Trustworthy Alignment of Retrieval-Augmented Large Language Models via Reinforcement Learning
This work addresses trustworthiness issues in retrieval-augmented language models for real-world applications, representing an incremental advancement by applying alignment techniques from human preference to evidence-based responses.
The paper tackles the problem of hallucinations in retrieval-augmented large language models by addressing conflicts between contextual and parametric knowledge, proposing a reinforcement learning algorithm that aligns models to rely solely on external evidence without explicit supervision, achieving improved trustworthiness as demonstrated theoretically and experimentally.
Trustworthiness is an essential prerequisite for the real-world application of large language models. In this paper, we focus on the trustworthiness of language models with respect to retrieval augmentation. Despite being supported with external evidence, retrieval-augmented generation still suffers from hallucinations, one primary cause of which is the conflict between contextual and parametric knowledge. We deem that retrieval-augmented language models have the inherent capabilities of supplying response according to both contextual and parametric knowledge. Inspired by aligning language models with human preference, we take the first step towards aligning retrieval-augmented language models to a status where it responds relying merely on the external evidence and disregards the interference of parametric knowledge. Specifically, we propose a reinforcement learning based algorithm Trustworthy-Alignment, theoretically and experimentally demonstrating large language models' capability of reaching a trustworthy status without explicit supervision on how to respond. Our work highlights the potential of large language models on exploring its intrinsic abilities by its own and expands the application scenarios of alignment from fulfilling human preference to creating trustworthy agents.