Hankyul Kang

Hankyul Kang

Ph.D. Candidate, Department of Artificial Intelligence
Ajou University, Suwon, Korea
hankyulkang at ajou.ac.kr / hankyulkang1997 at gmail.com
Google Scholar · LinkedIn · hankyul2.github.io

About

I am Hankyul Kang, a Ph.D. candidate in the Department of Artificial Intelligence at Ajou University in Suwon, Korea, where I am advised by Prof. Jongbin Ryu in the Multi-Modal Artificial Intelligence Lab. I joined the lab in January 2021 as an undergraduate researcher and have stayed there since, moving into the doctoral program in March 2022.

I am interested in advancing the frontiers of state-of-the-art algorithms to achieve breakthrough performance. My expertise lies in improving efficiency through architectures, compression, and learning algorithms for vision models and large language models. My work spans four directions:

Education

Research Experience

News

Publications (C: conference, J: journal, S: submission or preprint)

Sequential pruning-then-quantization compared with differentiable bit-width co-optimization
[C.3] Hankyul Kang and Jongbin Ryu. Differentiable Bit-Widths: Co-Optimizing Pruning and Quantization via SVD for Ultra-Efficient LLM Compression.
Advances in Neural Information Processing Systems (NeurIPS), 2026

Keywords: “LLM compression, pruning, quantization, differentiable bit-width, ultra-low-bit memory footprint, batch-size agnostic throughput gain”

Forgetting curves for short-term, optimal, and long-term recall intervals
[C.2] Hankyul Kang, Gregor Seifer, Donghyun Lee, and Jongbin Ryu. Do Your Best and Get Enough Rest for Continual Learning.
IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

Keywords: “continual learning, forgetting curve, recall interval, long-term memory”

Multi-token attention pooling with inter- and intra-group distillation across two token branches
[C.1] Hankyul Kang and Jongbin Ryu. Enriching Local Patterns with Multi-Token Attention for Broad-Sight Neural Networks.
IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), 2025

Keywords: “global average pooling, multi-token attention, attention pooling, high throughput network, low parameter overhead”

U-Net with a deconvolution decoder compared with ADNet's attentional decoder and harmonic pooling
[J.1] Hankyul Kang, Namkug Kim, and Jongbin Ryu. Attentional Decoder Networks for Chest X-Ray Image Recognition on High-Resolution Features.
Computer Methods and Programs in Biomedicine, 2024

Keywords: “chest X-ray recognition, encoder-decoder, attentional upsampling, resource efficient fine-tuning”

Naive multi-layer model compressed into a merged layer using layer imitation
[S.2] Zihao Ye, Hankyul Kang, et al. Evaluation of Winning Solutions of 2025 Low Power Computer Vision Challenge.
arXiv preprint, 2026

Keywords: “low-power computer vision, model compression, layer imitation, mobile NPU”

Baseline multi-head self-attention compared with decomposed and interactive attention variants
[S.1] Hankyul Kang, Ming-Hsuan Yang, and Jongbin Ryu. Interactive Multi-Head Self-Attention with Linear Complexity.
arXiv preprint, 2024

Keywords: “multi-head self-attention, cross-head interaction, linear complexity, scalable architecture, real-world constraint optimization, runtime, memory”

Working Papers

Token throughput comparison on GPQA and GSM8K across five quantized model configurations
[S.7] Anonymous Authors. LoadMatch: Balancing Memory and Compute Load in Quantized Models Using Speculative Decoding.
Working paper

Keywords: “quantized models, speculative decoding, memory bandwidth, compute load balancing, inference throughput”

Conceptual illustration of the QLoRA learning trajectory drifting away from the stable optimal region
[S.6] Anonymous Authors. An Empirical Study on Learning Trajectories of LoRA Adapters for Quantized Large Language Models.
Working paper

Keywords: “quantized LLM, QLoRA, exponential moving average, low-bit low-rank fine-tuning”

Success rate versus historical frames for streamlined VLA models of different sizes
[S.5] Anonymous Authors. Less Redundancy, More Dynamics: Efficient Multi-Frame Action Generation via Streamlined VLA Models.
Working paper

Keywords: “vision-language-action models, multi-frame action generation, temporal redundancy, efficient inference”

Decision boundaries with and without auxiliary out-of-distribution samples across continual learning tasks
[S.4] Anonymous Authors. Utilizing Auxiliary Datasets for Robust Decision Boundaries in Continual Learning Scenarios.
Working paper

Keywords: “continual learning, decision boundary, out-of-distribution data, long-term memory”

Accuracy versus runtime across four devices for slimmable networks
[S.3] Anonymous Authors. Pluggable Decoder for Slimmable Neural Networks.
Working paper

Keywords: “slimmable networks, layer slimming, pluggable decoder, on-device inference, heterogeneous devices, mobile accelerators, inference latency, training time”

Research Projects

Academic Service

Reviewer: CVPR · ECCV · NeurIPS · BMVC