Reproduce “Functionality learning through specification instructions” 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
Pedro Henrique Luz de Araujo, Benjamin Roth
The authors test whether plain-language behavior rules added to prompts can address NLP test-suite failures without updating model weights. They evaluate four English-language tasks across seven models, comparing task-only prompts with specification-augmented prompts. Effects vary by task and model. On HATECHECK with Davidson, Flan-T5-XXL’s aggregate score is 29.00% versus 24.39%, but this single-run result does not establish broad real-world generalization.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.