Tiberiu Mușat

Tiberiu Mușat

AI Researcher, Software Engineer, EdTech Enthusiast

Hello! I’m Tiberiu, a second-year master’s student at ETH Zurich. My research is about understanding the inner workings of deep learning models, with a focus on mechanistic interpretability and training dynamics. I want to understand what models learn and how they learn it. I believe this work can lead to more efficient and reliable AI systems. In the long term, I believe it could also help us understand the nature of intelligence, reasoning, and consciousness. In my free time, I develop BacPlus.ro, a website about Romanian schools.

Google ScholarSemantic Scholar

Publications

On the Emergence of Induction Heads for In-Context Learning

On the Emergence of Induction Heads for In‑Context Learning

Preprint (2025)

Tiberiu Musat, Tiago Pimentel, Lorenzo Noci, Alessandro Stolfo, Mrinmaya Sachan, Thomas Hofmann

arxiv
The Geometry of Grokking: Norm Minimization on the Zero-Loss Manifold

The Geometry of Grokking: Norm Minimization on the Zero‑Loss Manifold

Preprint (2025)

Tiberiu Musat

arxiv
Mechanism and Emergence of Stacked Attention Heads in Multi-Layer Transformers

Mechanism and Emergence of Stacked Attention Heads in Multi‑Layer Transformers

ICLR 2025

Tiberiu Musat

arxivpostersummary
Clustering and Alignment: Understanding the Training Dynamics in Modular Addition

Clustering and Alignment: Understanding the Training Dynamics in Modular Addition

Interpretable AI Workshop, NeurIPS 2024

Tiberiu Musat

arxivdemo
© 2026 Tiberiu Musat