Reproduce “Exploring the Limits of the Delta Learning Hypothesis: Single-Pair Generator Biases in Direct Preference Optimization” 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
Giuseppe Curci, Stefano Simonazzi, Andrea Molinari, Andrea Zugarini
The paper tests fixed model pairs as preference data for Direct Preference Optimization. A larger model’s answer is labeled chosen and a smaller model’s rejected, but differences in their writing can become shortcuts. For OLMo2-1B-SFT, original preference data improves the average across four benchmarks by 4.80 percentage points, versus −0.14 for five single pairs; style causality remains unproven.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.