Silver-evidence demonstrations improve ICL by filtering noisy retrieval
October 26, 2026
This method replaces raw, noisy RAG passages with LLM-generated, target-conditioned evidence summaries during in-context learning. By pairing these silver-evidence demonstrations with expert labels, the approach mitigates the impact of redundant or distracting information in long contexts.
HOW THIS AFFECTS YOU
●
builderYou can improve RAG performance by using an LLM to pre-process retrieved context into structured summaries.
●
researcherThe method provides a way to move from simple answer-side rationales to active supervision for retrieval filtering.