Publications

Asterisk (*) denotes corresponding author. Curly braces ({}) denote co-first authorship.

2026

Core idea. Retrieves evidence for unresolved window dependencies and refines it into evidence-dense memory for streaming dialogue summarization under a fixed budget.

EMNLP 2026 NLP

Don't Scroll Back: Missing-Evidence Memory for Streaming Dialogue Summarization

Hyangsuk Min, Hwanjun Song*

Conference on Empirical Methods in Natural Language Processing, 2026

Streaming Summarization Long-Context Dialogue Missing Evidence Memory Construction Retrieval

[arXiv]

DREAM pipeline showing multi-agent debate, automatic agreement labeling, and human escalation for disagreements

Core idea. Uses adversarial multi-agent debate to fill missing relevance judgments in IR benchmarks, auto-labeling agreement cases and escalating only disagreements to human annotators.

ICLR 2026 NLP

Completing Missing Annotation: Multi-Agent Debate for Accurate and Scalable Relevant Assessment for IR Benchmarks

{Minjeong Ban, Jeonghwan Choi, Hyangsuk Min}, Nicole Hee-Yeon Kim, Minseok Kim, Jae-Gil Lee, and Hwanjun Song*

International Conference on Learning Representations, 2026

Information Retrieval Missing Relevance Judgments Multi-Agent Debate Human-in-the-Loop Benchmark Refinement

[Paper]

2025
HAMLET pipeline from hierarchical key-fact queries through book-length summary generation and multi-level evaluation

Core idea. Builds a hierarchical key-fact evaluation pipeline that probes recall and faithfulness at multiple abstraction levels in book-length contexts.

EMNLP 2025 NLP

Towards a Holistic and Automated Evaluation Framework for Multi-Level Comprehension of LLMs in Book-Length Contexts

{Yuho Lee, Jiaqi Deng}, Nicole Hee-Yeon Kim, Hyangsuk Min, Taewon Yun, Minjeong Ban, Yul Kim, and Hwanjun Song*

International Conference on Empirical Methods in Natural Language Processing (Main), 2025

Long-Context Evaluation Book-Length Comprehension Key-Fact Hierarchy Query-Focused Summarization Faithfulness

[Paper]

ReFeed method using reflective reasoning on multi-dimensional feedback and SumFeed-CoT data construction

Core idea. Refines summaries across faithfulness, completeness, and conciseness by using reflective reasoning to reconcile multi-dimensional feedback and resist order and noise effects.

COLM 2025 NLP

ReFeed: Multi-Dimensional Summarization Refinement with Reflective Reasoning on Feedback

Taewon Yun, Jihwan Oh, Hyangsuk Min, Yuho Lee, Jihwan Bang, Jason Cai, and Hwanjun Song*

Conference on Language Modeling, 2025

Summary Refinement Reflective Reasoning Multi-Dimensional Feedback Long Chain-of-Thought Feedback Robustness

[Paper]

MSumBench overview across domains, English and Chinese summaries, multi-agent assistance, and three evaluation dimensions

Core idea. Benchmarks summarization across six domains and two languages with domain-specific key facts and multi-agent-assisted human judgments of faithfulness, completeness, and conciseness.

ACL 2025 NLP

Towards Multi-dimensional Evaluation of LLM Summarization across Domains and Languages

{Hyangsuk Min, Yuho Lee}, Minjeong Ban, Jiaqi Deng, Nicole Hee-Yeon Kim, Taewon Yun, Hang Su, Jason Cai, and Hwanjun Song*

Annual Meeting of the Association for Computational Linguistics (Main), 2025

Summarization Evaluation Multi-Domain Benchmark Multilingual NLP Multi-Agent Debate Human Evaluation

[Paper]

2023
TCTS architecture combining temporal convolutional pattern learning with clustering-based classification

Core idea. Jointly learns multi-scale temporal patterns with a temporal convolutional network and cluster-separable representations for unsupervised time-series segmentation.

BigComp 2023 Time-series DL

Temporal Convolutional Network-based Time-Series Segmentation

Hyangsuk Min, Jae-Gil Lee*

International Conference on Big Data and Smart Computing, 2023

Time-Series Segmentation Temporal Convolutional Network Temporal Clustering Unsupervised Learning

[Paper]

2022
COVID-EENet architecture with microscopic economy, geography, and epidemic encoders and a macroscopic aggregator

Core idea. Models how concurrent infection events affect district- and business-level sales using microscopic multi-view encoders and a macroscopic gated aggregator.

AAAI 2022 · Oral Time-series DL

Covid-EENet: Predicting Fine-Grained Impact of COVID-19 on Local Economies

Doyoung Kim, Hyangsuk Min, Youngeun Nam, Hwanjun Song, Susik Yoon, Minseok Kim, and Jae-Gil Lee*

AAAI Conference on Artificial Intelligence, 2022

Economic Impact Forecasting COVID-19 Multi-View Learning Fine-Grained Prediction Credit Card Data

[Paper]

2020
Hi-COVIDNet two-level architecture with country-level and continent-level encoders for imported case prediction

Core idea. Predicts imported COVID-19 cases with country- and continent-level encoders that combine epidemic signals with international inflow patterns.

KDD 2020 Time-series DL

Hi-COVIDNet: Deep Learning Approach to Predict Inbound COVID-19 Patients and Case Study in South Korea

Minseok Kim, Junhyeok Kang, Doyoung Kim, Hwanjun Song, Hyangsuk Min, Youngeun Nam, Dongmin Park, and Jae-Gil Lee*

International Conference on Knowledge Discovery and Data Mining, 2020

Epidemic Forecasting Hierarchical Modeling Imported Cases Transformer LSTM

[Paper]