Reproduce “In-Context Learning Amplifies a Latent Symbolic Circuit” 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
Melissa Wessel
The authors test whether demonstrations strengthen an existing computation for pattern following, rather than create a new one. They trace a three-stage circuit: Symbolic Abstraction maps examples to variables, Symbolic Induction infers a rule, and Retrieval selects a token. In Gemma, patching raises zero-shot ABA accuracy from 1% to 56% using downstream stages, supporting this account on a narrow task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.