Reproduce “Scaling Laws for Safety Probes” 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
Yuchen Wu, Jerry Wei
The authors develop scaling laws that predict safety-probe error from model size and the number of labeled examples. They construct six minimal-pair datasets and fit relationships using smaller dense base models, then test forecasts on held-out larger models. At the largest label budget, predictions fall within 0.02 of measured error in 72% of cases, on the studied models and tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.