IAAIFewer Loops, Less Refusal
Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models
Shows that running a recurrent-depth model with fewer loops than trained, a tempting deployment shortcut, erodes safety much faster than capability: harmfulness is still detected internally, but the refusal that should follow is missing.
Inside the method
[ Shared block ] × r with r < trained depth
Simplified conceptual schematic. Consult the paper for the complete architecture.
- Recurrence family
- Full-stack recurrence
- Depth control
- Inference-time loop reduction
- KV / state strategy
- See paper
Reading note
At half depth, Ouro-1.4B-Thinking keeps 89% of its GSM8K accuracy while harmful-help rate rises 5.3×. Early-exit rules and per-request depth do not repair it; training at randomly chosen depths does, including on Huginn-0125 across 1–32 loops.
Sources checked 2026-10-28. This catalog does not imply independent reproduction.
Related works
Cite this work
@misc{fewerloopsrefusal2026,
title = {Running Fewer Loops Removes Refusal Behaviour in Recurrent-Depth Language Models},
author = {Seyedarmin Azizi and Erfan Baghaei Potraghloo and Arya Fayyazi and Saeid Shokoufa and Massoud Pedram},
year = {2026},
eprint = {2610.15218},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.15218}
}