A new paper from Google DeepMind introduces a training-free method to enhance model architectures by enabling them to inform their own modifications. This approach, called recirculation, allows feedforward transformers to act as dynamical systems, feeding activations back through themselves during prefill to track belief states without retraining. This method has shown significant improvements, with an adaptive variant cutting perplexity by 23% and boosting GSM8k accuracy by 21% on the Gemma3 family of models, using only hyperparameter tuning on original weights. AI
IMPACT This training-free approach could lead to more efficient model development and improved performance without extensive retraining.
RANK_REASON The cluster describes a research paper detailing a novel method for improving LLM architectures. [lever_c_demoted from research: ic=1 ai=1.0]
Read on X — Omar Sanseviero (HF research) →
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
