Reproduce “Safin-1.5: Safety-Native Architecture and Verifiable Reasoning for Recursive Safety Improvement” 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

Jingnan Zheng, Cheng Jin, Ming Zhang, Huan Sun, Zhongtian Ma, Kaisen Yang, Ziyuan Zhou, Tong Wu and 23 more

Safin-1.5 combines removable experts, a recursive safety-training loop, and an optional formal-reasoning harness. It relocates targeted hazardous knowledge into experts that are excised for release, while updating the full model. After matched 250-million-token post-training, its WMDP-Bio score was 40.8%, versus 81.7% for ordinary unlearning; results cover tested domains and do not establish universal removal.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.