Large language models (LLMs) are powerful, but costly to train, adapt, and deploy when treated as monolithic systems. This thesis studies reusability and efficiency under explicit constraints on parameters, floating-point operations (FLOPs), memory, data, and deployment cost. It treats modularity as an empirical perspective on reuse under constraints: preserving, transferring, or selectively invoking useful behaviour. The thesis develops this perspective through two sets of experiments. The first concerns external parameter-efficient fine-tuning (PEFT) modules. I present a reference architecture and typology for describing PEFT methods as comparable artefacts, and I evaluate portability between corresponding same-scale T5 v1.1 and Flan-T5 checkpoints that share an architectural and training lineage. In this controlled setting, transferred modules sometimes retain useful task-specific information, but the result depends on the PEFT method, porting direction, task, and post-transfer budget. These experiments do not establish portability between independently trained same-architecture hosts, whose parameter spaces need not be aligned. I then study Budgeted LoRA, which organises dense retention and low-rank capacity under a shared budget. Its cost-only controller produces sampled quality–cost operating points within the studied distillation setup; these points are not claimed to be Pareto-optimal, and the larger-teacher budgeted students do not beat the smaller-teacher fully distilled student in absolute perplexity. The second set of experiments concerns internal computation, especially in-context learning (ICL) without parameter updates. Using Bi-Induct, a synthetic-data intervention, I test whether induction-like signatures can be strengthened under matched compute. The signatures become more visible, but targeted ablations do not show a corresponding reliable improvement in few-shot ICL. Within the settings studied, the results therefore distinguish structural or correlational evidence of reuse from evidence that a component is causally important for behaviour.