Abstract
Compression changes both the information retained by a large language model and the resources required to complete a request. This narrative review examines quantization, pruning and sparsity, knowledge distillation, low-rank factorization, and token, context, and key–value cache compression for autoregressive text generation, principally with decoder-only LLMs. Its central contribution is a synthesis of changes relative to each study’s baseline, connecting compression mechanisms to capability and deployment choices. Four numerical tables distinguish quality changes at stated compression budgets, distillation gains over same-size students, and paired or independent deployment measurements. Reporting is appraised by source identity, comparison conditions, measurement scope, artefacts, and uncertainty. The reviewed cases illustrate how capacity constraints change the feasible precision choices, how allocating a shared budget changes different task scores, and how compact-model construction and request compression address complementary resource demands. A deployment-validation sequence turns these findings into candidate checks for a target workload. Remaining questions concern transfer across capabilities, delayed use of compressed information, interactions between compression stages, and complete-task cost, including agent interactions.