#llm-architecture

Dual-Flow Transformers: A New Architecture That Decouples Reasoning from Generation

Researchers propose a transformer variant that separates the reasoning stream from the generation stream, aiming to reduce inference cost and improve interpretability.