Contextual Conditioning Does Not Erase LLM Memorization
October 9, 2026
Investigation of open-weight instruction-tuned models reveals that retrieving documents or adding instructions does not mitigate memorization. Instead, contextual conditioning shifts which specific memorized samples from the training set become extractable via prefix-based prompting.
HOW THIS AFFECTS YOU
●
researcherThis changes how you evaluate the safety and privacy boundaries of retrieval-augmented systems.
●
policyYou should account for the fact that RAG and instruction tuning do not inherently solve data privacy/memorization risks.