PretrainRL: Alleviating Factuality Hallucination of Large Language Models at the Beginning
Langming Liu, Kangtao Lv, Haibin Chen, Weidong Zhang, Yejing Wang, Shilei Liu, Xin Tong, Yujin Yuan, Yongwei Wang, Wenbo Su, Bo Zheng
TL;DR
This work identifies factual hallucination in LLMs as a consequence of imbalanced pretraining data that overemphasizes high-probability falsehoods while underrepresenting low-probability truths. It proposes PretrainRL, a pretraining-time reinforcement learning framework that first debiases the distribution and then trains factual knowledge using Direct Preference Optimization (DPO) with an efficient negative-sampling strategy to discover head knowledge. By combining DPO with a stabilizing CT loss and evaluating with beam-based metrics, PretrainRL achieves substantial reductions in hallucinations across POPQA, Wikidata-knowledge infusion, and EntityQuestions, scaling across model sizes and preserving downstream capabilities. This pretrain-and-align paradigm demonstrates that consolidating factual knowledge during pretraining can yield robust, generalizable improvements in factual reliability for LLMs.
Abstract
Large language models (LLMs), despite their powerful capabilities, suffer from factual hallucinations where they generate verifiable falsehoods. We identify a root of this issue: the imbalanced data distribution in the pretraining corpus, which leads to a state of "low-probability truth" and "high-probability falsehood". Recent approaches, such as teaching models to say "I don't know" or post-hoc knowledge editing, either evade the problem or face catastrophic forgetting. To address this issue from its root, we propose \textbf{PretrainRL}, a novel framework that integrates reinforcement learning into the pretraining phase to consolidate factual knowledge. The core principle of PretrainRL is "\textbf{debiasing then learning}." It actively reshapes the model's probability distribution by down-weighting high-probability falsehoods, thereby making "room" for low-probability truths to be learned effectively. To enable this, we design an efficient negative sampling strategy to discover these high-probability falsehoods and introduce novel metrics to evaluate the model's probabilistic state concerning factual knowledge. Extensive experiments on three public benchmarks demonstrate that PretrainRL significantly alleviates factual hallucinations and outperforms state-of-the-art methods.