Stanford and Meta Researchers Develop New AI Inference Technique for Large Language Models

A research team from Stanford University, in collaboration with Meta and other institutions, has introduced a novel AI inference technique designed to enhance the speed and reduce the computational costs of large language models (LLMs). The method processes intermediate reasoning in an internal latent space rather than generating text tokens, combining latent and text modes with a 'HybridLink' for continuous reasoning. Tests on Qwen3-based models demonstrated reduced token generation and increased inference speed, alongside improved problem-solving performance. This work has not yet been peer-reviewed.