The authors compared query-aware sentence selection (using typed decision models to keep sentences needed for the known question) with random deletion and word-level compressors like LLMLingua-2.
Compaction tuned for one query fails to aid later queries
Word-level compression methods consistently outperform sentence selection for questions asked after compaction, though the ranking depends on the scenario.
Independent
Junwei Yao · Yongxi Zhou · Zike Yang · Zhijun Wang · Jianan Liu · Yanxian Chen · +1 more
Research Digest··2 min read
The authors evaluate context compaction methods on later, unseen questions across four multi-query corpora.
§