IAAISafeBridge
Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models
Evaluates jailbreak robustness of looped language models at different recurrent depths and finds that safety can differ by depth and attacks transfer across depths. SafeBridge aligns safety across depths with depth-specific control of shared layers, selective state bridging and joint supervision.
Inside the method
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Multiple readout depths
- KV / state strategy
- See paper
Reading note
SFT and preference alignment alone leave depth-wise safety gaps. SafeBridge reduces matched- and cross-depth attack success with comparable over-refusal. Code and checkpoints are promised upon acceptance.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{safebridge2026,
title = {Safe at One Loop, Risky at Another: Aligning Safety Across Recurrent Depths in Looped Language Models},
author = {Yi Wang and Xiuyuan Qi and Dongqi Han and Dongsheng Li and Wenjie Wang},
year = {2026},
eprint = {2610.10625},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.10625}
}