AI can now compress its own thinking process — cutting computation costs by 70% while actually getting smarter

Researchers built a system that lets language models automatically shrink their internal reasoning steps into compact summaries without losing the ability to solve hard problems. This means running complex AI reasoning on less powerful hardware, or keeping the same hardware and solving harder problems faster.

Language models spend enormous computational energy on intermediate thinking steps — the working-out that happens between input and output. Every token of extra thought is a cost: slower response, higher power use, more expensive deployment. This paper shows you can cut that overhead by 70% while actually improving accuracy on reasoning tasks, which flips the usual tradeoff from a choice between speed and correctness into a straight win. If this holds at scale, the practical effect is immediate: AI systems become cheaper to run, more accessible to smaller companies, and more feasible to deploy on devices that currently can't handle them.