
LTL Verification of Memoryful Neural Agents
Mehran Hosseini*, Alessio Lomuscio, Nicola Paoletti - AAMAS 2025 - PDF coming soon.
I'm a postdoctoral researcher at King's College London's Informatics Department and a founder of NaPolleon, an AI-powered survey and prediction market backed by Oxford University Innovation. My research focuses on developing mathematical foundations for safe and verifiable AI systems, with particular emphasis on scalable verification tools and novel deep learning architectures that are more capable and amenable to formal verification.
At the KCL, I lead the research on safety verification of autonomous neural agents as part of the EU Horizon's REXASI-PRO project, collaborating with Nicola Paoletti and DDV Lab. Before joining KCL, I was at Imperial College London's Safe Artificial Intelligence Lab (SAIL) collaborating with Alessio Lomuscio on scalable verification of recurrent neural nets. Prior to this, I completed my DPhil in computer science at Oxford University, where I solved the universal termination problem for linear loop programs. At Oxford, I was lucky to have James Worrell and Joël Ouaknine as my supervisors.
Mehran Hosseini*, Alessio Lomuscio, Nicola Paoletti - AAMAS 2025 - PDF coming soon.
Francesca Cairoli, Francesco Giacomarra, Mehran Hosseini, Nicola Paoletti - AAMAS 2025 - PDF coming soon.
Mehran Hosseini*, Peyman Hosseini - arXiv Preprint 2024 - PDF.
Linus Jeary, Tom Kuipers, Mehran Hosseini, Nicola Paoletti - NeurIPS 2024 - PDF.
Mehran Hosseini*, Peyman Hosseini - WACV 2025 - PDF.
Ben Batten, Mehran Hosseini*, Alessio Lomuscio - AISTATS 2024 - PDF.
Mehran Hosseini*, Alessio Lomuscio - AAMAS 2023 - PDF.
Mehran Hosseini* - University of Oxford 2021 - PDF.
Mehran Hosseini*, Joël Ouaknine, James Worrell - ICALP 2019 - PDF.
Shaull Almagor, Brynmor Chapman, Mehran Hosseini*, Joël Ouaknine, James Worrell - CONCUR 2018 - PDF.
Mehran Hosseini*, Reza Rezaeian Farashahi - Finite Fields and Their Applications 2016 - PDF.
Vern is verifier for neural networks which supports memroyful models and LTL specifications.
Hummingbird is a proof-of-concept small & efficient language model based on Efficient Attention, trained on only 15B token.