Recirculation
The paper claims inference-time recurrence can improve foundation models without retraining their weights.
Mozer and coauthors describe “recirculation,” an architectural change applied to off-the-shelf models during inference. It lets the model update internal belief states beyond the fixed depth limits of a feedforward transformer, while adding little generation latency but serial work in prefill. Their adaptive version freezes the base model and reports gains on Gemma3, including lower perplexity and higher GSM8k accuracy. ArXiv · AI/CL/LG's note
Mozer and coauthors describe “recirculation,” an architectural change applied to off-the-shelf models during inference. It lets the model update internal belief states beyond the fixed depth limits of a feedforward transformer, while adding little generation latency but serial work in prefill. Their adaptive version freezes the base model and reports gains on Gemma3, including lower perplexity and higher GSM8k accuracy. ArXiv · AI/CL/LG's note
score 5