Language models just got 40% more efficient by rejecting irrelevant information instead of weighing everything

Researchers built a language model architecture that evaluates whether each piece of information is worth paying attention to, rather than forcing the model to distribute attention across everything whether relevant or not. This means models can use far fewer parameters to do the same job, run faster on long documents, and don't degrade when asked about information beyond their training data.

Every language model since 2017 has worked by weighing every input against every other input — a global popularity contest where even garbage information gets a vote. This paper shows that absolute thresholds work better: if something is clearly irrelevant, throw it out. The practical effect is massive efficiency gains, which matters because inference latency and parameter count are the actual costs that determine whether AI tools stay profitable or become too expensive to run. If this generalizes beyond the preprint, you're looking at faster, cheaper models doing the same work.