Peak-then-decline behavior occurs when additional retrieved context improves answer quality only up to a small optimum, after which accumulated noise reduces performance. It makes retrieval depth a model-specific production decision rather than a simple recall maximization problem. section 15.0
The pattern changes how a RAG system should be evaluated and tuned:
- Measure the reader's optimal retrieval depth instead of assuming that larger k produces better answers. section 15.0
- Separate stability from scalability: performance may remain consistent near an optimum without continuing to improve as context grows. section 15.0 section 15.9
- Treat higher retriever recall cautiously when the reader is noise-sensitive; semantically similar passages can distract more than less relevant passages. section 15.0
A learned retrieval gate
A lightweight gate can adjust retrieval influence per token, suppressing passages that add noise and helping a noise-sensitive reader avoid the declining side of its performance curve. section 15.4.1
