arXiv 2509.244912025-09-29cs.CVcs.AI

通过语义课程偏好优化缓解 MLLM 中的视觉幻觉

Mitigating Visual Hallucinations via Semantic Curriculum Preference Optimization in MLLMs

School of Engineering, Westlake University, Hangzhou, China（西湖大学工程学院）；School of Cyberspace Security, Nanjing University of Science and Technology, Nanjing, China（南京理工大学网络安全学院）

Yuanshuai Li, Yuping Yan, Junfeng Tang, Yunxuan Li, Zeqi Zheng, Yaochu Jin

阅读 PDF

arXiv 页面 ↗HTML 原文 ↗

AI总结：

针对 MLLM 视觉幻觉及 DPO 难捕捉细粒度语义、易捷径学习的问题，提出 SCPO 框架，用语义难度课程、动态参考模型和对称双向目标对齐，最高降低幻觉率 62.9%。

摘要：

多模态大语言模型（MLLM）显著提升了多种任务的性能，但仍持续受到视觉幻觉的困扰；这是一个关键问题，指生成的回答与视觉证据相矛盾。尽管 Direct Preference Optimization（DPO，直接偏好优化）被广泛用于对齐，但将其应用于 MLLM 时，往往无法捕捉细粒度语义差异，并会助长捷径学习。为应对这些挑战，我们提出 Semantic Curriculum Preference Optimization（SCPO，语义课程偏好优化），这是一个用于 MLLM 对齐的新颖框架。SCPO 采用由易到难的渐进式课程，该课程建立在我们的 Semantic Curriculum Preference Pairs 数据集之上；该数据集提供按难度排序的细粒度语义对比。此课程通过动态参考模型以及新颖的对称、双向目标进行训练，以促进同时从文本偏好和视觉偏好中学习。据我们所知，SCPO 是首个将语义、对称性和课程统一用于 MLLM 对齐的框架，可有效缓解视觉幻觉。在不同规模和版本的 LLaVA 模型上进行的大量实验验证，SCPO 在多个幻觉基准上相比基线模型表现出更优性能，将幻觉率最高降低 62.9%。此外，在泛化基准上的评估表明，SCPO 在保持通用能力的同时提升了事实性，并且其性能在通用视觉-语言基准上保持稳定。

英文摘要：

Multimodal Large Language Models (MLLMs) have significantly improved the performance of various tasks, but continue to suffer from visual hallucinations, a critical issue where generated responses contradict visual evidence. While Direct Preference Optimization(DPO) is widely used for alignment, its application to MLLMs often fails to capture fine-grained semantic differences and encourages shortcut learning. To address these challenges, we propose Semantic Curriculum Preference Optimization (SCPO), a novel framework for MLLM alignment. SCPO employs a progressive, easy-to-hard curriculum built upon our Semantic Curriculum Preference Pairs dataset, which provides fine-grained semantic contrasts sorted by difficulty. This curriculum is trained with a dynamic reference model and a novel symmetric, bidirectional objective to facilitate simultaneous learning from both textual and visual preferences. To our knowledge, SCPO is the first framework to unify semantics, symmetry, and curriculum for MLLMs alignment, effectively mitigating visual hallucinations. Extensive experiments on LLaVA models across various scales and versions validate that SCPO demonstrates superior performance compared to baseline models on multiple hallucination benchmarks, reducing the hallucination rate by up to 62.9%. Moreover, evaluations on generalized benchmarks show that SCPO improves factuality while preserving general capabilities, with its performance remaining stable across general vision-language benchmarks.

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。