AI reasoning models can now get 67% shorter while staying more accurate — if they think right instead of just longer

Researchers found that the best chain-of-thought reasoning in language models isn't about reducing uncertainty everywhere — it's about following a specific pattern where uncertainty drops steadily as the model thinks through a problem. A new training method called Entropy Trend Reward pushes models toward this pattern, resulting in shorter reasoning steps that are actually more accurate than longer ones.

Most AI reasoning tools today assume that longer thinking is better thinking, so they either penalize length crudely or try to reduce uncertainty everywhere at once. This paper shows the pattern actually matters more than the volume — a model that gradually gains confidence toward an answer beats one that just keeps exploring. In practice, this means AI reasoning could become faster and more efficient without sacrificing accuracy, which matters because reasoning tokens currently consume the most compute during inference. The tradeoff that seemed locked (short but dumb, long but accurate) just became moveable.