Rethinking LoRA Memory Through the Lens of KV Cache Compression

In three linesStudy of LoRA and KV cache interaction in question-answering. LoRA adapters become useful under aggressive cache compression (recovering 13-21 ROUGE-L points), functioning as parametric memory at decoding time rather than document encoder. QA-style supervision produces stronger adapters than raw next-token prediction.

Summary generated by Claude — human-verified