based on CLIP attention scores, all in a training-free manner. Extensive experiments demonstrate that TIE enhances semantic fidelity in multi-concept scenarios with minimal sampling overhead.

Backbone Augmented Training for Adaptations

Jae Wan Park, Junhyeok Kim, Youngjun Jun, Hyunah Ko, Seong Jae Hwang

AAAI 2026 Workshop (AIBSD) Oral Presentation

Abstract arXiv

Adaptations facilitate efficient training of large backbone models, including diffusion models for image generation and transformer-based language models. While various adaptation techniques enhance performance with minimal computational resources, limited adaptation data often leads to challenges in training. To address this, we focus on the enormous amount of backbone data used to pre-train the backbone models. We propose Backbone Augmented Training (BAT), a method that leverages backbone data to augment the adaptation dataset. First, we formulate and prove two mathematical key propositions: one establishes the validity of BAT, while the other identifies a condition under which BAT benefits adaptation. Furthermore, we introduce an advanced data selection scheme that satisfies these propositions and present ALBAT algorithm to implement this approach. ALBAT efficiently enhances adaptation training in both personalization and language generation tasks with scarce data.

2025

WAVE: Warp-Based View Guidance for Consistent Novel View Synthesis Using a Single Image

Jiwoo Park, Tae Eun Choi, Youngjun Jun, Seong Jae Hwang

ICCV 2025

Abstract arXiv Project Code

Generating high-quality novel views of a scene from a single image requires maintaining structural coherence across different views, referred to as view consistency. While diffusion models have driven advancements in novel view synthesis, they still struggle to preserve spatial continuity across views. Diffusion models have been combined with 3D models to address the issue, but such approaches lack efficiency due to their complex multi-step pipelines. This paper proposes a novel view-consistent image generation method which utilizes diffusion models without additional modules. Our key idea is to enhance diffusion models with a training-free method that enables adaptive attention manipulation and noise reinitialization by leveraging view-guided warping to ensure view consistency. Through our comprehensive metric framework suitable for novel-view datasets, we show that our method improves view consistency across various diffusion models, demonstrating its broader applicability.

PRGNN: Pyramidal Region Graph Neural Network for Region-Based Brain PET Classification

Daesung Kim, Seungbeom Seo, Boosung Kim, Kyobin Choo, Youngjun Jun, Mijin Yun

MICCAI 2025

Abstract Paper Code

Brain positron emission tomography (PET) has been widely used for the diagnosis of various neurodegenerative diseases. To assist physicians, convolutional neural networks (CNNs) and transformers have been explored for prediction of diseases based on brain PET images. While these models show promising performance, they are designed to process the entire image, which facilitates shortcut learning by extracting irrelevant features. To alleviate shortcut learning, we observe that brain images share the same structure, and regions of interest (ROIs) can be defined for relevant regions. In this regard, we propose Pyramidal Region Graph Neural Network (PRGNN), which employs a 3D convolutional backbone to learn multi-level feature representations and constructs nodes that correspond to anatomical ROIs. Using ROI-based node embeddings, PRGNN extracts metabolic patterns in functionally relevant regions and performs explicit inter-regional reasoning. We evaluate PRGNN on classifying 18F-fluorodeoxyglucose (FDG) and amyloid PET, outperforming models based on CNN, transformer, and GNN.

PRETI: Patient-Aware Retinal Foundation Model via Metadata-Guided Representation Learning

Yeonkyung Lee, Woojung Han, Youngjun Jun, Hyeonmin Kim, Jungkyung Cho, Seong Jae Hwang

MICCAI 2025 Early Accept (Top 9%)

Abstract arXiv Code

Retinal foundation models have significantly advanced retinal image analysis by leveraging self-supervised learning to reduce dependence on labeled data while achieving strong generalization. Many recent approaches enhance retinal image understanding using report supervision, but obtaining clinical reports is often costly and challenging. In contrast, metadata (e.g., age, gender) is widely available and serves as a valuable resource for analyzing disease progression. To effectively incorporate patient-specific information, we propose PRETI, a retinal foundation model that integrates metadata-aware learning with robust self-supervised representation learning. We introduce Learnable Metadata Embedding (LME), which dynamically refines metadata representations. Additionally, we construct patient-level data pairs, associating images from the same individual to improve robustness against non-clinical variations. To further optimize retinal image representation, we propose Retina-Aware Adaptive Masking (RAAM), a strategy that selectively applies masking within the retinal region and dynamically adjusts the masking ratio during training. PRETI captures both global structures and fine-grained pathological details, resulting in superior diagnostic performance.

Disentangling Disentangled Representations: Towards Improved Latent Units via Diffusion Models

Youngjun Jun, Jiwoo Park, Kyobin Choo, Tae Eun Choi, Seong Jae Hwang

WACV 2025

Abstract arXiv Project Code

Disentangled representation learning (DRL) aims to break down observed data into core intrinsic factors for a profound understanding of the data. In real-world scenarios, manually defining and labeling these factors are non-trivial, making unsupervised methods attractive. Recently, there have been limited explorations of utilizing diffusion models (DMs), which are already mainstream in generative modeling, for unsupervised DRL. They implement their own inductive bias to ensure that each latent unit input to the DM expresses only one distinct factor. In this context, we design Dynamic Gaussian Anchoring to enforce attribute-separated latent units for more interpretable DRL. This unconventional inductive bias explicitly delineates the decision boundaries between attributes while also promoting the independence among latent units. Additionally, we also propose Skip Dropout technique, which easily modifies the denoising U-Net to be more DRL-friendly, addressing its uncooperative nature with the disentangling feature extractor. Our methods, which carefully consider the latent unit semantics and the distinct DM structure, enhance the practicality of DM-based disentangled representations, demonstrating state-of-the-art disentanglement performance on both synthetic and real data, as well as advantages in downstream tasks.

2024

Slice-Consistent 3D Volumetric Brain CT-to-MRI Translation with 2D Brownian Bridge Diffusion Model

Kyobin Choo, Youngjun Jun, Mijin Yun, Seong Jae Hwang

MICCAI 2024 Early Accept (Top 11%)

Abstract arXiv Project Code

In neuroimaging, generally, brain CT is a more cost-effective and accessible imaging option compared to MRI. Nevertheless, CT exhibits inferior soft-tissue contrast and higher noise levels, yielding less precise structural clarity. In response, leveraging more readily available CT to construct its counterpart MRI, namely, medical image-to-image translation (I2I), serves as a promising solution. Particularly, while diffusion models (DMs) have recently risen as a powerhouse, they also come with a few practical caveats for medical I2I. First, DMs' inherent stochasticity from random noise sampling cannot guarantee consistent MRI generation that faithfully reflects its CT. Second, for 3D volumetric images which are prevalent in medical imaging, naively using 2D DMs leads to slice inconsistency. While 3D DMs do exist, significant training costs and data dependency bring hesitation. As a solution, we propose novel style key conditioning (SKC) and inter-slice trajectory alignment (ISTA) sampling for the 2D Brownian Bridge diffusion model. Specifically, SKC ensures a consistent imaging style across slices, and ISTA interconnects the independent sampling of each slice, deterministically achieving style and shape consistent 3D CT-to-MRI translation.