Inner Workings of Language Models
Understanding how language models represent information, organize internal computation, and change their mechanisms through training, post-training, and adaptation.
Diffusion · Language Models · Inner Workings
Injin Kong
Rethinking language generation through diffusion and inner mechanisms.
I am a master's student in Data Science at Seoul National University and a member of HOLI LAB. I received my B.S. in Mathematics from Seoul National University. My research focuses on diffusion-based language generation, the internal mechanisms of language models, and reliable AI applications.
Current Affiliation
HOLI LAB
/ Seoul National University
About
Understanding AI systems from the inside out.
I am a master's student in Data Science at Seoul National University and a member of HOLI LAB. I received my B.S. in Mathematics from Seoul National University.
My research studies the internal mechanisms of language models and alternative approaches to generation, especially diffusion-based language modeling. I am interested in how models represent information, how their behavior changes through training and post-training, and how these insights can lead to more reliable AI systems.
Research Interests
Language models, inner workings, and diffusion-based generation.
Diffusion-based Language Generation
Exploring diffusion-based generative paradigms for language, including masked diffusion models, decoding dynamics, and hidden-state diffusion interfaces.
AI Agents and Applications
Building reliable language-model-based systems that connect research ideas to real-world workflows, educational tools, and agentic applications.
Selected Papers
Research output and manuscripts.
Where Should Diffusion Enter a Language Model? Geometry-Guided Hidden-State Replacement
Injin Kong, Hyoungjoon Lee, Yohan Jo · ICML FoGen Workshop 2026
Studies where diffusion mechanisms should enter language models through geometry-guided hidden-state replacement.
Mechanism Shift During Post-training from Autoregressive to Masked Diffusion Language Models
Injin Kong*, Hyoungjoon Lee*, Yohan Jo · EMNLP 2026
Studies how model mechanisms shift during post-training from autoregressive language models to masked diffusion language models.
Dual Mechanisms of Value Expression: Intrinsic vs. Prompted Values in Large Language Models
Jongwook Han*, Jongwon Lim*, Injin Kong, Yohan Jo · ICML 2026
Investigates how large language models express values through intrinsic behavior and prompt-induced responses.
Show more papers
Can MLLMs Reason About Visual Persuasion? Evaluating the Efficacy and Faithfulness of Reasoning
Naeun Lee, Hyunjong Kim, Sunghwan Choi, Injin Kong, Yohan Jo · ECCV Visper Workshop 2026 oral presentation
Evaluates whether multimodal large language models can reason effectively and faithfully about visual persuasion.
Style Extraction on Text Embeddings Using VAE and Parallel Dataset
Injin Kong, Shinyee Kang, Yuna Park, Sooyong Kim, Sanghyun Park · arXiv preprint · 2024
Proposes a VAE-based approach for extracting style information from text embeddings using a parallel dataset.
CV
Education, awards, and experience.
Education
2025 - Present
Seoul National University
M.S. in Data Science
2020 - 2025
Seoul National University
B.S. in Mathematics
2019
Yonsei University
Withdrew to enroll in Seoul National University
2016 - 2018
Hana Academy Seoul
High School
Awards
-
2023
Second Place
CJ Logistics Future Technology Challenge · Image-Based Volume Estimation of Parcels
-
2019
Academic Excellence Award
Yonsei University · Highest Academic Achievement
-
2018
Gold Prize
Korean Mathematical Olympiad (KMO)
-
2018
Grand Prize (1st Place)
Earth Science Competition · Hana Academy Seoul
-
2018
Gold Prize (2nd Place)
Mathematics Research Presentation Contest · Hana Academy Seoul
Experience
2023.11 - 2025.08
Chief Science Officer
Aardvark
2019
Proofreading Committee Member
Black Label Geometry
Contact
Feel free to contact me.
For research discussions, collaboration, or opportunities, feel free to reach out by email.