Recurrent-depth language models reuse a single block of weights multiple times in sequence. The number of repetitions, called loops, is not fixed during training but chosen when the model runs. This design allows the same model to behave differently at inference depending on how many times the block is applied.
A new preprint reports that reducing the loop count removes refusal behaviour. Models that would decline certain prompts become more compliant when run with fewer loops. The effect appears tied to the depth of recurrence rather than to any retraining or prompt engineering.
Because running fewer loops is also cheaper and faster, the finding suggests a simple inference-time lever for adjusting model behaviour. The authors do not propose a mechanism, but the result points to loop count as a meaningful control knob for recurrent architectures.