This page organizes papers related to artificial intelligence published around the world. This page is summarized using Google Gemini and is operated on a non-profit basis. The copyright of the paper belongs to the author and the relevant institution. When sharing, simply cite the source.
To address the increased computational cost associated with long input lengths in Retrieval-Augmented Generation (RAG) models, we propose a lossless context compression technique called CORE. Without predefined compression guidelines, CORE optimizes the compression policy end-to-end by utilizing downstream task performance as a feedback signal. Even at a high compression ratio of 3%, it prevents performance degradation and improves the Exact Match (EM) score by 3.3 points.
Takeaways, Limitations
•
Takeaways:
◦
A novel end-to-end learning technique for context compression in RAG models is proposed.
◦
Learning compression policies by leveraging downstream task performance without predefined compression labels.
◦
Achieve performance improvements even at high compression ratios.
•
Limitations:
◦
Code release is planned, but currently not implemented or usable.
◦
Additional information about the specific dataset and experimental results is needed.
◦
Absence of comparative analysis information with other compression techniques.
