Measuring and Enhancing Trustworthiness of LLMs in RAG through Grounded Attributions and Learning to Refuse

2024-09-17

arXiv.org

AI Summary

To address the low trustworthiness of large language models (LLMs) in retrieval-augmented generation (RAG) systems, this paper proposes Trust-Score—a quantitative framework for multi-dimensional trust assessment—and Trust-Align—a lightweight, fine-tuning-free alignment method. Trust-Score jointly models factual citation accuracy, refusal capability, and attribution grounding to holistically quantify LLM trustworthiness. Trust-Align enhances robustness in unanswerable question detection and evidence attribution through learned refusal mechanisms, multi-task prompt alignment, and cross-model generalization adaptation. Evaluated on ASQA, QAMPARI, and ELI5 benchmarks, our approach outperforms 26 out of 27 open-source models—including a +12.56% improvement for LLaMA-3-8B on ASQA—and efficiently adapts to models ranging from 1B to 8B parameters. To our knowledge, this is the first work to establish an interpretable, generalizable, and fine-tuning-free paradigm for enhancing LLM trustworthiness in RAG settings.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Data Mining & Knowledge Management: Representing, Reasoning, and Using Provenance, Trust

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

LLMs are an integral component of retrieval-augmented generation (RAG) systems. While many studies focus on evaluating the overall quality of end-to-end RAG systems, there is a gap in understanding the appropriateness of LLMs for the RAG task. To address this, we introduce Trust-Score, a holistic metric that evaluates the trustworthiness of LLMs within the RAG framework. Our results show that various prompting methods, such as in-context learning, fail to effectively adapt LLMs to the RAG task as measured by Trust-Score. Consequently, we propose Trust-Align, a method to align LLMs for improved Trust-Score performance. 26 out of 27 models aligned using Trust-Align substantially outperform competitive baselines on ASQA, QAMPARI, and ELI5. Specifically, in LLaMA-3-8b, Trust-Align outperforms FRONT on ASQA (up 12.56), QAMPARI (up 36.04), and ELI5 (up 17.69). Trust-Align also significantly enhances models' ability to correctly refuse and provide quality citations. We also demonstrate the effectiveness of Trust-Align across different open-weight models, including the LLaMA series (1b to 8b), Qwen-2.5 series (0.5b to 7b), and Phi3.5 (3.8b). We release our code at https://github.com/declare-lab/trust-align.

Problem

• Evaluating LLM trustworthiness in RAG systems.
• Improving LLM alignment for better task performance.
• Enhancing LLM ability to refuse and cite correctly.

Innovation

• Introduces Trust-Score for LLM trustworthiness evaluation.
• Proposes Trust-Align to enhance LLM performance in RAG.
• Demonstrates Trust-Align effectiveness across various open-weight models.