Seungpil Lee

Gwangju Institute of Science and Technology (GIST)

Program: Master's in Artificial Intelligence

Email | Google Scholar | CV

"From humans to AI, and from AI back to humans."

That is my research motto. My long-standing question has been how an intelligent system works, so rather than simply raising performance, I do research to understand how intelligence works at its core. Guided by this philosophy, my earlier work modeled human intelligence and used what I learned along the way to improve AI. Today, building on that philosophy and under the guidance of Prof. Sundong Kim, I mainly work on ARC-AGI, a visual puzzle benchmark in which the task is to find a rule from a few examples.

I am open to every kind of research on the nature of intelligence, not only AI but also STS, literature, and philosophy (mostly philosophy of mind). Even so, the area that interests me most right now is metacognition. From my research so far, I have come to think that, unlike people, language models lack metacognition, the ability to quickly notice problems in their own thinking. Going forward, I want to test this hypothesis by building metacognition into models. Further out, I want to extend the question of how intelligence works to the problems of consciousness and the self. If you are curious how my philosophy took shape, what my earlier work covered, and what I want to work on next, my Research Philosophy page lays it out at more length. And away from research, I am also a passionate fiction writer; some of my stories live under Writing.

Education

Gwangju Institute of Science and Technology (GIST)

March 2025 - Present

• Master's Degree in AI Convergence
• Research focus: Cognition of Intelligent Systems via LLMs, Mechanistic Interpretability, Memory-Augmented Agents
• Advisor: Prof. Sundong Kim

Gwangju Institute of Science and Technology (GIST)

March 2018 - December 2024

• Bachelor's Degree in Electrical Engineering and Computer Science
• Minor in Mathematics
• Minor in Literature

Research Experience

LG AI Research, Seoul

September 2026 - Present

Research Intern

• Improving the coding capability of K-Exaone, a Korean sovereign foundation model

Microsoft Research Asia (MSRA), Shanghai

February 2026 - August 2026

Research Intern, Collaborative Research Program

• Developed test-time skill optimization for LLM agents that acquire and reuse task skills at inference time without parameter retraining
• Extended the Agent-Lightning framework with memory generation so agents carry learned context across long-horizon tasks

Data Science and Artificial Intelligence Lab, GIST

September 2023 - Present

Graduate Research Assistant, Supervised by Prof. Sundong Kim

• Working on Abstraction and Reasoning Corpus (ARC) challenge
• Developing novel approaches for visual reasoning and abstract problem solving
• Studying the cognition of intelligent systems through LLMs: failure modes, mechanistic interpretability, and memory-augmented agents that learn at test time

Publications

* denotes equal contribution

Journal Articles

Solution Augmentation for ARC-AGI Problems Using GFlowNet: A Probabilistic Exploration Approach

Sanha Hwang, Seungpil Lee, Sejin Kim, and Sundong Kim†

Transactions on Machine Learning Research, 2025

PAPER

Reasoning Abilities of Large Language Models: In-Depth Analysis on the Abstraction and Reasoning Corpus

Seungpil Lee*, Woochang Sim*, Donghyeon Shin*, Wongyu Seo, Jiwon Park, Seokki Lee, Sanha Hwang, Sejin Kim and Sundong Kim†

ACM Transactions on Intelligent Systems and Technology, vol. 16, no. 6, 2025, 2025

CODE PAPER SITE

@article{LoTH_TIST2025,
  title = {{Reasoning Abilities of Large Language Models: In-Depth Analysis on the Abstraction and Reasoning Corpus}},
  author = {Lee, Seungpil and Sim, Woochang and Shin, Donghyeon and Seo, Wongyu and Park, Jiwon and Lee, Seokki and Hwang, Sanha and Kim, Sejin and Kim, Sundong},
  journal = {ACM Transactions on Intelligent Systems and Technology},
  volume = {16},
  number = {6},
  pages = {1--52},
  year = {2025},
  doi = {10.1145/3712701},
}

Conference Papers

From Noise to Diversity: Random Embedding Injection in LLM Reasoning

Heejun Kim*, Seungpil Lee*, Jewon Yeom, Jaewon Sok, Seonghyeon Park, Jeongjae Park, Taesup Kim†, Sundong Kim†

In NeurIPS (earlier version at ICML Mechanistic Interpretability Workshop 2026, Spotlight), 2026

CODE PAPER

@inproceedings{RSP_NeurIPS2026,
  title = {{From Noise to Diversity: Random Embedding Injection in LLM Reasoning}},
  author = {Kim, Heejun and Lee, Seungpil and Yeom, Jewon and Sok, Jaewon and Park, Seonghyeon and Park, Jeongjae and Kim, Taesup and Kim, Sundong},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2026},
}

Can Large Language Models Develop Gambling Addiction?

Seungpil Lee*, Donghyeon Shin, Yunjeong Lee and Sundong Kim†

In NeurIPS, 2026

PAPER

@inproceedings{GamblingAddiction_NeurIPS2026,
  title = {{Can Large Language Models Develop Gambling Addiction?}},
  author = {Lee, Seungpil and Shin, Donghyeon and Lee, Yunjeong and Kim, Sundong},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year = {2026},
}

From Generation to Selection: Findings of Converting Analogical Problem-Solving into Multiple-Choice Questions

Donghyeon Shin*, Seungpil Lee*, Klea Lena Kovacec, and Sundong Kim†

In EMNLP Findings, 2024

CODE PAPER SITE

@article{Revisit_KAIS19,
  author = {Kim, Sundong and Lee, Jae-Gil},
  title = {{From Generation to Selection: Findings of Converting Analogical Problem-Solving into Multiple-Choice Questions}},
  author = {Shin, Donghyeon and Lee, Seungpil and Kovacec, Klea Lena and Kim, Sundong},
  booktitle = {EMNLP Findings},
  year = {2024},
  }

ARCLE: The Abstract and Reasoning Corpus Learning Environment for Reinforcement Learning

Hosung Lee*, Sejin Kim*, Seungpil Lee, Sanha Hwang, Jihwan Lee, Byung-Jun Lee†, and Sundong Kim†

In CoLLAs, 2024

CODE PAPER SITE

@article{arcle2024lee,
  author = {Hosung Lee, Sejin Kim, Seungpil Lee, Sanha Hwang, Jihwan Lee, Byung-Jun Lee, and Sundong Kim},
  title = {{ARCLE: The Abstraction and Reasoning Corpus Learning Environment for Reinforcement Learning}},
  author = {Lee, Hosung and Kim, Sejin and Lee, Seungpil and Hwang, Sanha and Lee, Jihwan and Lee, Byung-Jun and Kim, Sundong},
  booktitle = {CoLLAs},
  year = {2024}
}

Workshop Papers

Reasoning Abilities of Large Language Models through the Lens of Abstraction and Reasoning (Extended Abstract of the below paper)

Seungpil Lee*, Woochang Sim*, Donghyeon Shin*, Sejin Kim and Sundong Kim†

In NeurIPS Workshop on System-2 Reasoning at Scale, 2025

CODE PAPER SITE

@article{LoTHneurips2024Lee,
  title = {{Reasoning Abilities of Large Language Models through the Lens of Abstraction and Reasoning (Extended Abstract of the below paper)}},
  author = {Lee, Seungpil and Sim, Woochang and Shin, Donghyeon and Kim, Sejin and Kim, Sundong},
  booktitle = {NeurIPS Workshop on System-2 Reasoning at Scale},
  year = {2024},
  }

Regulation Using Large Language Models to Generate Synthetic Data for Evaluating Analogical Ability

Donghyeon Shin*, Seungpil Lee*, Klea Lena Kovacec, and Sundong Kim†

In IJCAI Workshop on Analogical Abstraction in Cognition, Perception, and Language, 2024

CODE PAPER SITE

@article{regulation2024shin,
  author = {Donghyeon Shin, Seungpil Lee, Klea Lena Kovacec, and Sundong Kim},
  title = {{Regulation Using Large Language Models to Generate Synthetic Data for Evaluating Analogical Ability}},
  author = {Shin, Donghyeon and Lee, Seungpil and Kovacec, Klea Lena and Kim, Sundong},
  booktitle = {IJCAI Workshop on Analogical Abstraction in Cognition, Perception, and Language},
  year = {2024},
  }

Domestic Conference Papers

Extracting the core knowledge of ARC with the World Model

KSC, 2023

PAPER

Evaluating Prior Knowledge of ARC Using World Models

KSC, 2023

PAPER

Awards & Honors

National Research Foundation of Korea Funding, NRF

September 2025 - September 2026

• Funding offered to outstanding master's students for conducting research projects

Korean Government Scholarships, GIST College

March 2018 - August 2024

• Scholarship awarded to students studying in GIST

Scholarship for Summer Session Abroad

June 2019 - August 2019

• Scholarship awarded to students studying abroad during a summer session

Navy AI Competition

June 2023 - August 2023

• Awarded drone object detecting competition hosted by Korean Navy

K-StartUp finals

March 2023 - November 2023

• Advanced to the finals of Korean largest start-up contest, 'K-StartUp', as ML researcher and back-end developer

Media Coverage

Can AI become a tool for safer gambling?

SiGMA News, 2026

ARTICLE

AI Bots Show Signs of Gambling Addiction, Study Finds

Newsweek, 2026

ARTICLE

AI can show gambling addiction-like behaviour under risk: Expert

SiGMA News, 2026

ARTICLE

New Data Augmentation Method Improves ARC Accuracy by 5.85%p

Dong-A Science, 2025

ARTICLE

Quantitative Evaluation Framework for LLM Reasoning Abilities

News1, 2025

ARTICLE