Answers are grounded in this paper and cite their sources.
Serving large language models at scale is a distributed systems problem whose overload behavior is governed at the ingress. We conduct an adapted-PRISMA systematic review of fifty-five admission and admission-adjacent works published between 2022 and 2026, and report three findings. First, admission sits at a different layer from the one the serving literature studies: of the nine LLM serving systems that the admission literature most often builds on, all place the overload response inside the engine, through batching, migration, placement, autoscaling or memory management, and none takes a decision at the ingress, while the corpus work that does gate at ingress is a control-plane contribution rather than an engine one. Second, of sixteen admission-primary papers, no two share a baseline set: policies are measured against the absence of admission control and almost never against one another. Third, placing all sixteen on our axes succeeds on six of seven properties and fails on one. Unit, setting, objective, the quantity a policy watches, the metadata it assumes, and whether it is deployable in front of an unmodified engine are all recoverable from the published text; the observed range of the quantity the policy reads is reported by none of them. Selection therefore fails on a single missing fact, and it is the fact that determines whether a reported improvement is interpretable at all. We contribute systematic coverage of fifty-five works under an adapted-PRISMA protocol; a five-axis taxonomy of unit, setting, SLOawareness, signal quantity and portability, with a master classification of the twenty-eight works that propose or embed an admission mechanism, each carrying a GRADE-adapted evidence rating; formal definitions of the admission-decision, goodput-under-admission and fairness-index primitives; a cross-paper synthesis of admission-signal overhead; an artifact-availability audit and seven properties on which the corpus diverges, which a cross-paper comparison would have to recon cile; a mapping of seven public production incidents to admission failure modes; and seven open problems, each grounded in published peer review or in a primary source’s own limitations.
No takes yet. Share an insight, caveat, or question.
Jain et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: