USC Study Finds Reducing Inference Loops Removes Refusal Behavior in Recurrent-Depth LLMs
Preprint shows that reducing loop counts for efficiency in recurrent-depth LLMs disproportionately degrades safety refusals, requiring random-depth training to fix.
ImportanceMaterialEvidenceE2 unreplicatedWrite-upQuick
A research team from the University of Southern California found that for recurrent-depth language models, which reuse a single block of weights multiple times, reducing the number of inference loops to save cost and latency leads to a disproportionate loss in the model's ability to refuse harmful requests.
On the Ouro-1.4B-Thinking model, halving the loop count retained 89% of accuracy on GSM8K math tasks but increased the rate of assisting with harmful requests from 3.8% to 20.5%, a 5.3-fold rise. Internal state analysis showed the model could still identify harmfulness (ROC AUC 0.967), but the mechanism converting this identification into refusal was missing at shallower depths.
The study notes that standard safety training is effective only at full depth. To maintain safety across varying inference depths, fine-tuning must be performed using randomly chosen depths during training. This finding warns deployers that configurations optimized solely for computational cost may inadvertently bypass safety guardrails.