#parallel reasoning

'Researchers introduce Soft Thinking, a training-free method that allows large language models to reason with continuous concept embeddings, enhancing accuracy and efficiency in math and coding tasks.'