Unified Hallucination Fuzzing for Multimodal Large Language Models

多模态大语言模型的统一幻觉模糊测试

Abstract: Hallucination remains a persistent challenge for Multimodal Large Language Models (MLLMs), severely limiting their reliability in high-stakes applications. Existing evaluations, predominantly based on static benchmarks, suffer from narrow taxonomical coverage and rapid performance saturation, failing to reflect model robustness in evolving real-world scenarios.

摘要： 幻觉仍然是多模态大语言模型（MLLM）面临的持续挑战，严重限制了其在高风险应用中的可靠性。现有的评估主要基于静态基准测试，存在分类覆盖范围狭窄和性能快速饱和的问题，无法反映模型在不断演变的现实场景中的鲁棒性。

To bridge this gap, we present a systematic evaluation framework integrating a comprehensive benchmark with self-evolving stress testing. First, we introduce UniHall, a fine-grained dataset grounded in a unified taxonomy spanning Object, Instruction, and Knowledge dimensions.

为了弥补这一差距，我们提出了一个系统性的评估框架，将综合基准测试与自我演进的压力测试相结合。首先，我们引入了 UniHall，这是一个基于统一分类法（涵盖对象、指令和知识维度）的细粒度数据集。

Second, to address benchmark saturation, we propose Self-Adaptive Multimodal Fuzzing (SAMF), a self-adaptive framework that employs evolutionary mutation strategies to explore the boundaries of model hallucinations. Crucially, to ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles.

其次，为了解决基准测试饱和的问题，我们提出了自适应多模态模糊测试（SAMF），这是一个采用进化变异策略来探索模型幻觉边界的自适应框架。至关重要的是，为了确保对动态输入的可靠评估，SAMF 结合了一套由多模态预言机集成驱动的结构化指标体系。

Our extensive experiments reveal that state-of-the-art MLLMs exhibit significant performance degradation under fuzzing compared to conventional settings, exposing a dissociation between reasoning capabilities and factual grounding. Furthermore, we identify a helpfulness-hallucination trade-off, where reinforcement learning alignment inadvertently exacerbates sycophancy in instruction-following tasks. The framework, code and benchmark are available at this https URL.

我们的大量实验表明，与传统设置相比，最先进的 MLLM 在模糊测试下表现出显著的性能下降，揭示了推理能力与事实基础之间的脱节。此外，我们发现了一种“有用性与幻觉”之间的权衡，即强化学习对齐在指令遵循任务中无意中加剧了谄媚行为。该框架、代码和基准测试可在以下网址获取：[链接]。