About
I'm an MS student at KAIST, currently working on AI systems. Before that, I spent three years as a Research Scientist at T-Tech, first on in-context RL and then on LLM reasoning, alignment and interpretability. There I first-authored Headless-AD (ICML 2024) and showed that a single steering vector per layer can match full RL fine-tuning for reasoning (EMNLP 2025). I got my Bachelor's in Computer Science at Innopolis University, where I did my first research as an intern in Prof. Adil Khan's group.
Papers & projects
-
Engram Training with Offloaded TablesOffloading Engram's memory tables to CPU RAM during training with zero step-time overhead, freeing 6.4 GiB of GPU memory.
-
A reproduction of Anthropic's “When Models Manipulate Manifolds: The Geometry of a Counting Task” on open-source models, with an interactive explorer.
-
RL-trained steering vectors suppress language mixing, act largely as first-token substitution, work through MLP rather than attention, transfer across models and compose across layers.
-
One RL-trained steering vector per layer, with all weights frozen, matches fully RL-tuned reasoning models on math benchmarks using ~0.0016% extra parameters on an 8B model.
-
Headless-AD: a generalization of Algorithm Distillation that, trained once, adapts in context to discrete action spaces of varying size, meaning and order.
Full and up-to-date list on Google Scholar.
Experience & Education
- KAIST2026 – present MS student
- T-Tech2023 – 2026 Research Scientist · in-context RL, then the Alignment team; reasoning and interpretability
- Innopolis University2020 – 2024 B.Sc. (Hons) in Computer Science
- Innopolis University2022 – 2023 Research Intern with Prof. Adil Khan · unsupervised domain adaptation
- Oil and Gas Technology Center2022 – 2024 Junior ML Engineer, then Consultant · rock classification, pipeline defect detection
Teaching & talks
- Central University Lecture on online methods in LLM alignment
- Innopolis University Taught algorithms to first-year students
- Hackathon mentor Mentored teams of informatics teachers