#Context-Compaction

You don’t need long context

The paper argues that frontier models can keep contexts short without losing much accuracy on long-horizon and long-input tasks, and that selective reading plus compaction can replace long contexts.