Luca Zhou
About Me
I am a first-year PhD student in Computer Science at GLADIA, Sapienza University of Rome. Previously, I was a visiting graduate student researcher at UC San Diego.
My research primarily focuses on model merging, activation steering, and mechanistic interpretability. I aim to build methods and tools that make large models more adaptable, composable, and interpretable. I’m also keen on discovering weaknesses of SOTA models through challenging benchmarks.
Outside the lab, I stay active through sports 🥊🤸♂️ and explore my creativity through photography 📷 and music 🎹🎸. I’m always happy to collaborate, exchange ideas, or just have a chat. Feel free to reach out!
Research Interests
- Model Merging
- Activation Steering
- Mechanistic Interpretability
Education
Experience
News
-
[06-2026] 🎉 Are questions where models fail inherently more difficult? Not always, check out our latest paper accepted to AI4Math @ ICML 2026!
-
[04-2026] 🎉 Our paper on interpretable mergeability prediction for model merging has been accepted to ICML 2026, check it out here!
-
[04-2026] 🎉 CaTS-Bench, our benchmark for time series captioning and reasoning, has been accepted to ACL 2026 Findings, read it here!
- [09-2025] 🎉 Excited to share that our paper “On Task Vectors and Gradients” will be presented at UniReps @ NeurIPS 2025 in San Diego!
-
[07-2025] 🎉 I’ve graduated with an MSc Degree from Sapienza University of Rome with honors! Check out my thesis here.
-
[07-2025] 🎉 Our paper “ATM: Improving Model Merging by Alternating Tuning and Merging” has been accepted for presentation at the Breaking the Monolith: 1st ICIAP Workshop on Advances in Modular Deep Learning!
- [03-2025] 🚀 I’m visiting UC San Diego as a visiting student researcher, working on a Multimodal Benchmark on Time Series Captioning & Reasoning alongside Rose STL Lab.