Reproduce “Learning to Extrapolate to New Tasks: A Relational Approach to Task Extrapolation” 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
Adam Ousherovitch, Yixin Wang
RTE extrapolates to unseen tasks by reusing a familiar task and applying a learned relation, rather than predicting directly. For continuous shifts it uses task embeddings; for discrete changes, it proposes anchor–transformation pairs and checks them against target examples. On 200 held-out CodeIO compositions, Mistral-7B reaches 45.3% exact-match accuracy versus 30.2% for chain-of-thought majority vote, below the 76.0% oracle.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.