Reproduce “Rules Amortize, Pairings Don't: Linguistic Structure Determines What Latent Task Representations Can Replace In-Context Learning” 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
Gunmay Jhingran
MCT reduces repeated in-context learning prompt costs by caching a summary of support examples and using a query-conditioned network to update a frozen language model’s internal representation. On tested inflection tasks, it reaches 0.85 ± 0.02 held-out-pair accuracy for past-to-present on GPT-2-XL, versus 0.13 ± 0.01 for 10-shot ICL. On antonyms, it trails ICL, so results do not establish generality beyond tested tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.