Recirculation

TL;DR - Recirculation adds inference-time recurrence to pretrained foundation models so they can iteratively update belief states without retraining. Adaptive recirculation substantially improves Gemma 3 perplexity and reasoning accuracy with essentially no added generation latency, though prefill becomes serial.

• Recirculation addresses transformer depth limits by turning the model into a recurrent dynamical system for state tracking.
• The method is distinct from chain-of-thought reasoning, depth looping, and training recurrent transformer architectures.
• Adaptive recirculation freezes the original model weights and requires only light hyperparameter tuning.
• On Gemma 3 models, it reduces perplexity by 23% across a dataset suite and increases GSM8K accuracy by 21% relative to off-the-shelf baselines.