Measuring Epistemic Humility in Multimodal Large Language Models

Abstract: Hallucinations in multimodal LLMs (MLLMs) -- where the model generates content inconsistent with the input image -- pose significant risks in real-world applications, from misinformation in visual question answering to unsafe errors in decision-making. Existing benchmarks primarily test recognition accuracy, i.e., evaluating whether models can select the correct answer among distractors. This overlooks an equally critical capability for trustworthy AI: recognizing when none of the provided options are correct, a behavior reflecting epistemic humility. We present HumbleBench, a new hallucination benchmark designed to evaluate MLLMs' ability to reject plausible but incorrect answers across three hallucination types: object, relation, and attribute. Built from a panoptic scene graph dataset, we leverage fine-grained scene graph annotations to extract ground-truth entities and relations, and prompt GPT-4-Turbo to generate multiple-choice questions, followed by a rigorous manual filtering process. Each question includes a "None of the above" option, requiring models not only to recognize correct visual information but also to identify when no provided answer is valid. We evaluate a variety of state-of-the-art MLLMs -- including both general-purpose and specialized reasoning models -- on HumbleBench and share valuable findings and insights with the community. By incorporating explicit false-option rejection, HumbleBench fills a key gap in current evaluation suites, providing a more realistic measure of MLLM reliability in safety-critical settings. Our code and dataset are released publicly and can be accessed at https://github.com/maifoundations/HumbleBench.

• The paper introduces HumbleBench, a benchmark that measures epistemic humility by requiring MLLMs to correctly reject false options, including 'None of the above'.
• It employs a robust dataset constructed from panoptic segmentation and GPT-4-Turbo generated questions, refined through extensive manual curation.
• Results reveal that most MLLMs struggle with uncertainty, highlighting the need for architectural and data-driven innovations to reduce hallucinations.

Measuring Epistemic Humility in Multimodal LLMs

Introduction

This paper introduces HumbleBench, a discriminative benchmark specifically constructed for evaluating epistemic humility in multimodal LLMs (MLLMs) (2509.09658). Unlike conventional hallucination benchmarks, which focus solely on factual recall and recognition accuracy, HumbleBench explicitly measures models’ capability for false-option rejection—that is, their ability to identify when none of the provided answers are valid. The inclusion of a "None of the above" (NOTA) option in every multiple-choice question directly targets whether an MLLM can abstain from overconfident hallucinations and demonstrates appropriate uncertainty, an emergent property critical for robust, trustworthy AI systems.

Figure 1: Examples from HumbleBench; the correct (including NOTA) answers are marked green, illustrating the challenge of explicit false-option rejection.

HumbleBench Benchmark Design

Dataset Construction

HumbleBench is built upon the Panoptic Scene Graph (PSG) dataset, leveraging its pixel-level segmentation and scene graph annotations to generate fine-grained and verifiable object, relation, and attribute data. For each of 4,500 sampled images, the pipeline extracts:

• Objects and relations: Direct from PSG scene graphs.
• Attributes: Automatically extracted using InstructBLIP, which generates concise visual descriptions per object.
• Questions and distractors: GPT-4-Turbo synthesizes multiple-choice questions (five options, with the last being NOTA), guided by rigorous prompts and logical constraints to maximize challenge and minimize ambiguity.

A custom PyQt5 GUI enables high-throughput, robust manual curation, ultimately filtering 41,843 initial questions down to 22,831 high-quality items, making HumbleBench the largest discriminative hallucination benchmark to date.

Figure 2: The HumanBench construction pipeline, combining high-precision panoptic annotation, automated question/option generation with GPT-4-Turbo, and manual curation.

Task and Diversity

Each question in HumbleBench requires the MLLM to select the most accurate answer from five candidates, with the NOTA option present in every case. The benchmark covers three hallucination types (object, relation, attribute) in balanced proportions and ensures answer distributions avoid bias toward any particular choice.

Figure 3: Distribution over answer choices (left) and question types (right), underscoring the benchmark’s structural balance.

Comparison with Existing Benchmarks

A key differentiator lies in HumbleBench’s epistemic humility assessment: prior datasets either omit the NOTA option or assume a valid answer is always present, thereby implicitly rewarding models for guessing rather than acknowledging uncertainty.

Evaluation of State-of-the-Art MLLMs

Model Selection

Nineteen MLLMs are evaluated, with diversity across general-purpose and reasoning-specialized architectures (e.g., Qwen2.5-VL, InternVL3, LLaVA-Next, GLM-4.1V-Thinking, R1-Onevision, Visionary-R1). Both scaling strategies and reasoning-oriented finetuning are represented to assess robustness under the HumbleBench regime.

Metrics and Settings

Three main conditions are used:

1. HumbleBench: Standard benchmark (single correct or NOTA as correct).
2. HumbleBench-E: Only NOTA is correct (all others removed/invalid).
3. HumbleBench-GN: All images replaced by pure Gaussian noise (only NOTA can be correct; visual cues are unavailable).

Main Results

• Recognition vs. Humility: The leading model (GLM-4.1V-Thinking) reaches only 73.46% overall accuracy, with most high-performers plateauing in the low 70% range—well above random guess (20%), but markedly short of high-reliability operation.
• NOTA selection is rarely performed correctly; in HumbleBench-E, nearly all models collapse, frequently scoring at or below random chance—even top performers like GLM-4.1V-Thinking score near zero, revealing a severe deficiency in epistemic humility.
• Scaling Law Violation: There is no monotonic relationship between model size and robustness; architecture and training data quality dominate over parameter count.
• Reasoning Fine-tuning: No universal improvement—several reasoning-specialized models underperform their base models, indicating that post-hoc reasoning signals do not guarantee hallucination avoidance.

  Figure 4: HumbleBench-E results; nearly all SOTA models perform below random guessing on NOTA-only items, revealing consistent failure to exhibit epistemic humility.

Robustness: Visual Faithfulness Under Noise

HumbleBench-GN critically tests visual grounding. With all input images replaced by Gaussian noise, the optimal behavior is to always select NOTA. Many models demonstrate catastrophic performance, defaulting to linguistic priors and hallucinating plausible-sounding answers disconnected from visual evidence.

• Some models (e.g., Qwen2.5-VL) approach high accuracy (over 90%) in the pure-noise regime, but others (e.g., Phi-4, InternVL3) have surprisingly low accuracies, indicating over-reliance on language biases.
• Performance is uncorrelated with HumbleBench standard accuracy, further emphasizing the nontrivial gap between recognizing correct visual content and robustly rejecting hallucinations.

  Figure 5: Correlation between HumbleBench accuracy and HumbleBench-GN (noise); robust visual reasoning models would appear in the top-right, yet many high performers in the main task fail in the noise regime.

Qualitative Error Analysis

Major failure modes identified include:

• Inability to select NOTA: Persistent selection of plausible but false options, even when NOTA is correct.
• Relation and attribute hallucination: Hallucinated spatial relationships, object existence, attribute identification, and simple counting errors remain frequent.
• Noise-driven fabrication: In noise-only scenarios, models often hallucinate based on language priors (e.g., assigning a color to a "sky" in a non-informative noise image).

  Figure 6: Error analysis exemplifying failures to select NOTA (hallucinated object, relation, or attribute), counting errors, and noise-induced hallucination, where answers are untethered from visual evidence.

Implications and Future Directions

The HumbleBench findings indicate that MLLMs exhibit limited epistemic humility: current training pipelines heavily bias toward providing an answer, even in the absence of evidence, and scaling or naïve reasoning enhancement strategies do not reliably support robust false-option rejection. Thus, reliable deployment of MLLMs in safety-critical environments is not attainable with recognitional accuracy alone as an evaluation criterion.

Practical implications include:

• Benchmarking and Evaluation: Future benchmarks must explicitly include abstention or uncertainty options to drive progress on reliable AI decision-making.
• Training Strategies: Supervised and reinforcement learning pipelines need to incorporate explicit negative samples—scenarios where abstention is optimal.
• Alignment Objective: Approaches that bind visual evidence to linguistic output, penalizing confident hallucinations and rewarding appropriate abstention, should become the norm in robust MLLM training methodologies.
• Model Selection: Parameter scaling is insufficient; architectural and data-centric innovations are required to address visual faithfulness and uncertainty modeling.

Conclusion

HumbleBench sets a new standard in hallucination evaluation by directly measuring epistemic humility. The results underscore that recognition accuracy, while necessary, is not sufficient for trustworthy multimodal reasoning. The persistent inability of SOTA MLLMs to abstain in the face of uncertainty highlights the need for new training objectives, more incisive benchmarks, and a paradigm shift towards AI systems that know when not to answer.