Mátyás Vincze

PhD researcher / language agents / reinforcement learning

I am a PhD researcher at the University of Trento , conducting research at Fondazione Bruno Kessler with Bruno Lepri and Giovanni Iacca . I am currently visiting Michiel Bakker 's group at MIT .

I work on post-training for long-horizon language agents, combining distillation, multi-agent feedback, and reinforcement learning with verifiable rewards. My earlier work on interpretable reinforcement learning led to SMoSE , a sparse mixture of shallow policy experts.

Updates

[05/26] Visiting Michiel Bakker 's group at MIT to work on post-training for long-horizon language agents 🚀
[03/26] ISCRA compute grant from CINECA — more A100s go brrr.
[01/26] Driving license secured ✅
[04/25] ISCRA compute grant from CINECA — A100s go brrr, again.
[03/25] Attended the ALPS 2025 winter school in Aussois, France.
[02/25] Presented SMoSE at AAAI 2025 in Philadelphia.
[12/24] SMoSE accepted to AAAI 2025 .
[07/24] Placed 1st in the Interpretable Control Competition at GECCO 2024.
[06/24] ISCRA compute grant from CINECA — A100s go brrr.

Selected publications

see all

A benchmark of expert-level academic questions to assess AI capabilities (Humanity’s Last Exam)
Center for AI Safety, Scale AI, HLE Contributors Consortium

Contributor to the expert-level AI benchmark published in Nature.
@ Nature · 2026 |
- Paper
- Website
SMoSE: Sparse Mixture of Shallow Experts for Interpretable Reinforcement Learning in Continuous Control Tasks
M. Vincze, L. Ferrarotti, L. L. Custode, B. Lepri, G. Iacca

First-author work on sparse, interpretable reinforcement-learning policies with only 108–672 active actor parameters.
@ AAAI · 2025 |