NVIDIA paper: frequent context compaction preserves accuracy, long-context models may be unnecessary

nvidia · hf · 2026-11-04

NVIDIA's "You don't need long context" paper argues that frequent context compaction preserves task accuracy on long-horizon and long-input tasks.

It compares three regimes: accumulating full history (FHA), compacting at a threshold (EoCC), and semantic state compaction after every action (SAC), where an updatable semantic state replaces the growing conversation.
The findings suggest dedicated long-context LLMs may be unnecessary, motivating semantic state space models instead.

More from coding & agent