ELION Lab
Language Intelligence & Representation
Our research group explores the frontiers of natural language processing and AI systems, guided by the vision of Elucidating Language Intelligence & RepresentatiON (ELION), to understand how language models represent knowledge, reason, and generate meaning toward interpretable and controllable language intelligence for real-world interaction.
μ°κ΅¬ λΆμΌ (Research Areas)
Toward the vision of All Languages, One Mind, we focus on the following research topics.
Language Models & World Models
We develop and analyze language models to study how scale, structure, and representations support reasoning, generalization, controllability, and world modeling.
Thinking & Reasoning
We study the internal representations and processes that enable language models to perform multi-step thinking and reasoning, aiming to interpret how conclusions are formed and fail.
Hallucination Detection & Mitigation
We investigate hallucination as a representational and epistemic failure in language models, developing methods to detect, analyze, and mitigate ungrounded or misleading generations.
Model Editing
We explore model editing as a means of modifying internal knowledge representations, enabling correction, updating, and control of language model behavior without retraining.
Multilingual & Multimodality
We study how language models represent and align meaning across languages and modalities, with the goal of building unified models that generalize beyond linguistic and modal boundaries.
AI Agents
We design and analyze language-model-based AI agents, focusing on how internal representations support planning, decision-making, and interaction in dynamic environments.
λͺ¨μ§ μ€! (NOW HIRING!)
π Ongoing Projects
ELION Labμ μΈμ΄ λͺ¨λΈμ νκ³λ₯Ό λμ΄, μΈμμ μ΄ν΄νκ³ νλνλ μ°¨μΈλ AIλ₯Ό ꡬμΆνκΈ° μν΄ λ€μμ ν΅μ¬ κ³Όμ λ€μ μννκ³ μμ΄μ.
1. Continual Representation Learning π§
μΈμ΄ λͺ¨λΈμ΄ μλ‘μ΄ μ 보λ₯Ό νμ΅ν λ κΈ°μ‘΄ μ§μμ μμ΄λ²λ¦¬λ 'μΉλͺ μ λ§κ°(Catastrophic Forgetting)' μμ΄ μ§μμ μ§μμ μΌλ‘ μ λ°μ΄νΈνκ³ μ κ΅ννλ λ°©λ²μ μ°κ΅¬νκ³ μμ΄μ. μ΄λ₯Ό Life2Vec μ€νμΌμ μ°κ΅¬λ‘ νμ₯νμ¬, AnyType λ°μ΄ν°λ₯Ό νμ©ν΄ μΈκ°κ³Ό λͺ¨λΈμ μ μμ μ£ΌκΈ°μ κ±ΈμΉ μ¬κ±΄κ³Ό 리μ€ν¬λ₯Ό μμΈ‘νλ μ§μνμ΅(Lifelong Learning) μμ€ν ꡬμΆμ λͺ©νλ‘ ν΄μ.
* Status: λ€μμ λ Όλ¬Έμ΄ κΈλ‘λ² νν°μ΄ νν μ¬μ¬ μ€ (Under Review)
2. Reducing Hallucination & RAG π‘οΈ
LLMμ΄ μλͺ»λ μ 보λ₯Ό μμ±νλ νκ°(Hallucination) νμμ νμ§νκ³ μννλ μ΅μ²¨λ¨ λ°©λ²λ‘ μ νꡬνκ³ μμ΄μ. λ΄λΆμ μΌλ‘λ Calibrationκ³Ό Uncertainty Estimationμ ν΅ν΄ λͺ¨λΈμ λ΅λ³ νμ λλ₯Ό μΈ‘μ νκ³ , μΈλΆμ μΌλ‘λ μ λ°ν κ²μ μ¦κ°(RAG) κ³Όμ μμμ Grounding κΈ°μ μ κ°λ°νμ¬ μμ± κ²°κ³Όμ μ λ’°μ±κ³Ό μΌκ΄μ±μ κ·Ήλννκ³ μμ΄μ.
* Project: IITP κ³Όμ μν μ€ (2024-2026, μμ±ν AI μ±κ³Όλ¬Όμ μ λ’°μ± λ° μΌκ΄μ± μ°κ΅¬)
3. Multi-Agent Systems & Orchestration π€
λ¨μΌ λͺ¨λΈμ νκ³λ₯Ό λμ΄, μ¬λ¬ AI μμ΄μ νΈκ° νλ ₯νλ μ€μ©μ μΈ λ©ν°μμ΄μ νΈ μμ€ν κ°λ°μ μ§ν₯νκ³ μμ΄μ. ν¨μ¨μ μΈ μ€μΌμ€νΈλ μ΄μ (Orchestration) κΈ°μ μ ν΅ν΄ Planningκ³Ό μ¨ν¨λ‘μ§ νμ κ³Όμ μ ν΅ν©νλ©°, κ²½λνλ λͺ¨λΈμμλ 볡μ‘ν νμ€ν¬λ₯Ό μνν μ μλ κ³ μ±λ₯ μμ΄μ νΈ μμ€ν μ ꡬννλ κ²μ΄ λͺ©νμμ.
* Status: μμ΄μ νΈ κ΄λ ¨ μ κ· μ°κ΅¬ κ³Όμ μ§μ λ° μΆμ§ μ€
* Collaboration: κ³ λ €λνκ΅ Agentic AI Teamκ³Ό κΈ΄λ°ν νλ ₯ μ°κ΅¬ μ§ν
4. Multimodal LM & World Understanding ποΈ
ν μ€νΈλ₯Ό λμ΄ μκ° λ± λ€μν λͺ¨λ¬λ¦¬ν°λ₯Ό ν΅ν©νμ¬ λ¬Όλ¦¬μ /μ¬νμ λ§₯λ½μ μ΄ν΄νλ μ°κ΅¬λ₯Ό μννκ³ μμ΄μ. Vision-Language λͺ¨λΈμ μμ‘΄μ± κ΅¬λ¬Έ λΆμ(Dependency Parsing)κ³Ό μλ©ν± μ²νΉ(Semantic Chunking) μ±λ₯μ λμ¬ λ©ν°λͺ¨λ¬ λ¬Έμ κ²μ λ° μ΄ν΄λλ₯Ό νμ νλ©°, AIκ° νμ€ μΈκ³μ μμμ μΆλ‘ ν μ μλ μ§λ₯μ κ°μΆλλ‘ ν΄μ.
* Collaboration: κ³ λ €λνκ΅ Document AI & Multimodal νκ³Ό 곡λ μ°κ΅¬ μν μ€
5. World-Interactive Data Augmentation π
μΉμμ λ°μ΄ν°κ° ν¬ν μνμ μ΄λ₯Έ 'λ°μ΄ν° κ³ κ°' μλμ λμνκΈ° μν΄, νμ€ μΈκ³ νκ²½κ³Όμ μνΈμμ©μ ν΅ν λ°μ΄ν° μ¦κ° κΈ°λ²μ νꡬν΄μ. νκ²½κ³Όμ νΌλλ°±μ ν΅ν΄ κ³ μ°¨μμ μΈ μΆλ‘ (High-level reasoning)κ³Ό λ€μ°¨μμ μΈμ΄ λ°μ΄ν°λ₯Ό μ€μ€λ‘ μμ±ν΄λ΄λ νμ μ μΈ λ°μ΄ν° μμ§ κΈ°μ μ μ°κ΅¬ν΄μ.
* Collaboration: μ±κ°ν¬λ₯΄ A*STAR Research λ° Microsoft Research Asia (MSRA)μ κΈλ‘λ² νμ μμ
ELION Lab is dedicated to pushing the boundaries of language models to build next-generation AI that interacts with the real world.
1. Continual Representation Learning π§
We investigate methods for language models to continuously update and refine their internal knowledge without experiencing "catastrophic forgetting." Expanding this into Life2Vec-style research, we aim to build lifelong learning systems that utilize AnyType data to predict events and risks across both human and model lifecycles.
* Status: Multiple papers currently under review at top-tier global conferences.
2. Reducing Hallucination & RAG π‘οΈ
We explore cutting-edge methodologies to detect and mitigate the hallucination phenomenon in LLMs. Internally, we focus on measuring model confidence through calibration and uncertainty estimation. Externally, we strive to maximize the reliability and consistency of generated outputs by developing advanced grounding techniques within precise Retrieval-Augmented Generation (RAG) pipelines.
* Project: Supported by IITP (2024β2026, Research on the Reliability and Coherence of Generative AI Outcomes).
3. Multi-Agent Systems & Orchestration π€
Moving beyond the limitations of single models, we aim to develop practical and precise multi-agent systems. Through efficient orchestration, we integrate planning and ontology exploration to achieve high performance even with lightweight models, enabling them to execute complex, multi-step tasks.
* Status: New research proposals in progress; actively expanding the agentic AI agenda.
* Collaboration: Joint research with the Korea University Agentic AI Team.
4. Multimodal LM & World Understanding ποΈ
We conduct research on integrating diverse modalities, such as vision, to understand physical and social contexts beyond text. By enhancing dependency parsing and semantic chunking in Vision-Language Models (VLMs), we innovate multimodal document retrieval and empower AI to reason with real-world common sense.
* Collaboration: Ongoing joint research with the Korea University Document AI & Multimodal Team.
5. World-Interactive Data Augmentation π
To address the era of "data exhaustion" as online data becomes saturated, we explore world-interactive data augmentation. We research innovative data engines that autonomously generate high-level reasoning and multi-dimensional language data through direct feedback and interaction with real-world environments.
* Collaboration: Upcoming international collaboration with Singapore A*STAR Research and Microsoft Research Asia (MSRA).
ELION Labμμ μΈκ³΅μ§λ₯ μ°κ΅¬μ μ΄μ μ μ§λ μΈν΄, μμ¬, λ°μ¬κ³Όμ νμμ λͺ¨μ§ν©λλ€
(We are now looking for talented M.S/Ph.D students, and research interns.)
μ΅μ λ΄μ€ (Latest News)
Stay updated with our recent achievements and announcements.
- Mar 2026 π₯ 2 papers are accepted at CVPR 2026.
- Mar 2026 π Established the ELION Lab at Konkuk University
- Aug 2025 π₯ 5 papers are accepted at EMNLP 2025.
- May 2025 π₯ 1 paper is accepted at ACL 2025.
- Feb 2025 π₯ 1 paper is accepted at ICLR 2025.
- Feb 2025 π₯ 2 papers are accepted at NAACL 2025.
μ£Όμ λ Όλ¬Έ (Featured Publications)
Selected recent papers from our research group.
π Evidential Transformation Network: Turning Pretrained Models into Evidential Models for Uncertainty Estimation
Conference on Computer Vision and Pattern Recognition (CVPR), 2026
π M3DocDep: Multi-modal, Multi-page, Multi-document Dependency Chunking with Large Vision-Language Models
Conference on Computer Vision and Pattern Recognition (CVPR), 2026
π The Impact of Negated Text on Hallucination with Large Language Models
Empirical Methods in Natural Language Processing (EMNLP), 2025
π KoLEG: On-the-Fly Korean Legal Knowledge Editing with Continuous Retrieval
Empirical Methods in Natural Language Processing (EMNLP) Findings, 2025
π MultiDocFusion: Hierarchical and Multimodal Chunking Pipeline for Enhanced RAG on Long Industrial Documents
Empirical Methods in Natural Language Processing (EMNLP), 2025
π Metric Calculating Benchmark: Code-Verifiable Complicate Instruction Following Benchmark for Large Language Models
Empirical Methods in Natural Language Processing (EMNLP), 2025
π K-HALU: Multiple Answer Korean Hallucination Benchmark for Large Language Models
International Conference on Learning Representations (ICLR), 2025
π Post-negation Text Induce New Hallucinations in Large Language Models
Annual Conference on Human and Cognitive Language Technology (HCLT), 2024
π KoCommonGEN v2: A Benchmark for Navigating Korean Commonsense Reasoning Challenges in Large Language Models
Annual Meeting of the Association for Computational Linguistics (ACL) Findings, 2024
π CHEF in the Language Kitchen: A Generative Data Augmentation Leveraging Korean Morpheme Ingredients
Empirical Methods in Natural Language Processing (EMNLP), 2023
π PU-GEN: Enhancing generative commonsense reasoning for language models with human-centered knowledge
Knowledge-Based Systems, 2022
π Plain Template Insertion: Korean-Prompt-Based Engineering for Few-Shot Learners
IEEE Access, 2022
π A Dog Is Passing Over The Jet? A Text-Generation Dataset for Korean Commonsense Reasoning and Evaluation
North American Chapter of the ACL (NAACL) Findings, 2022
π Dense-to-Question and Sparse-to-Answer: Hybrid Retriever System for Industrial Frequently Asked Questions
Mathematics, 2022
π KommonGen: A Dataset for Korean Generative Commonsense Reasoning Evaluation
Annual Conference on Human and Cognitive Language Technology (HCLT), 2021