Chanwoong Yoon

Integrated MS/PhD at Korea University under the supervision of Professor Jaewoo Kang.

I'm deeply interested in Natural Language Processing (NLP) and Large Language Model (LLM).

Email / CV / Google Scholar / LinkedIn / X / Github

Interpretability: Recently, I've explored tools (e.g., SAE, circuit analysis) to gain insights into the internal mechanisms of LLMs. Beyond simply controlling these models without understanding their inner workings, I aim to comprehend how LLMs make decisions and leverage pretrained world knowledge to perform complex tasks. Also, I'm interested in verifying, editing, and controlling the knowledge and behavior of LLMs. (ChroKnowledge)
Retrieval-Augmented Generation: I’ve tackled RAG bottlenecks, focusing on improving retriever and reader components. For example, RetPO optimizes a query reformulation model to better align with retriever preferences, enhancing off-the-shelf retrievers in conversational search, while CompAct actively compresses retrieved documents to help reader models to handle long contexts more efficiently.

10.2024. propose a long-context benchmark to assess LLMs’ ability to leverage the entire context fully: ETHIC.
09.2024. Unveiling chronological knowledge in 5 domains released: ChroKnowledge.
09.2024. Novel framework for context compression paper accepted to @ EMNLP'24: CompAct.
07.2024. Language Model-Augmented Police Investigation System paper accepted to @ CIKM'24: LAPIS.
02.2024. present RetPO (Retriever’s Preference Optimization) to optimize a LM for reformulating search queries in line with the preferences of the target retrieval systems: RetPO.

Selected Publications

Question Answering Retrieval-Augmented Generation Prompt Compression
CompAct: Compressing Retrieved Documents Actively for Question Answering
Chanwoong Yoon, Taewhoo Lee, Hyeon Hwang, Minbyul Jeong, Jaewoo Kang
EMNLP 2024 Main.

arXiv / code / Google Scholar

We propose CompAct (Compressing Retrieved Documents Actively for Question Answering), a novel framework that employs an active strategy for compressing extensive documents. CompAct dynamically preserves query-related contexts, focusing on the integration of information across documents.

Information Retrieval Conversational Question Answering Retrieval-Augmented Generation
Ask Optimal Questions: Aligning Large Language Models with Retriever's Preference in Conversational Search
Chanwoong Yoon*, Gangwoo Kim*, Byeongguk Jeon, Sungdong Kim, Yohan Jo, Jaewoo Kang
Preprint.

arXiv / Google Scholar

we present RetPO (Retriever’s Preference Optimization), which is designed to optimize a language model (LM) for reformulating search queries in line with the preferences of the target retrieval systems. The process begins by prompting a large LM to produce various potential rewrites and then collects retrieval performance for these rewrites as the retrievers’ preferences.

Template based on Jon Barron's website.

Research Interests

News

Selected Publications