SGIC:面向 RAG 的自指式迭代校准框架
计算与语言
2025-06-23 v1
摘要
最近的检索增强生成(RAG)研究集中于从候选文档中检索有用信息。然而,许多方法忽略了大型语言模型(LLM)的校准能力,这种能力依赖于其强大的上下文推理能力。本工作表明,为 LLM 提供特定线索可显著提升其校准效果,尤其是在多轮校准中。我们提出了新的 SGIC:自指式迭代校准框架,该框架利用不确定性得分作为工具。首先,本框架计算不确定性得分,以确定每个文档与查询的相关性以及 LLM 生成响应的置信水平。随后,它迭代重新评估这些得分,将其与先前响应融合,以细化校准。 Furthermore, we introduce an innovative approach for constructing an iterative self-calibration training set, which optimizes LLMs to efficiently harness uncertainty scores for capturing critical information and enhancing response accuracy. Our proposed framework significantly improves performance on both closed-source and open-weight LLMs.
引用
@article{arxiv.2506.16172,
title = {SGIC: A Self-Guided Iterative Calibration Framework for RAG},
author = {Guanhua Chen and Yutong Yao and Lidia S. Chao and Xuebo Liu and Derek F. Wong},
journal= {arXiv preprint arXiv:2506.16172},
year = {2025}
}