Dehallucinating Parallel Context Extension for Retrieval-Augmented Generation

Zexiong Ma, Shengnan An, Zeqi Lin, Yanzhen Zou, Jian-Guang Lou, Bing Xie

Keywords: retrieval-augmented generation, hallucination, parallel context extension

TL;DR: This work introduces DePaC to alleviate the hallucination issue of parallel context extension on RAG.

Abstract: Large language models (LLMs) are susceptible to generating hallucinated information, despite the integration of retrieval-augmented generation (RAG). Parallel context extension (PCE) is a line of research attempting to effectively integrating parallel (unordered) contexts, while it still suffers from in-context hallucinations when adapted to RAG scenarios. In this paper, we propose DePaC (Dehallucinating Parallel Context Extension), which alleviates the in-context hallucination problem with context-aware negative training and information-calibrated aggregation. DePaC is designed to alleviate two types of in-context hallucination: fact fabrication (i.e., LLMs present claims that are not supported by the contexts) and fact omission (i.e., LLMs fail to present claims that can be supported by the contexts). Specifically, (1) for fact fabrication, we apply the context-aware negative training that fine-tunes the LLMs with negative supervisions, thus explicitly guiding the LLMs to refuse to answer when contexts are not related to questions; (2) for fact omission, we propose the information-calibrated aggregation which prioritizes context windows with higher information increment from their contexts. The experimental results on nine RAG tasks demonstrate that DePaC significantly alleviates the two types of in-context hallucination and consistently achieves better performances on these tasks.

Primary Area: foundation or frontier models, including LLMs

Submission Number: 6771