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.
Resisting Contextual Interference in RAG via Parametric-Knowledge Reinforcement
Created by
Haebom
Author
Chenyu Lin, Yilin Wen, Du Su, Hexiang Tan, Fei Sun, Muhan Chen, Chenfu Bao, Zhonghou Lyu
Knowledgeable-R1: A Reinforcement Learning-Based RAG Framework for Contextual Interference Resistance
Outline
Knowledgeable-R1 is a reinforcement learning-based Retrieval-Augmented Generation (RAG) framework proposed to improve the performance of knowledge-intensive tasks. It addresses the problem of models relying on inaccurate information and introducing errors due to incorrect or irrelevant text. Knowledgeable-R1 leverages parametric knowledge (PK) to resist contextual interference and explicitly trains a large-scale language model to leverage external context when it is reliable. The framework uses a joint sampling approach to generate response pairs based on retrieval and non-retrieval, and learns both local and global advantages to quantify when misleading context should be ignored or adopted. It also employs an asymmetric advantage transformation to amplify parametrically oriented exploration behavior.
Takeaways, Limitations
•
Takeaways:
◦
Improves robustness and inference accuracy in knowledge conflict situations.
◦
23% performance improvement over existing state-of-the-art (SOTA) in counterfactual scenarios.
◦
No performance degradation when the retrieved context is accurate.
