Google DeepMind just published a paper that might quietly change how language models process text. The technique, called “recirculation,” takes activations from a transformer’s deeper layers and mixes them back into its shallower layers during inference. The paper, co-authored with researchers from the University of Texas at Austin, demonstrates that this lightweight recurrent connection delivers performance gains that rival, and in some cases surpass, full fine-tuning. No retraining required. No major architectural surgery.
What recirculation actually does
Recirculation breaks the one-way flow of transformer processing. A fraction of the activations computed in the model’s deeper layers gets fed back into the shallower layers during token generation. The model essentially gets a second pass at understanding its own internal representations, allowing it to refine what the paper calls its “belief states” across multiple steps.
The key distinction from existing approaches like chain-of-thought prompting or looped transformer architectures is that recirculation doesn’t require extra reasoning tokens or added architectural depth. It’s a bolt-on modification to how inference runs, not a redesign of the model itself.
The numbers are hard to ignore
The DeepMind team tested recirculation across the Gemma3 model family, including 1B, 4B, and 12B parameter variants. Basic recirculation delivered an approximate 8.5% reduction in perplexity across nine language modeling datasets, with zero added latency during generation.
Adaptive recirculation pushed further, achieving a 23% mean reduction in perplexity across those same nine datasets. For context, full fine-tuning only managed a 21.6% reduction. On reasoning tasks, the GSM8K benchmark saw a 21% relative increase in accuracy with adaptive recirculation.
There is a tradeoff. Prefill processing, the stage where the model digests the initial prompt, becomes serial under recirculation, increasing the upfront cost of processing a prompt.
Why the AI community is paying attention
The paper, listed as arXiv:2608.17981, has already been independently reproduced. GitHub implementations have confirmed the gains on models outside the Gemma family, including Llama 3.2 1B. Discussions have spread across research communities on X and Chinese tech media platforms.
Recirculation operates at the activation level, below the token-generation surface, which makes it complementary to rather than competitive with prompting techniques like chain-of-thought reasoning, which consume output tokens and add latency.