Decisions and Feedback Across the Lifecycle of Hardware-Efficient Language Model Systems

Hardware-efficient Transformer-based language-model systems make consequential choices across development/adaptation, deployment preparation, and runtime/serving, yet mechanism-centric organization can obscure their dependencies. This critical survey uses two-stage Scopus retrieval, structured screening, detailed assessment of eligible cross-lifecycle evidence, and targeted contextual enrichment. Its decision-centric framework organizes Evidence → Decision → Action relations through lifecycle locus, evidence, action, hardware role, timescale, forward coupling, and feedback. The reconciled evidence shows upstream-to-runtime coupling through prepared choice spaces, conditional capabilities, and profile/information interfaces, which may overlap. Feedback is distinguished as search/configuration feedback (FB-S), runtime-local feedback (FB-R), operational evidence reused upstream (FB-O), and strict lifecycle closure (FB-C). Industrial RAG verifies FB-O through operational records reused in upstream parameter-efficient adaptation; the inspected evidence does not establish subsequent deployment of the revised artifact with an observed operational consequence. No source in the reconciled corpus was confirmed as strict FB-C. Distinguishing hardware-evidence origin (H_SPEC, H_PROFILE, H_LIVE) from functional role (H0–H4), the seven-dimensional taxonomy supports structural comparison and a corpus-bounded agenda addressing upstream evidence reuse, operationally validated closure, hardware evidence crossing upstream, and cross-timescale escalation. It does not imply mechanism rankings or field-wide prevalence.