Reproduce “SAVeR$^2$: Reasoning-based Safety Alignment for Large Reasoning Models via Verifiable Rewards” 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
Yichen Sun, Qirui Liu, Yan Wang, Jianan Lin, Linbo Jiang, Shiyu Wang, Zhibo Wang, Zhixuan Chu
SAVeR² trains reasoning models to assess intent and keep both their reasoning traces and final answers safe. It combines safety rewards with general-reasoning training, projecting safety updates that conflict with reasoning and selecting less-conflicting examples. On BSA, DeepSeek-R1-Distill-Qwen-7B reaches 77.92 Think@1 versus 45.38 for STAR-1; this comparison measures reasoning safety, not universal protection.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.