Robustness of Fine-Tuned Llms Under Noisy Retrieval Inputs: Failure Modes and Error Containment

Abstract

Aggregate performance can conceal concentrated failures, so errors must be classified by cause, consequence, and the controls available to contain them. This structured evidence review evaluates "Robustness of Fine-Tuned Llms Under Noisy Retrieval Inputs" alongside nine author-disjoint, topically matched publications in reliable retrieval-augmented generation. It compares construct definitions, evaluation choices, operating assumptions, and reported limitations instead of treating bibliographic similarity as empirical equivalence. Viewed through error taxonomy and failure containment, the map separates claims supported by the available record from questions that still require full-text extraction, replication, or new experiments. The synthesis is interpretive rather than meta-analytic and therefore does not present a pooled effect estimate or a new causal result. The resulting agenda pairs a documented error taxonomy with stress tests, escalation rules, and safeguards for high-consequence failures.