Reproduce “Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models” 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
Yi Wang, Xiuyuan Qi, Dongqi Han, Dongsheng Li, Wenjie Wang
SafeBridge aligns safety across recurrent depths in looped language models, where jailbreak responses can vary with the readout depth. It combines depth-specific controls, a gated state bridge, preference training, and safety labels. On Ouro-1.4B, it reduces overall attack success rate from 21.06% with DPO to 4.56%; this result covers the tested benchmarks and attacks, not a guarantee elsewhere.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.