Reproduce “Generative Monoculture in Large 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
Fan Wu, Emily Black, Varun Chandrasekaran
The authors measure generative monoculture: narrower spread in task-relevant attributes of model outputs than in human source data. They examine book reviews and code generation, comparing extracted attributes for the same tasks. Goodreads and CodeContests are proxies, not verified training data. Generated reviews concentrate in positive sentiment, while generated code is more accurate but less diverse on several measures.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.