Jongwon Ryu

Mr. Jongwon Ryu


Research Focus & Metrics

Academic Career Length

2Years (2025-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
20 October 2025 in IEEE Access
Verified
Journal Article
Open access
JRJongwon Ryu
Jongwon Ryu
MKMingi Kim
Mingi Kim
JKJunyeong Kim
Junyeong Kim
We present a sparse Mixture-of-Experts student architecture for knowledge distillation that replaces conventional dense students with a modular design of lightweight, class-agnostic experts dynamically routed by input-conditioned gating. Our approach provides three key benefits: (i) accuracy–efficiency Pareto superiority over dense students by activating only a subset of experts per input, (ii) loss-agnostic plug-and-play compatibility, consistently improving performance across diverse Knowledge distillation objectives, and (iii) class-agnostic expert decomposition, enabling generalization without label supervision. To ensure fairness, we construct both dense and Mixture-of-Experts students with comparable parameter budgets and systematically evaluate multiple gating strategies, including soft, top-k, and attention-based variants. Experiments on CIFAR-10 and CIFAR-100 under four distillation methods show that attention-gated MoE students consistently outperform dense baselines at equal capacity, achieving up to +1.7% accuracy gains with an average 25% FLOPs reduction. These findings demonstrate the structural advantages of sparse Mixture-of-Experts students as efficient, generalizable, and loss-agnostic alternatives to dense models in knowledge distillation.

11 December 2025 in IEEE Access
Verified
Journal Article
Open access
MKMingi Kim
Mingi Kim
JRJongwon Ryu
Jongwon Ryu
JKJunyeong Kim
Junyeong Kim
Spurious correlations in training data often cause models to underperform on minority groups. Existing methods that improve worst-group accuracy (WGA) typically rely on group annotations, complex optimization objectives or multi-stage pipelines, limiting their practical application.We make a key observation: robust classifiers utilize a sparse set of features, suggesting they selectively focus on core, task-essential features. Based on this insight, we propose a single-stage, annotation-free method that co-trains a learnable binary mask to structurally filter input features. End-to-end training is enabled via a Straight-Through Estimator (STE) and a Gradient Adjustment technique, with the mask learned solely from the main task loss. While many de-biasing methods improve WGA, they often overlook the critical trade-off with overall accuracy. We therefore propose a novel Trade-off Ratio (ToR) metric, which quantifies the efficiency of WGA gains relative to any loss in overall accuracy. On four benchmarks (Waterbirds, CelebA, CivilComments, MultiNLI), our method achieves state-of-the-art WGA on CivilComments of 81.9% on CivilComments and the highest ToR score of 12.8 on Waterbirds.

1 January 2026 in IEEE Access
Verified
Journal Article
Open access
JRJongwon Ryu
Jongwon Ryu
JGJaehoon Go
Jaehoon Go
TPTrung X. Pham
Trung X. Pham
JKJunyeong Kim
Junyeong Kim

6 August 2026
Verified
Book Chapter
Closed access
PDF file
JYJisoo Yang
Jisoo Yang
JRJongwon Ryu
Jongwon Ryu
MMMinuk Ma
Minuk Ma
TPTrung Pham
Trung Pham
JKJunyeong Kim
Junyeong Kim

1 January 2026 in IEEE Access
Verified
Journal Article
Open access
JYJisoo Yang
Jisoo Yang
JRJongwon Ryu
Jongwon Ryu
TPTrung X. Pham
Trung X. Pham
JKJunyeong Kim
Junyeong Kim

28 October 2025 in IEEE Access
Verified
Journal Article
Open access
JRJongwon Ryu
Jongwon Ryu
MKMingi Kim
Mingi Kim
JKJunyeong Kim
Junyeong Kim
Despite significant advances in large language models, many reasoning datasets are still built from a fixed set of predefined relations, manually curated types such as cause, effect, and intent found in knowledge graph datasets such as ATOMIC and COMET. While these predefined relations provide essential structure, the fixed schema limits relational coverage and adaptability to novel contexts. We present DYNA-SKILL, a dual-triple knowledge graph framework that preserves 35 predefined relations consolidated and refined from existing commonsense knowledge graph datasets while augmenting them with 133 additional schema-free dynamic relations generated via a self-prompting mechanism. Each instance consists of two linked triples (Head–Predefined Relation–Tail) and (Tail–Dynamic Relation–Additional Tail) used as independent training samples while retaining linkages for extended reasoning paths. Across reasoning-intensive benchmarks, including CommonsenseQA, RiddleSense, and ARC Challenge, the Hybrid configuration, which combines predefined and dynamically generated relations, achieves performance comparable to or slightly higher than Predefined-only settings and yields up to 3.2% higher accuracy than baseline BERT models. By expanding the relation set from 35 predefined types to a total of 168 relations, DYNA-SKILL enriches relational diversity and improves multi-step logical reasoning, which can enhance performance in real-world scenarios such as complex question answering, multi-document analysis, and causal reasoning, where accurate and adaptable inference is critical.