Reinforcement Learning for Prompt Optimization in Language Models: A Comprehensive Survey of Methods, Representations, and Evaluation Challenges

The growing prominence of prompt engineering as a means of controlling large language models has given rise to a diverse set of methods, ranging from handcrafted templates to embedding-level tuning. Yet, as prompts increasingly serve not merely as input scaffolds but as adaptive interfaces between users and models, the question of how to systematically optimize them remains unresolved. Reinforcement learning, with its capacity for sequential decision-making and reward-driven adaptation, has been proposed as a possible framework for discovering effective prompting strategies. This survey explores the emerging intersection of RL and prompt engineering, organizing existing research along three interdependent axes: the representation of prompts (symbolic, soft, and hybrid), the design of RL-based optimization mechanisms, and the challenges of evaluating and generalizing learned prompt policies. Rather than presenting a single unified framework, the discussion reflects the fragmented, often experimental nature of current approaches, many of which remain constrained by unstable reward signals, limited generalizability, and a lack of reproducible evaluation standards. By analyzing methodological innovations and points of friction alike, this work aims to foster a more critical and reflective understanding of what it means to "learn to prompt" in complex, real-world language modeling contexts.