#SwiReasoning

SwiReasoning alternates between silent latent reasoning and explicit chain-of-thought based on next-token entropy, improving accuracy and dramatically cutting token usage on math and STEM benchmarks.