No-Training Depth Boost? Recirculation Paper

teortaxesTex · x · 2026-08-19

A paper titled 'Recirculation' proposes an inference-time architectural enhancement that injects top-layer activations into bottom layers at the next step. This allows the model to act as a dynamical system to track belief states without retraining. It achieves a 23% perplexity reduction and a 21% accuracy boost on GSM8k for Gemma3, with minimal latency overhead during generation.

Related event: Recirculation Boosts LLM Reasoning Without Any Retraining(3 posts)→

Original post →

More from Research

Research channel →