Learning dynamics and LoRA fine-tuning
I study how pre-training shapes downstream adaptation, and why stronger representations can sometimes delay or hurt Low-Rank Adaptation.
Yale · 2026
ML theory · Statistical physics
Gibbs Nwemadji
PhD student at SISSA, currently visiting Yale. I work on learning dynamics, scaling laws, superposition, LoRA fine-tuning, and teacher-student models.
Scaling laws
LoRA fine-tuning
Superposition
Teacher-student models
Quantum ML
Research
I use theoretical models and experiments to understand why neural networks learn, specialize, adapt, and scale.
Scaling laws and superposition
I am interested in whether scaling laws can emerge from superposition, and how feature geometry changes the observed decay of error.
Teacher-student models
I use teacher-student settings to isolate mechanisms of specialization, approximation, representation learning, and generalization.
Quantum machine learning
My earlier work connects quantum information, quantum algorithms, and learning theory.
Preprints and Publications
Recent work around learning theory, neural networks, statistical physics, and quantum learning.
2026
When pre-training hurts LoRA fine-tuning: a dynamical analysis via single-index models
arXiv
2026
Generalization performance of narrow one-hidden layer networks in the teacher-student setting
2024
On the Capacity of a Quantum Perceptron for storing Biased Patterns
Talks, Posters, Schools & Conferences
A timeline of recent and past scientific activities, including invited talks, poster presentations, schools, workshops, and conference participation.
Aug 31 – Sept 5, 2026
ROCKIN' AI 3.0
Roccella Ionica, Italy.
Talk
Jul 20 – 23, 2026
Youth in High-Dimensions 2026
ICTP, Trieste.
Selected participant
Late May 2026
Invited talk at Boston College
Talk on learning dynamics, scaling laws, and related questions.
Invited talk
Sept 1 – 5, 2025
ROCKIN' AI 2025
Roccella Ionica, Italy.
Poster / participation
Jul 3 – 10, 2025
Youth in High-Dimensions & How Creative is Generative AI?
Recent Progress in Machine Learning, High-Dimensional Statistics and Inference; Perspectives from Science and Philosophy. ICTP, Trieste.
Poster / participation
Jan 27 – Feb 7, 2025
Towards a Theory for Typical-Case Algorithmic Hardness
Les Houches, France.
School
Sept 2 – 6, 2024
ROCKIN' AI 2024
Roccella Ionica, Italy.
Poster / participation
May 20 – 24, 2024
Youth in High-Dimensions
Recent Progress in Machine Learning, High-Dimensional Statistics and Inference. ICTP, Trieste.
Poster / participation
Mar 24 – 29, 2024
Complex Systems, Statistical Mechanics and Machine Learning Crossover
Les Houches, France.
Workshop
Jul 3 – 7, 2023
Learning and Inference from Structured Data
Universality, correlations and beyond — ICTP, Trieste.
Workshop
May 21 – Jun 2, 2023
Youth in High-Dimensions
Recent Progress in Machine Learning, High-Dimensional Statistics and Inference. ICTP, Trieste.
Poster / participation
Nov 7 – 11, 2022
Ethical and Societal Challenge of Machine Learning
ICTP, Trieste.
Workshop
Feb 21 – Mar 15, 2022
Spring College on Physics of Complex Systems
SISSA–ICTP, Trieste.
School
Path
From physics and data science to theoretical machine learning.
2026
Yale University — New Haven, CT
Visiting PhD student working with Prof. John Sous on superposition, scaling laws, and representations induced by LoRA fine-tuning.
2025
École Normale Supérieure — Paris
Visiting PhD student working with Prof. Bruno Loureiro on a theory for Low-Rank Adaptation fine-tuning.
2022–2026
SISSA — Trieste
PhD in Theoretical and Scientific Data Science. Supervisor: Prof. Jean Barbier.
2021–2022
ICTP — Trieste
Postgraduate Diploma in Physics, Quantitative Life Sciences.
2020–2021
AIMS Cameroon
Structured Master's Degree in Mathematical Sciences, Data Science.
2015–2020
University of Douala — Cameroon
Master's degree in Physics, majoring in Science and Radiation Matter, and Bachelor's degree in Physics, majoring in Fundamental Physics.
Contact
I am happy to discuss machine learning theory, statistical physics, scaling laws, LoRA fine-tuning, and related questions.