Towards Faithful and Robust LLM Specialists for Evidence-Based Question-Answering

TL;DR

The paper tackles faithfulness in Evidence-Based QA by formalizing source-citation requirements and introducing a scalable data-generation pipeline (SynSciQA) with automated quality filters. It defines three quality dimensions—source quality, format quality, and answer attributability—and develops metrics to quantify them, then demonstrates that data quality surpasses quantity in fine-tuning open-source LLMs. Through SynSciQA and four evaluation sets, the authors show synthetic data can improve both in-domain and out-of-domain performance, while attributability scores align well with human and GPT-4 judgments. The work also provides a pathway for validating OOD performance via synthetic development data and discusses limitations and future directions toward production-ready, faithful RAG systems.

Abstract

Advances towards more faithful and traceable answers of Large Language Models (LLMs) are crucial for various research and practical endeavors. One avenue in reaching this goal is basing the answers on reliable sources. However, this Evidence-Based QA has proven to work insufficiently with LLMs in terms of citing the correct sources (source quality) and truthfully representing the information within sources (answer attributability). In this work, we systematically investigate how to robustly fine-tune LLMs for better source quality and answer attributability. Specifically, we introduce a data generation pipeline with automated data quality filters, which can synthesize diversified high-quality training and testing data at scale. We further introduce four test sets to benchmark the robustness of fine-tuned specialist models. Extensive evaluation shows that fine-tuning on synthetic data improves performance on both in- and out-of-distribution. Furthermore, we show that data quality, which can be drastically improved by proposed quality filters, matters more than quantity in improving Evidence-Based QA.

Figures (12)

• Figure 1: Synthetic data generation pipeline and Evaluation for Evidence-Based QA.
• Figure 2: Quality Dimensions of Evidence-Based Question Answering.
• Figure 3: Evaluation Dataset's orientation towards real-world use case scenarios vs. their distribution's proximity to the trainsets.
• Figure 4: Controlling quantity, Source Quality scores vs. number of epoch, caused by different quality.
• Figure 5: Controlling quantity, Attributability scores vs. number of epoch, caused by different quality.