Google DeepMind published a paper describing an inference technique called « recirculation » that reinjects activations from a model’s deeper layers back into its shallower layers during token generation. Tested on the Gemma3 model family (1B, 4B, and 12B parameters), this method achieved a 23% reduction in perplexity without requiring any retraining. Adaptive recirculation even outperformed full fine-tuning, which only delivered a 21.6% reduction. On the GSM8K mathematical reasoning benchmark, the same technique produced a 21% relative increase in accuracy. The tradeoff is a more expensive initial prompt processing, as prefill becomes serial under recirculation.
Source: Read the original article
