Large Language Models (LLMs) often produce factoid hallucinations - plausible yet incorrect answers. A common mitigation strategy is model alignment, which improves factual accuracy by training on curated (factual, non-factual) pairs. However, this approach often relies on a stronger model (e.g., GPT-4) or an external knowledge base to assess factual correctness that may not always be accessible. Addressing this, we propose Atomic Consistency Preference Optimization (ACPO), a self-supervised preference-tuning method that enhances factual accuracy without external supervision. ACPO leverages atomic consistency signals (i.e., the agreement of individual facts across multiple stochastic responses) to identify high- and low-quality data pairs for model alignment. Despite being fully self-supervised, ACPO outperforms the strong supervised alignment baseline by 1.95 points averaged across Phi-3 and Llama3 on the LongFact and BioGen datasets, demonstrating its effectiveness in improving factual reliability without relying on external models or knowledge bases.

“Extractive Question Answering (EQA) in the few-shot learning scenario is one of the most chal-lenging tasks of Machine Reading Comprehension (MRC). Some previous works employ exter-nal knowledge for data augmentation to improve the performance of few-shot extractive ques-tion answering. However, there are not always available external knowledge or language- anddomain-specific NLP tools to deal with external knowledge such as part-of-speech taggers, syn-tactic parsers, and named-entity recognizers. In this paper, we present a novel Plug-and-PlayData Augmentation Component (PPDAC) for the few-shot extractive question answering, whichincludes a paraphrase generator and a paraphrase selector. Specifically, we generate multipleparaphrases of the question in the (question, passage, answer) triples using the paraphrase gener-ator and then obtain highly similar statements via paraphrase selector to form more training datafor fine-tuning. Extensive experiments on multiple EQA datasets show that our proposed plug-and-play data augmentation component significantly improves question-answering performance,and consistently outperforms state-of-the-art approaches in few-shot settings by a large margin.”