cv
Basics
| Name | Matteo Merler |
| Label | Research Scientist |
| mmerler@fbk.eu | |
| Phone | +39 342 3646961 |
| Url | https://merlerm.github.io/ |
| Summary | I am a pre-doctoral researcher at FBK NLP in Trento, Italy, with an MSc in Machine Learning, Data Science, and AI from Aalto University in Helsinki. My research focuses on bridging large-scale models (such as LLMs and VLMs) with methods like Reinforcement Learning (RL) and Symbolic Planning. I aim to explore how RL's ability to learn from experience can complement the strengths of LLMs, which currently lack this capacity. I'm particularly interested in model-based RL and world models, investigating how large-scale models can be trained to plan ahead and imagine future states, helping them to adapt to complex reasoning and long-horizon planning beyond pattern matching. |
Work
-
2025.01 - Present Trento, Italy
Researcher
Fondazione Bruno Kessler
Working as a pre-doctoral researcher ("Researcher 4 contract") in the Natural Language Processing Lab under the supervision of Bernardo Magnini.
- Lead researcher on FBK’s side of a joint project on planning with Vision-Language Models (with Aalto University): the resulting work is available as a preprint.
- Led a second project on selectively integrating VLM guidance with model-free Reinforcement Learning based on the agent's own uncertainty, manuscript currently under review.
- Participating on private research for collaborations with companies.
- Contributing to the draft of a Horizon Europe funding proposal, leading my lab's team for Language and Reinforcement Learning.
- Active participation to the research group’s scientific activities, including monthly meetings and collaborative discussions.
-
2024.08 - 2024.09 Helsinki, Finland
Research Assistant
Aalto University
MSc Thesis worker in the Foundation Models for Language and Reinforcement Learning (Larel) group, supervised by Prof. Pekka Marttinen.
- Conducted a research project on Symbolic Regression with Large Language Models, presented at the ACL Student Research Workshop.
- Co-led a research project on Code World Models with PhD candidate Nicola Dainese, forming the basis of my Master’s Thesis, accepted at NeurIPS 2024.
- Started a project on planning with Vision-Language Models, later continued by both Aalto University and my current group at FBK.
- Regularly attended lab meetings and participated in broader academic activities, including the Seminar on Large Language Models.
-
2022.02 - 2022.06 Trento, Italy
Internship
ThinkIN
Worked as a software developer intern in collaboration with the University of Trento.
- Designed and built a prototype for an indoor navigation system specifically targeted for supermarkets, to be used with the company's own location tracking system.
Education
-
2022.09 - 2024.09 Helsinki, Finland
Master of Science
Aalto University
Machine Learning, Data Science and Artificial Intelligence
Final Score: 5/5, graduated with honors
- Thesis: Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search, supervised by Prof. Pekka Marttinen.
- Thesis awarded as one of the three best MSc theses of 2024 at the School of Science.
-
2019.09 - 2022.07 Trento, Italy
Bachelor of Science
University of Trento
Computer Science
Final Score: 110 cum laude/110
- Thesis: Route Optimization for Indoor Spaces, supervised by Prof. Alberto Montresor (in Italian).
-
2014.09 - 2019.07 Trento, Italy
Awards
- 2025.03.05
Master's Thesis Award
Aalto University School of Science
My Master's thesis was awarded as one of the three best MSc theses of 2024 at the School of Science.
Publications
-
2025.05.19 ViPlan: A Benchmark for Visual Planning with Symbolic Predicates and Vision-Language Models
arXiv preprint
We introduce ViPlan, a benchmark for evaluating the capabilities of Vision-Language Models for visual planning tasks, under two different settings (VLM-as-planner and VLM-as-grounder) and in two different domains.
-
2024.12.13 Generating Code World Models with Large Language Models Guided by Monte Carlo Tree Search
Advances in Neural Information Processing Systems 37
We investigate the use of Large Language Models (LLMs) for generating world models for Reinforcement Learning. Our approach uses the LLM to generate a Python program that acts as a simulator for the environment, using a novel code generation strategy we call GIF-MCTS
-
2024.08.01 In-Context Symbolic Regression: Leveraging Large Language Models for Function Discovery
The 62nd Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop)
We propose a novel approach to Symbolic Regression that leverages Large Language Models (LLMs) to propose candidate functions that are iteratively scored and refined using the LLM itself.
Volunteer
-
2018.08 - 2018.08 Summer Camp Volunteer
Associazione Volontaria "Per un Mondo Migliore"
I participated in a volunteering camp experience in Rijeka, Croatia where we helped different families and associations in need around the city and in mainland Croatia while living together in a shared home.
-
2017.07 - 2017.07 Summer Camp Volunteer
Libera Associazioni Nomi e Numeri Contro le Mafie
I participated in a volunteering camp experience in southern Italy where we cultivated the fields of mafia confiscated house for the Libera association, while also taking part to various activities on the territory and listening to testimonies from victims and policemen.