Deep Learning Engineer · Dnotitia
I work on multimodal models and LLM inference at Dnotitia, currently focused on KV cache compression and multilingual multimodal retrieval.
I received my M.S. in AI at Seoul National University with Prof. Nojun Kwak at the MIPAL lab, and my B.S. in Computer Science & International Studies at KyungHee University.
Research interests: LLM inference efficiency, multimodal foundation models, multilingual/multimodal alignment, video understanding.
Updates
Publications
Preprint, 2026
project · arXiv · model (coming soon)
Conditional Program stores each n-gram memory row as a two-bit trellis code plus a norm-restoring scale, decoding only the rows an input requests; it lowers output KL by 41–63% vs. 2-bit RTN and shrinks DeepSeek's table from 203 GB to 51 GB.
Preprint, 2026
PAV predicts the consequences of a candidate action prefix in the VLA's own representation space and scores whether it stays compatible with task success, choosing how much of each action chunk to execute; it improves success-query trade-offs for π0 and π0.5 on LIBERO, RoboTwin 2.0 and real-world manipulation.
Preprint, 2026
CF-Prune is a training-free method that prunes future-frame tokens in world action models using instruction and action attention, keeping dense computation in precision phases; at 35% pruning it speeds up inference by 1.19–1.27× while improving task success.
NeurIPS 2026, 2026
project · arXiv (coming soon)
CoSE leaves the deployed embedding structurally untouched and appends a jointly trained LoRA side channel, so legacy indices stay valid while a single α knob trades general retention for domain retrieval.
EMNLP 2026 (Main Conference), 2026
The low-resource retrieval gap is not the linear language direction but the EOS hidden state's forward-path trajectory: swapping that single row three blocks before the projector lifts Swahili R@1 from 22.1% to 69.1%.
Preprint, 2026
We apply Okapi BM25 scoring to sparse visual words from a Sparse Autoencoder on ViT patch tokens, enabling efficient, interpretable image retrieval that matches dense accuracy via a two-stage pipeline.
Under Review, 2025
CacheFocus improves long-context handling in LLMs by optimizing cache reuse and reducing inference latency without additional training. Layer-Adaptive Cache Pruning and Adaptive Positional Allocation maintain strong performance even with extended input lengths.
Expert Systems with Applications, 2025
GAIL adjusts per-layer gradient updates by priority for few-shot learning, aligning gradients across tasks to improve convergence speed and accuracy over prior inner-loop adaptation methods.
EMNLP 2024 Industry Track, 2024
MERLIN uses LLMs in a training-free iterative feedback loop to refine text-video retrieval, boosting Recall@1 by aligning user queries with video content across datasets.
CVPR Workshop on Efficient Large Vision Models, 2024
We adapt the pre-trained multimodal CLIP encoder for video highlight detection via fine-tuning and a saliency pooling technique, achieving state-of-the-art performance.
NeurIPS 2023 Workshop on Advancing Neural Network Training, 2023 Oral
We expedite ViT inference by concatenating abstract visual tokens from multiple images along dim=1 and processing them collectively.
CVPR, 2023
project · code · video · arXiv
We model a ray with a mixture density model to learn density distributions efficiently with sparse inputs, and propose ray-depth estimation as an auxiliary task for few-shot novel view synthesis.
ECCV, 2022
For few-shot (n<100) image synthesis, we propose latent-space smoothing regularizations that empower the generator to produce diverse, perceptually continuous samples — without directly combatting memorization.
IEEE Access, 2023
We expose a limitation of multi-domain image translation models in performing recursive translations, and resolve it via additive perturbations during training that also improve overall translation quality.
ICPRAM, 2022
A new method for video frame interpolation using an invertible U-Net based Generative Flow that avoids optical-flow techniques while preserving temporal consistency and image quality.
Background
LLM inference efficiency and multilingual multimodal retrieval.
Multimodal short-form video AI and product-search embedding models for enterprise Korean retailers.
Advised by Prof. Nojun Kwak at MIPAL.