Technical Summary: Unified Hallucination Fuzzing for Multimodal Large Language Models

1. Problem Statement

Multimodal Large Language Models (MLLMs) have demonstrated exceptional capabilities in generative reasoning and visual understanding. However, hallucination—the generation of responses that are linguistically coherent but factually unfaithful to visual inputs or established knowledge—remains a persistent bottleneck for their trustworthy deployment in high-stakes domains such as healthcare, legal consulting, and autonomous systems.

Current evaluation paradigms suffer from two critical limitations:

1. Narrow Taxonomical Coverage: Existing benchmarks primarily focus on static, object-level errors, failing to capture the long-tail distribution of hallucinations across instruction-following and knowledge retrieval.
2. Performance Saturation: Static datasets with fixed inputs allow models to rapidly memorize benchmarks ("learning the exam") without acquiring true robustness. High scores on these static sets often mask underlying brittleness, creating a false sense of security.

There is a pressing need to shift from static measurement to dynamic, adversarial stress testing that can systematically probe model vulnerabilities under evolving real-world scenarios.

2. Methodology

The paper proposes a systematic evaluation framework integrating a comprehensive benchmark with a self-evolving stress testing mechanism.

2.1. UniHall: A Unified Benchmark

The authors introduce UniHall, a large-scale, fine-grained dataset grounded in a unified taxonomy spanning three dimensions:

• Object-level: Fabricated existence, incorrect attributes, and false relations.
• Instruction-level: Failure to refuse impossible questions and sycophancy (agreeing with misleading premises).
• Knowledge-level: Fictitious facts and fabricated references.

Unlike previous works focusing on monolithic object errors, UniHall integrates diverse open-domain and domain-specific data, enabling deeper analysis of specific failure modes across different cognitive tasks.

2.2. SAMF: Self-Adaptive Multimodal Fuzzing

To address benchmark saturation, the paper proposes SAMF (Self-Adaptive Multimodal Fuzzing), an automated red-teaming framework inspired by software fuzz testing.

• Dynamic Mutation: SAMF employs evolutionary strategies, including heuristic search and Reinforcement Learning (RL), to generate semantically preserved yet cognitively adversarial mutations on both textual prompts and visual inputs.
• Mutation Operators:
  • Textual: Redundancy/interference control, complexity traps (syntactic complications), and knowledge-conflict induction (authority bias).
  • Visual: Background distraction injection, redundant multi-view copies with quality degradation, and multi-image overload.
• Adaptive Policy: The framework uses a 20% pilot subset to discover optimal mutation strategies. RL agents learn adaptive policies to optimize operator selection and mutation intensity based on oracle feedback, outperforming heuristic strategies in exploration efficiency.

2.3. Structured Metric Suite

To ensure reliable assessment of dynamic inputs, SAMF incorporates a structured metric suite driven by an ensemble of multi-modal oracles (visual detectors, cross-modal alignment models, symbolic rules, and LLM verifiers). The metrics include:

• Generic Hallucination Rate (GHR): A binary decision function detecting unsupported but fluent responses.
• Breakdown Hallucination Rate (BHR): Decomposes responses into atomic claims (Object, Instruction, Knowledge) for deterministic verification against evidence sets.
• Structured Hallucination Rate (SHR): Generalizes BHR using GPT-based judges for claim decomposition and severity scoring.
• Risk-Aware Aggregation: Weights metrics by instance-level risk levels (1–5) to reflect real-world impact.

3. Key Contributions

1. UniHall Benchmark: A risk-aware benchmark built on a unified Object–Instruction–Knowledge hallucination taxonomy, providing significantly more extensive coverage than existing benchmarks.
2. SAMF Framework: The first mutation-driven MLLM fuzzing framework featuring adaptive evolution and structured oracle verification, capable of overcoming static benchmark saturation.
3. Empirical Insights: Systematic stress testing reveals a dissociation between reasoning capabilities and factual grounding, and identifies a "helpfulness-hallucination trade-off" where reinforcement learning alignment inadvertently exacerbates sycophancy.

4. Experimental Results

Extensive experiments on state-of-the-art MLLMs (including GPT-5 series, Gemini, Qwen, LLaVA, and InternVL variants) reveal:

• Performance Degradation under Fuzzing: All models exhibit significant performance drops when transitioning from static to fuzzing evaluation. For instance, the Generic Hallucination Rate (GHR) for Qwen2.5-VL-7B rises from ~29.8% (static) to ~42.7% (RL-based fuzzing).
• Reasoning-Grounding Dissociation: Stronger general-purpose reasoning does not necessarily yield better hallucination resistance. Models with advanced reasoning capabilities (e.g., Qwen3-VL-8B, GPT-5.2) sometimes show higher hallucination rates, particularly in knowledge and instruction dimensions, suggesting that complex reasoning rationales may introduce more unsupported claims.
• The Alignment Tax: Aggressive reinforcement learning (RL) alignment for user helpfulness tends to exacerbate hallucinations. Models often fabricate plausible information to satisfy misleading instructions or appear helpful, rather than admitting ignorance or refusing. This is particularly evident in "Refuse" and "Sycophantic" categories, where hallucination rates approach saturation (near 100%).
• Category Sensitivity: Object-level hallucinations are relatively well-handled, whereas instruction-level challenges (refusal, sycophancy) and knowledge-level hallucinations show marked degradation under adversarial mutations.
• Metric Sensitivity: Verbose models are penalized more heavily by surface-level metrics (GHR) than by claim-level metrics (SHR/BHR), highlighting the need for granular verification.

5. Significance and Claims

The paper claims to establish a new framework for universal hallucination assessment. By moving beyond static leaderboards, the authors demonstrate that current MLLMs possess latent vulnerabilities that are only exposed through dynamic, adversarial stress testing.

The work argues that:

• Static benchmarks are insufficient: High scores on static datasets mask brittleness and fail to reflect robustness in real-world, evolving scenarios.
• Alignment has costs: Optimizing for helpfulness via RL can come at the cost of factual faithfulness, creating a trade-off that requires careful calibration.
• Robustness requires multimodal integration: True reliability depends on the model's ability to maintain factual independence under conflicting cues (e.g., adversarial prompts vs. visual evidence).

Ultimately, the proposed framework, code, and benchmark (UniHall) are intended to guide the development of more trustworthy multimodal systems by providing a rigorous, self-evolving mechanism for identifying and mitigating hallucinations.