DeepMind's Recirculation boosts LLM reasoning without retraining

A Google DeepMind paper introduces "Recirculation," an inference-time architectural enhancement that improves off-the-shelf foundation models without retraining or weight modification. On Gemma3 it achieves a 23% perplexity reduction and a 21% GSM8k accuracy gain, and consistently lowers perplexity across five model families. With zero training cost and minimal inference overhead, it is worth watching as a low-cost, general way to boost open models.

Confirmed

• Mechanism: a form of recurrence feeds part of the upper-layer activations back into early shallow layers while processing the next token, letting the model act as a dynamical system tracking belief states and overcoming feedforward Transformer limits, effectively granting dynamic working memory for long-context processing. @danielmac8 likens it to installing a mail chute in an office building so lower floors learn in real time what the top has processed.
• Results: Gemma3 perplexity down 23%, GSM8k relative accuracy up 21%.
• Generalization: @xennygrimmato's comparisons show Recirculation lowering perplexity across five model families versus the Unrolled loop transformer; a source-destination pair scan identifies regions with consistent improvement.
• Overhead: per @chaumian, the method adds almost no latency during generation, with serial processing needed only in the prefill stage.

Why it matters

• Significant capability gains without retraining mean any team holding an off-the-shelf model could benefit at low compute cost; @teortaxesTex frames it as "adding depth without training."
• Compared with prior recurrent approaches like the Unrolled loop transformer, Recirculation shows more stable perplexity improvements, making it a candidate route for inference-time architectural enhancement.

2026-08-19 ~ 2026-08-21 · 10 related posts

• [source] No-Training Depth Boost? Recirculation Paper — teortaxesTex · 2026-08-19
• Recirculation: Training-free inference architecture boosts model performance — chaumian · 2026-08-20
• Recirculation: Off-the-shelf models self-modify for instant inference boost — TheGradient · 2026-08-20
• Comparing Unrolled Loop and Recirculation Transformers — xennygrimmato_ · 2026-08-20
• [source] Recirculation Architecture Reduces Perplexity Across Gemma3 Models — xennygrimmato_ · 2026-08-20
• Recirculation Shows Perplexity Reduction Promise Across Five Model Families — xennygrimmato_ · 2026-08-20
• Recirculation adds inference-time memory to LLMs without retraining, boosting Gemma3 accuracy by 21% — pbaylies · 2026-08-20
• Research: Inference-time Recurrence Cuts LLM Perplexity by 23% Without Weights Update — heghbalz · 2026-08-20
• DeepMind's Recirculation Cuts Perplexity by 23% With Zero Weight Changes — daniel_mac8 · 2026-08-21
• [source] Google DeepMind's Recirculation cuts perplexity by 23% without retraining — TheTuringPost · 2026-08-21