Interests
My goal is to realize Human-Level AI, which is defined as AGI capable of performing all tasks that humans can do. I consider it promising to apply the computational basis of the human mind to AI, thereby endowing cognitive functions. Under the hypothesis that explicit memory is the key to higher-order cognitive functions such as continual learning and metacognition, I am currently focusing on the necessary components to build it in LLMs.
Education
University of Texas at Austin 2025 - Present
Ph.D. in Computer Science Austin, Texas, USA
Expected Graduation: May 2030
Advisor: Professor Risto Miikkulainen
Sungkyunkwan University 2017 - 2024
Bachelor of Science in Department of Computer Science and Engineering Seoul, Korea
Bachelor of Arts in Department of Psychology
GPA: 4.24/4.5 | Magna Cum Laude | Transcript
Publications
-
Sangjun Park. (2026). Position: Hippocampal Explicit Memory Is the Cornerstone for AGI. Proceedings of the 43rd International Conference on Machine Learning (ICML Position Track). | ArXiv
-
Minsoo Khang, Sangjun Park, Teakgyu Hong, and Dawoon Jung. (2025). CReSt: A Comprehensive Benchmark for Retrieval-Augmented Generation with Complex Reasoning over Structured Documents. arXiv preprint.
-
Yeonji Lee, Sangjun Park, Kyunghyun Cho, and JinYeong Bak. (2024). MentalAgora: A Gateway to Advanced Personalized Care in Mental Health through Multi-Agent Debating and Attribute Control. arXiv preprint.
-
Sangjun Park and JinYeong Bak. (2024). Memoria: Resolving Fateful Forgetting Problem through Human-Inspired Memory Architecture. Proceedings of the 41st International Conference on Machine Learning.
(ICML 2024 Spotlight: Acceptance Rate=3.5%) | ArXiv | OpenReview | Github
Awards & Honors
Wesley W. Calhoun Jr. Endowed Scholarship 2025, 2026
SKKU OpenSource SW Activity Top Award 2023, 2024
NAVER Representative Award 2021
Student's Success Scholarship 2019
TOPCIT Army Chief Staff Award 2019
2018 SKKU BugBounty Incentive Award 2018
2017 SKKU BugBounty Special Award 2017
SungKyun Software Scholarship 2017 - 2019, 2022, 2023
Dean's List 2017 - 2019
Work Experience
Cognizant AI Lab May 2026 - Aug 2026
AI Research Intern San Francisco, CA, USA
- Investigated whether LLMs internally form conjunctive, concept-level representations analogous to human explicit memory.
- Designed a causal attention-control paradigm on a paired-entity dataset, masking which tokens could attend to a knowledge block and greedily eliminating layers and attention heads to isolate the minimal circuit required to read out the correct entity.
- Found that entities are read out through a compact, stable circuit hubbed on a few dominant attention heads at a consistent depth late in the network rather than smeared across it.
Upstage May 2024 - May 2025
AI Research Engineer Suwon, Korea
- Participated in the supervised fine-tuning stage of the general-purpose LLM, Solar Pro 22B.
- Researched and developed methods such as multi-task pretraining to enhance key information extraction (KIE) performance, and built a KIE pipeline leveraging large language models.
- Investigated retrieval-augmented generation (RAG) QA techniques using graph-based evidence and contributed to establishing internal benchmarks for both RAG and KIE systems.
Scatterlab Feb 2022 - Aug 2022
Machine Learning Researcher Seoul, Korea
- Designed a way of memory utilization process such as extracting, memorizing, retrieving, and using the previous conversation context as a form of memory.
- Developed a memory reminder model to retrieve relevant information from memory pool based on the current context.
- Conducted pretraining BART to serve as the foundational model for memory extraction.
- Developed sketch image classifier and stroke sequence generation model based on Transformer architecture so that the stroke sequence can be converted into the sketch image.
Scatterlab Feb 2020 - Feb 2022
Machine Learning Engineer Seoul, Korea
- Played a key role in diverse projects involving the development of an open-domain conversational chatbot agent, Luda-Lee. The tasks included enhancing the data processing pipeline, implementing an automatic finetuning & evaluation system, conducting model distillation, and so on.
- Developed a person name recognition model to de-identify person names and designed a sophisticated de-identification algorithm.
- Built vector-similarity search engine using faiss and developed a response retrieval pipeline for the chatbot.
- Optimized BERT-based classification models through the application of knowledge distillation method to reduce model inference time while preserving the original performance.
- Constructed automated pipelines with Kubeflow to finetune and evaluate language models in order to get a trained model and result metrics easily.
Research Experience
Neural Networks Research Group Oct 2025 - Present
Graduate Researcher (Advisor: Professor Risto Miikkulainen) Austin, TX, USA
- Developed ESMA (Evolution Strategy for Metacognitive Alignment), a method that trains LLMs to know what they know by optimizing metacognition, which lacks a differentiable objective, with evolution strategies rather than gradient descent.
- Adapted d’type2, a psychometric index of metacognitive sensitivity from signal detection theory, to measure whether a model’s expressed meta answer predicts its actual correctness.
- Demonstrated that this metacognitive gain is broadly consistent, holding across multiple QA datasets, across languages, and in isolated out-of-distribution settings.
Visual Cognitive Neuroscience Laboratory Feb 2024 - Dec 2024
External Research Assistant (Advisor: Professor Min-Suk Kang) Suwon, Korea
- Analyzed the formation of event boundaries when subjects perceive narratives using a language model and recall text.
- Compared Memoria’s retrieval pattern to human behavior and attempted to determine event boundaries in the Sherlock drama.
- Researched an episodic memory system for deep neural networks, Memoria, based on theories of human memory formation and forgetting, with a particular focus on Hebbian theory.
Human Language Intelligence Lab Oct 2022 - Feb 2024
Undergraduate Researcher (Advisor: Professor JinYeong Bak) Suwon, Korea
- Researched an episodic memory system for deep neural networks, Memoria, which is based on the theories of human memory formulation and forgetting with a particular focus on Hebbian theory.
- Conducted extensive experiments to prove the effectiveness of Memoria in enhancing long-term dependency consideration, applying it to Transformer-based models such as BERT and GPT.
- Participate in research developing a mental disorder detection model with a large language model and play a role in providing advice for research orientation and plan.
UCI Center for Embedded Computer Systems Jul 2018 - Aug 2018
Undergraduate Researcher (Advisor: Professor Elaheh Bozorgzadeh) Irvine, California, USA
- Participated in a research project about locating robots within a non-GPS environment.
- Simulated a single robot localization with the dead-reckoning method in V-REP and Python and multi-robot localization with UKF filter Matlab.
Activities
AI + Human Writing Competition Jan 2023
AI Model Developer Suwon, Korea
- Participated in a writing competition where human writes an essay about a given topic using their essay generation models. [Github]
- Resolved long-term dependency problem to generate text over 20,000 characters by employing a summarization model to abstract the previous context and received good comments in technical aspects.
2021 Korean Voice and Natural Language AI competition Sep 2021 - Nov 2021
Team Leader Seoul, Korea
- Won the 1st prize in a dialogue summarization competition with a total prize of $27K. [Github]
- Researched and developed dialogue summarization model applying SOTA methods and original ideas.
SW Maestro: Legato May 2019 - Nov 2019
Team Leader, Developer, and ML Engineer Seoul, Korea
- Led a chatbot development government-funded team project aimed at improving mental health.
- Researched and developed multiple natural language processing models such as sentence similarity model or conditional response generation model, etc.
- Released StoryForest to Google Play Store, the application was downloaded over 1000 times.
2018 SKKU BugBounty Oct 2018
Participant Suwon, Korea
- Won the incentive award to find vulnerabilities such as cookie reuse and SQL injection.
Hacker's In inTrusion Dec 2017 - Jan 2019
President Suwon, Korea
- Operated university information security club and taught club members basic hacking and security.
2017 SKKU BugBounty May 2017
Participant Suwon, Korea
- Won the special award to find web vulnerabilities such as cross site scripting and web shell uploading.
Teaching Experience
Mentoring Foreign Students in Major Classes Sep 2022 - Dec 2022
Mentor Suwon, Korea
- Taught foreign students about the data structure course and provided answers to their questions.
Arduino Mentoring Dec 2019
Instructor Suwon, Korea
- Taught high school students Arduino (making RC car with Bluetooth, ultrasonic, and other modules).
Computational Thinking and Software Coding Mar 2019 - Jun 2019
Teaching Assistant Seoul, Korea
- Conducted practical lessons on Entry and basic Python syntax for students.
Freshman Python Education Feb 2019 & Feb 2018
Teaching Assistant Suwon, Korea
- Taught how to solve basic algorithm problems with Python to freshmen.
Competences
Languages Korean (native), English (fluent)
Techniques Python, C/C++, Go, Java, Javascript, Arduino, Visual Studio Code, Tensorflow, Pytorch, git, Docker, Flask, Kubernetes, Kubeflow, Faiss, AWS, GCP, LaTeX
Interests Writing, Calligraphy, Playing Piano, Board Game, Reading Books, Magic