What happened
Researchers built a smarter way for AI language models to compress long documents by recognizing that some parts contain much more useful information than others, and treating dense sections differently than repetitive ones. Instead of compressing every part equally, the model now picks from a set of predefined compression levels based on what it detects — which makes the system actually work in practice, whereas trying to adjust compression continuously breaks the model's reasoning.
Why it matters
This matters because long-context processing is one of the few remaining computational bottlenecks preventing language models from handling truly long documents cheaply — if you can compress intelligently instead of uniformly, you get better accuracy at lower cost, which changes what becomes economically feasible to deploy.