Unified Hallucination Fuzzing for Multimodal Large Language Models
Authors: Pengfei Zhou, Jiajun Song, Zhiwei Tang, Yixing Ma, Xiaopeng Peng, Donghui Si, Yuhang Xu, Huiqi Song, Yiyuan Miao, Yichen Qian, Weihua Chen, Wangbo Zhao, Bohan Zhuang, Jiasheng Tang, Yang You
Subject: Computer Science > Computation and Language
Submitted: 15 July 2026 (arXiv:2608.07525)
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. 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. 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.
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.
1. Introduction
As multimodal large language models (MLLMs) become increasingly integrated into real-world applications—from autonomous driving to medical diagnostics—the issue of hallucination (i.e., generating plausible but factually incorrect outputs) poses a critical barrier to deployment. While recent advances have greatly improved MLLM capabilities, the evaluation of their robustness against hallucinations has lagged behind. Traditional static benchmarks, though valuable, present several limitations. They often rely on a narrow set of hallucination categories, which fails to capture the diversity of failure modes in dynamic environments. Moreover, as models quickly achieve high scores on these fixed tasks, the benchmarks become saturated, providing little insight into how models will behave in novel or adversarial situations.
To address these shortcomings, we propose a systematic evaluation framework that combines a comprehensive benchmark with self-evolving stress testing. Our approach aims to mimic the unpredictable nature of real-world interactions, pushing models to their limits and uncovering latent weaknesses that static tests miss.
2. UniHall: A Fine-Grained Benchmark
We introduce UniHall, a large-scale dataset designed to cover a unified taxonomy of hallucination types. This taxonomy encompasses three primary dimensions:
- Object: Hallucinations related to the misidentification, omission, or false attribution of objects within an image or video.
- Instruction: Hallucinations arising from misunderstanding or misinterpreting user instructions, leading to inappropriate or irrelevant responses.
- Knowledge: Hallucinations involving factual errors or outdated information that is inconsistent with world knowledge or given context.
Each dimension is further subdivided into fine-grained categories, ensuring comprehensive coverage. UniHall serves as a robust foundation for measuring baseline performance and identifying specific areas of weakness in current MLLMs.
3. Self-Adaptive Multimodal Fuzzing (SAMF)
To overcome the limitations of static evaluation, we propose SAMF, a self-adaptive fuzzing framework that dynamically generates adversarial test cases. SAMF operates through an iterative process of mutation and selection, drawing inspiration from evolutionary algorithms. Key components include:
- Evolutionary Mutation Strategies: SAMF systematically perturbs inputs—altering objects, instructions, or knowledge elements—to generate novel and challenging scenarios. The mutations are guided by a fitness function designed to maximize the likelihood of inducing hallucinations, thereby exploring the boundaries of model competence.
- Ensemble of Multi-Modal Oracles: To reliably assess the correctness of model outputs on dynamic inputs, SAMF employs a panel of oracles that integrate multiple modalities (e.g., visual grounding, semantic consistency, and factual verification). These oracles work in concert to provide a structured metric suite, ensuring that evaluations are both accurate and comprehensive.
4. Experimental Findings
Our experiments, conducted on state-of-the-art MLLMs, yield two major insights:
- Performance Degradation Under Fuzzing: Models experience significant performance drops when subjected to SAMF-generated inputs, compared to their performance on static benchmarks. This result highlights a dissociation between reasoning ability and factual grounding—models may excel at logical reasoning but fail when confronted with subtle factual distortions.
- Helpfulness–Hallucination Trade-Off: We observe that reinforcement learning alignment, while increasing helpfulness, inadvertently exacerbates sycophancy in instruction-following tasks. This trade-off suggests that optimization for helpfulness alone may lead to a higher incidence of hallucinations, as models become more prone to fabricate answers to satisfy user expectations.
5. Conclusion and Future Work
Our unified framework provides a more rigorous and dynamic approach to evaluating MLLM hallucinations, addressing the critical gaps left by static benchmarks. By combining a comprehensive taxonomy with self-evolving fuzzing, we offer a tool for both assessing current models and guiding future improvements. The identified trade-offs between helpfulness and factual accuracy open new avenues for research in alignment strategies. Future efforts will expand UniHall to additional modalities and languages, and further refine SAMF's mutation and oracle mechanisms.
Availability: The full framework, code, and dataset are released at https://github.com/LanceZPF/EvalHall.
Additional Details: The paper spans 47 pages and includes 17 figures, offering a thorough examination of the methodologies and results.
via ArXiv CL+LG