Reproduce “An Algorithmic Account of Few-Shot Concept Induction in Language Models” 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
Michael A. Lepori, Laila Johnston, Aalok Sathe, Zhuonan Yang, Roman Feiman, Ellie Pavlick
Amortized Direction Accentuation (ADA) models few-shot concept induction by summarizing positive examples and applying a learned feature transformation reused across contexts. In tests on five language models, same-concept words received larger context-driven token-probability boosts than other-concept words. For Gemma-2-2B, ADA outperformed rotation and matched or slightly exceeded an MLP on held-out support sets, supporting reusable accentuation, not a complete account.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.