arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2510.00880cs.CL
收藏

HalluGuard:基于证据支撑的小型推理模型用于缓解检索增强生成中的幻觉问题

HalluGuard: Evidence-Grounded Small Reasoning Models to Mitigate Hallucinations in Retrieval-Augmented Generation

Banque de Luxembourg(卢森堡银行);Center for Data Science in Humanities, Chosun University(人文数据科学中心,全州大学);ATILF, University of Lorraine–CNRS(ATILF,洛林大学–CNRS);SnT, University of Luxembourg(SnT,卢森堡大学)

Loris Bergeron, Ioana Buhnila, Jérôme François, Radu State

AI总结:

本研究提出4B参数量的小型推理模型HalluGuard,通过合成数据集、偏好微调等技术缓解RAG中的幻觉,在基准测试中性能媲美更大参数量的模型,将以开源协议发布。

摘要:

大语言模型(LLM)在众多自然语言处理(NLP)任务中表现出色,但仍容易产生幻觉,限制了其在实际应用中的可信度。我们提出HalluGuard,这是一个参数量为4B的小型推理模型(SRM),用于缓解检索增强生成(RAG)中的幻觉问题。HalluGuard会将文档-声明对分类为有依据或幻觉,并生成基于证据支撑的理由以保证透明度。我们的方法结合了三部分内容:一是基于FineWeb生成、经多阶段筛选与数据重构优化的跨领域合成数据集;二是合成的有依据声明与幻觉声明;三是采用比值偏好优化(Odds Ratio Preference Optimization)的偏好微调,将大模型的推理能力蒸馏到更小的骨干模型中。在LLM-AggreFact基准的RAGTruth子集上,HalluGuard取得了84.0%的平衡准确率(BAcc),与专用模型MiniCheck(7B;84.0%)和Granite Guardian 3.3(8B;82.2%)相当,而参数量仅约为它们的一半。在完整基准上它达到了75.7%的平衡准确率,与GPT-4o(75.9%)等更大的通用大语言模型持平。我们将在论文录用后以Apache 2.0许可证发布HalluGuard及相关数据集。

英文摘要:

Large Language Models (LLMs) excel in many NLP tasks but remain prone to hallucinations, limiting trust in real-world applications. We present HalluGuard, a 4B-parameter Small Reasoning Model (SRM) for mitigating hallucinations in Retrieval-Augmented Generation (RAG). HalluGuard classifies document-claim pairs as grounded or hallucinated and produces evidence-grounded justifications for transparency. Our approach combines (i) a domain-agnostic synthetic dataset derived from FineWeb and refined through multi-stage curation and data reformation, (ii) synthetic grounded and hallucinated claims, and (iii) preference-based fine-tuning with Odds Ratio Preference Optimization to distill large-model reasoning into a smaller backbone. On the RAGTruth subset of the LLM-AggreFact benchmark, HalluGuard achieves 84.0% balanced accuracy (BAcc), rivaling specialized models, MiniCheck (7B; 84.0%) and Granite Guardian 3.3 (8B; 82.2%) while using roughly half their parameters. Over the full benchmark it reaches 75.7% BAcc, matching larger general-purpose LLMs such as GPT-4o (75.9%). We will release HalluGuard and datasets under Apache 2.0 upon acceptance.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑