The paper shows a fast, training-free way to boost an LLM’s step-by-step reasoning by smartly reusing the model’s own probabilities.
VOYAGER is a training-free way to make large language models (LLMs) create data that is truly different, not just slightly reworded.