About

I am a research scientist at Qualcomm AI Research in Amsterdam, where I lead a team of three working on RL for LLM post-training, efficient reasoning, and looped language models. I introduced the Conditional Information Bottleneck view of budget forcing (NeurIPS 2026) and a looped-transformer design that cuts the KV cache of Ouro models by 4x.

I came to machine learning from physics. I did my PhD at the University of Bologna and Columbia University, building Monte Carlo simulations and statistical analyses for the XENON dark-matter experiments, after a summer on CMS data at CERN. Before Qualcomm, I was a researcher at ISTI-CNR in Pisa and a visiting researcher at USC’s Information Sciences Institute, working on face recognition across resolutions and the detection of adversarial attacks.

I review for NeurIPS, ICML and ICLR, and I am a co-inventor on more than ten US patents and applications.

Short bio

Fabio Valerio Massoli is a research scientist at Qualcomm AI Research in Amsterdam, where he leads work on efficient reasoning for large language models, including RL for post-training and looped transformers. He holds a PhD in Physics from the University of Bologna and Columbia University.

Contact

Email fabiovaleriomassoli@gmail.com, or download my CV for the full list of roles, patents and publications.