This PhD position is integrated into the ERC-funded STEP-RL project, which focuses on the “specialization” of domain-independent temporal planners to achieve high-performance execution in complex environments. The research aims to explore the synergy between Large Language Models (LLMs) and Reinforcement Learning (RL) as a means to automate the transition from human-centric requirements to optimized solver guidance. The research will investigate one or more of the following promising directions: the use of LLMs for the synthesis of formal planning problems from informal or underspecified specifications; the application of generative models for the augmentation of training data to improve the robustness of RL agents; and the synthesis of domain-dependent heuristic guidance to accelerate symbolic search. Crucially, this is not a standard LLM-centric project focused on text generation; rather, it treats LLMs as a component within a rigorous neuro-symbolic pipeline, where the ultimate goal is to enhance the efficiency and formal correctness of symbolic temporal solvers. By investigating these intersections, the candidate will contribute to a framework where LLMs facilitate the modeling and learning phases, while RL and symbolic solvers ensure computational efficiency and formal rigor. This position offers the opportunity to define a unique research path at the forefront of automated planning and machine learning, directly addressing the challenge of making temporal planning both more accessible and more scalable for real-world application.