Reproduce “Tracing the Scientific Discovery Capacity of LLMs to Their Training Data” 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
Arda Uzunoglu, Mahler Revsine, Michael Schatz, Benjamin Van Durme, Mehrdad Farajtabar, Daniel Khashabi, Eric Nalisnick
The authors test whether the rate of training examples seen only once relates to scientific discovery. They vary this monofact rate and evaluate exact recovery of materials formulas and DNA sequences using two model families. Across both tasks, models, training stages, and sampling settings, higher monofact rates accompany higher discovery rates; this relationship is limited to the studied setups.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.