#AIME

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.

'TUMIX ensembles diverse tool-using agents that share notes and use an LLM judge for early stopping, improving accuracy on hard reasoning benchmarks while cutting inference cost.'

'Meta AI compresses recurring CoT reasoning into short named behaviors that models can retrieve or internalize, cutting reasoning tokens by up to 46% and improving accuracy in some settings.'