Exploring the Generalizability of Factual Hallucination Mitigation via Enhancing Precise Knowledge Utilization

TL;DR

This work addresses factual hallucinations in LLMs by focusing on precise knowledge utilization rather than injecting external data. It introduces PKUE, a self-aligned, direct-preference fine-tuning approach that uses self-generated, precise QA data and a large, multi-domain Chinese dataset, FactualBench, to improve factuality and general abilities. Through extensive experiments across multilingual benchmarks, PKUE demonstrates consistent improvement in factual tasks, general skills, and cross-lingual transfer, achieving up to several points of Avg gains and outperforming existing post-training baselines. The findings support the idea that sharpening knowledge utilization for simple, precise QA can drive broad, generalizable enhancements in LLM behavior with practical implications for safer, more reliable AI systems.

Abstract

Large Language Models (LLMs) often struggle to align their responses with objective facts, resulting in the issue of factual hallucinations, which can be difficult to detect and mislead users without relevant knowledge. Although post-training techniques have been employed to mitigate the issue, existing methods usually suffer from poor generalization and trade-offs in other different capabilities. In this paper, we propose to address these by directly augmenting LLM's fundamental ability to precisely leverage its knowledge and introduce PKUE (Precise Knowledge Utilization Enhancement), which fine-tunes the model on self-generated responses to precise and simple factual questions through preference optimization. Furthermore, we construct FactualBench, a comprehensive and precise factual QA dataset containing 181k Chinese data spanning 21 domains, to facilitate both evaluation and training. Extensive experiments demonstrate that PKUE significantly improves LLM overall performance, with consistent enhancement across factual tasks of various forms, general tasks beyond factuality, and tasks in different language.

Figures (17)

• Figure 1: Previous methods on factual hallucination mitigation exhibit poor generalizability across different factual tasks and suffer from degradations on comprehensive abilities and helpfulness, while our method PKUE improves model performance on all seven benchmarks, with a significant advantage on Avg.
• Figure 2: The framework of our work. Left: We first extract factual knowledge from the Internet encyclopedia and construct a large and comprehensive dataset, FactualBench. Several filtering strategies are adopted for higher quality. Right: Next, we align LLM on self-generated response pairs on FactualBench. We elicit diverse responses to the same question, verify each correctness comparing to the standard answer, and sample preference pairs for training.
• Figure 3: A comparison between Baichuan1 accuracy in low temp. BO1 and high temp. BO8. Significant gaps in all domains demonstrate the potential of the model. Each domain is represented by its first five letters.
• Figure 4: Changes of Baichuan1 alignment with Qwen2-72B-Instruct on four benchmarks after training.
• Figure 5: Baichuan1 performance improvement after DPO increases at a logarithmic rate with the training data size expanding.