Reproduce “Cache2LoRA: Understanding and Improving Context Internalization” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yi Fang, Gaurav Srivastava, Zhenyu Bi, Sirui Ding, Xuan Wang
Cache2LoRA writes query-specific LoRA updates from a reusable document cache, avoiding repeated document encoding for each answer. A trainable reader uses the question to select information and generate the update; varied teacher answers train response adaptation. On Qwen3-4B, macro token F1 is 0.663 versus 0.671 for full-context ICL across 17 tasks under standard short-answer instructions, not equivalence on every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.