Context Compaction Reduces Performance for Future Queries
October 21, 2026
Research shows that while context compaction improves accuracy for immediate questions, methods selecting tokens for the current query fail to outperform random deletion for subsequent questions. Word-level deletion via LLMLingua-2 proved more effective at maintaining long-term context than sentence-level deletion.
HOW THIS AFFECTS YOU
●
builderBe cautious when using context compaction for long-horizon agents, as current methods may degrade performance on later queries.