Recent advances in Large Language Models (LLMs) and their multimodal variants, have enabled the development of agents capable of perceiving, reasoning, and acting in complex environments. Despite this progress, current systems still face significant challenges in reliable decision making, long-horizon reasoning, and systematic evaluation across diverse tasks.

This PhD project aims to design, study, and evaluate LLM-based agents for decision-making tasks in multimodal settings. The research will investigate how LLMs, possibly in combination with other foundation models, can be used as core decision-making components within agentic systems. Different learning paradigms will be explored, including reinforcement learning (RL) as well as alternative approaches based on prompting, planning and language-based reasoning. A key objective is to understand how agents can improve over time through interaction and experience. Possible research directions include continual learning, world modeling, and the autonomous generation of reusable tools, such as code or recipes, to enhance reasoning, adaptability, and generalization.

In parallel, the project will address the challenge of evaluation by developing benchmarks and experimental protocols for agents operating in simulated or real-world-inspired multimodal environments. The goal is to enable rigorous comparison across agent architectures and learning strategies, contributing to the development of more reliable and scalable agentic systems.