Controlled Text Generation (CTG) aims to produce text that reliably adheres to specified attributes, such as sentiment, topic, or factual content. Despite the impressive generation capabilities of large language models (LLMs), standardised evaluation protocols are largely absent, simultaneous control over multiple attributes remains unreliable, and existing fine-tuning approaches are either computationally prohibitive or too rigid to accommodate new control requirements without retraining. This thesis addresses these challenges by developing principled evaluation methodology, systematically analysing the CTG landscape, and proposing modular, reusable control mechanisms. Evaluation in CTG research lacks principled foundations. We develop a level-playing-field (LPF) framework that addresses four key sources of evaluation bias: system selection, metric selection, dataset selection, and output-generation protocols. Applied to 12 CTG systems, the framework demonstrates that reported scores and system rankings vary substantially with methodological choices. We introduce a multi-dimensional categorisation scheme for CTG techniques along three dimensions: control attributes, control integration, and control implementation. Validated through manual annotation of 140 papers and extended to 178 further papers via an LLM-based annotation system, the scheme reveals the dominance of single-attribute control and the rise of prompting. Comparing prompting-based and Parameter-Efficient Fine-Tuning (PEFT) approaches for content-controlled generation, we find that LoRA-adapted (Low Rank Adaptation) Mistral 7B matches GPT-3.5 within 1% BLEU despite a 25-fold difference in parameter count, showing that PEFT can largely compensate for model scale. The central contribution is a plug-and-play framework of attribute-specialised LoRA modules, composable at inference time for multiple-attribute control without joint retraining. Experiments with LLaMA 3 and Mistral show that composing independently trained modules via output summing outperforms individually trained ones. Few-step adaptation enables module reuse across architectures at a fraction of retraining cost. This thesis demonstrates that modular, attribute-specialised PEFT approaches offer a principled and efficient path towards flexible, reusable control in large language models.