Retrieval-Augmented Models Reduce Lexical Frequency Sensitivity

Using k-nearest-neighbor retrieval-augmented language models can mitigate the performance gap seen in vanilla LLMs when processing rare lexical items. The approach tests the hypothesis that hippocampal-style episodic memory mechanisms improve robustness to syntactic contrasts regardless of word frequency.