Arbitrage: Efficient Reasoning via Advantage-Aware Speculation

Modern Large Language Models achieve impressive reasoning capabilities with long Chain of Thoughts, but they incur substantial computational…

Raccontata damachinelearning.apple.com