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.
This paper teaches video-making AIs to follow real-world physics better without retraining them.