Jiwon Kim

Visiting Researcher · Geometry for AI, AI for Math

Max Planck Institute for Security & Privacy (MPI-SP), Bochum · 2026

[email protected]·GitHub

Education

2016 – 2025

Ph.D. in Mathematics

Seoul National University

  • Advisor: Prof. Woong Kook
  • Dissertation: An Orbifold Framework for Point Scan
  • ~3 years of dedicated doctoral research; 6 concurrent years in industry R&D under Korea's alternative military service and a subsequent CTO role. Academic publication during the industry period was restricted by NDA; research output appears as 3 US patents.
2014 – 2016

M.S. in Mathematics

Seoul National University

  • Advisor: Prof. Cheolhyun Cho
  • Thesis: Survey of Quantum Entanglement and Geometry
2010 – 2014

B.S. in Mathematical Sciences

KAIST

Appointments

2026 – present

Visiting Researcher

Max Planck Institute for Security and Privacy (MPI-SP), Bochum

  • Scientific Host: Prof. Meeyoung (Mia) Cha
  • Collaborating with Dr. Jae Kwon and Dr. Dongkyum Kim on AI interpretability
  • Research focus: Mathematical formulation for memory localization in AI systems
  • Invited for joint research and fully funded by MPI-SP

Publications

J. Kwon*, D. Kim*, J. Kim*, Y. Kim, W. Kook, and M. Cha. “AI Engram: In Search of Memory Traces in Artificial Intelligence.” Oral, International Conference on Machine Learning (ICML), 2026.

*Equal contribution (Kwon, D. Kim, J. Kim)  |  Corresponding author: M. Cha (MPI-SP)

OpenReview

J. Kwon, J. Kim, D. Kim, and M. Cha. “Moir: Let the Model Direct Its Own Story for Robust Cross-Domain Knowledge Editing.” Under review at the Conference on Neural Information Processing Systems (NeurIPS), 2026.

Corresponding author: M. Cha (MPI-SP)

J. Kim, K. Ahn, and Y. Kim. “Categorical Abstract Reasoner.” In preparation, 2026.

Leading first author. Category-theoretic foundations for abstract reasoning, with mechanized proofs in Lean.

J. Kim, et al. “A Geometric Framework for Unlearning and Editing.” In preparation, 2026.

Engram-series follow-up: differential-geometric foundations unifying memory localization, knowledge editing, and machine unlearning.

Awards & Honors

2026

ICML 2026 Oral Presentation

For AI Engram: In Search of Memory Traces in Artificial Intelligence — 168 Orals selected from 23,918 ICML 2026 submissions (top 0.7%).

2026

MPI-SP Visiting Researcher Fellowship

Fully funded visiting position invited by Prof. Meeyoung Cha, Max Planck Institute for Security & Privacy.

2011 – 2013

President, Mathematical Sciences Student Council

KAIST. Elected by peers in the Department of Mathematical Sciences.

Invited Talks

Feb 2026

Memory Engram in AI

Prof. Krishna Gummadi's Group, Max Planck Institute for Software Systems (MPI-SWS), Kaiserslautern, Germany.

Patents

US 11,734,807

Core inventor. Modeled confocal microscopy signal space as an orbifold (T²/V₄) for real-time Lissajous phase correction, reducing calibration latency from 300 ms (2019 SOTA) to under 10 ms.

US 11,270,418

Core inventor. Developed gauge-theoretic calibration pipeline for reconstructing global manifolds from local probe scans with sub-micron precision.

US 12,387,292

Core inventor. Derived optimal frequency selection rule using ergodic theory to maximize fill factor in scanning acquisition.

Research Interests

Mechanistic interpretability of memory and editing

Memory localization, feature attribution, and surgical knowledge editing in deep neural networks. Differential geometry of parameter space (Fisher information, natural gradient) and covariance-based editors in the ROME / MEMIT / AlphaEdit lineage. Featured in AI Engram (ICML 2026 Oral) and Moir (NeurIPS 2026 submission).

Categorical foundations for abstract reasoning

Sheaf-theoretic and functorial approaches to neural representation and abstract reasoning. Current focus: category-theoretic foundations for reasoning agents with mechanized proofs in Lean (Categorical Abstract Reasoner, in preparation, leading first author).

Orbifold structures and Lie-theoretic optimization

Foundational geometric methods for imaging, calibration, and equivariant representation. Orbifold quotients (T²/V₄), Riemannian optimization on Lie groups (SO(4), SE(3)), and gauge-symmetric calibration. Doctoral framework: An Orbifold Framework for Point Scan; industrially validated through 3 US patents on confocal microscopy and Lissajous phase correction.

Industry Experience

Note: “Military Service” entries refer to Korea's alternative service program for technical research specialists.

2022 – 2024

CTO, Artiwealth

Led engineering team; built convex tax optimization engine with 10x convergence improvement.

2020 – 2022

Researcher, Tibero

Military Service

Designed computational geometry kernel for spatial database indexing with exact geometric computation.

2019 – 2020

Researcher, VPIX Medical

Military Service

Core inventor on 3 US patents; developed real-time geometric signal processing for confocal microscopy (~30× latency reduction: 300 ms → under 10 ms vs. 2019 SOTA Lissajous scanning baseline).

Teaching

2017 – 2018

Head Teaching Assistant, Natural Sciences

Seoul National University. Oversaw a team of TAs; coordinated curriculum delivery and standardized grading.

2015 – 2016

Teaching Assistant, College Mathematics I & II

Seoul National University (4 semesters).