Measuring Epistemic Humility in Multimodal Large Language Models

Bingkui Tong; Jiaer Xia; Sifeng Shang; Kaiyang Zhou

cs.CVSep 11, 2025

Measuring Epistemic Humility in Multimodal Large Language Models

Authors: Bingkui Tong, Jiaer Xia, Sifeng Shang, Kaiyang Zhou

Organizations: Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates. · Hong Kong Baptist University, Hong Kong.

Abstract

Hallucinations in multimodal large language models (MLLMs), where a model generates content inconsistent with the input image, pose significant risks in real-world applications. Existing hallucination benchmarks primarily test recognition, i.e., whether a model can select the correct answer among distractors. They overlook another capability that is essential for trustworthy AI: recognizing when none of the provided options is supported by the image and declining to commit to a false choice. We present HumbleBench, a hallucination benchmark that evaluates this false-option rejection behavior, a directly observable facet of epistemic humility, through multiple-choice questions that always include a "None of the above" (NOTA) option. HumbleBench is built from the Panoptic Scene Graph dataset: its object and relation annotations, together with candidate attribute cues, are used to prompt GPT-4-Turbo to generate questions, which are then manually verified, yielding 22,831 questions on object, relation, and attribute hallucinations. We evaluate 27 MLLMs, including general-purpose, reasoning-oriented, and proprietary models, and complement standard accuracy with NOTA-aware metrics, diagnostic stress tests, and a human evaluation. Even the strongest models select NOTA for only about half of the questions in which it is correct, cautious prompting shifts models toward NOTA without improving their ability to detect unsupported options, and humans reject unsupported options considerably more reliably than all evaluated models. These results show that recognition accuracy alone is insufficient for assessing hallucination robustness. Our code and dataset are publicly available at https://github.com/maifoundations/HumbleBench.

Figures & tables

Figure 1: Example questions from HumbleBench for object, relation, and attribute hallucinations, with correct options marked in green. Every question includes a “None of the above” option, so a model must also recognize when none of the provided answers is supported by the image.

Figure 2: Distributions of (a) the correct answer choices and (b) the question types in HumbleBench. NOTA denotes the “None of the above” option.

Table 3

Figure 3: Construction pipeline of HumbleBench. For each image, object and relation information is extracted from the panoptic scene graph annotations, and candidate attribute cues are generated by InstructBLIP from cropped objects. GPT-4-Turbo combines this information to generate multiple-choice questions with five options, the last of which is “None of the above”. Every question is then manually verified.

Metric	Definition
Overall accuracy	Accuracy over all samples.
Non-NOTA accuracy	Accuracy restricted to samples where $y_{i}\neq n_{i}$ .
NOTA hit rate	$P(\hat{y}_{i}=n_{i}\mid y_{i}=n_{i})$ , i.e., recall on true NOTA samples.
False-NOTA rate	$P(\hat{y}_{i}=n_{i}\mid y_{i}\neq n_{i})$ , i.e., the rate of incorrectly abstaining on non-NOTA samples.
NOTA selection rate	$P(\hat{y}_{i}=n_{i})$ over all samples.
Humility score	Balanced accuracy for NOTA detection: $\frac{1}{2}\left[\mathrm{NOTA\ hit}+(1-\mathrm{False\text{-}NOTA})\right]$ .

Table 3: Evaluation metrics used in the experiments. For sample

i

y_{i}

is the ground-truth label,

\hat{y}_{i}

is the predicted label, and

n_{i}

is the option label whose text corresponds to “None of the above”.

General-Purpose Models
Model	Object	Relation	Attribute	Overall	Non-NOTA	NOTA hit	False-NOTA	HS
Cambrian	49.39	53.17	63.31	55.56	53.90	65.32	28.64	68.34
Gemma-3	50.71	56.39	70.19	59.48	66.18	20.05	2.90	58.58
DeepSeek-VL2	58.36	57.31	69.54	61.97	70.53	11.68	1.30	55.19
VILA1.5	51.99	62.47	72.37	62.66	71.11	13.00	0.57	56.22
LLaVA-Next	61.03	61.61	72.53	65.29	73.78	15.44	1.52	56.96

Table 4: Results in the Original setting. All values are percentages, and HS denotes the humility score (Table 3 ). Bold marks the best result within each model group.

Setting	# Models	Avg. Overall	Avg. HS	Avg. Macro-F1	Avg. NOTA Sel.	Best Model(s)
Original	27	67.19	62.66	63.86	8.53	Gemini-3-Flash (77.26) / Gemini-2.5-Pro (73.75)
Shuffled NOTA	20	62.62	59.55	62.83	7.89	GLM-4.1V (72.85) / Cambrian (69.66)
Cautious prompt	20	60.63	63.14	58.79	14.97	Qwen2.5-VL (71.40) / GLM-4.1V (73.50)
Noise image	20	27.58	54.15	25.92	54.14	Qwen2.5-VL (NOTA Sel. 90.52)
Noise + grounding	20	22.53	52.29	18.59	66.45	Cambrian (NOTA Sel. 97.86)
NOTA-only	20	26.61	–	8.05	26.59	Cambrian (60.67) / –

Table 5: Summary of the evaluation settings. Values are percentages averaged over the evaluated models. The last column lists the models with the highest overall accuracy and HS; for the two noise settings, in which the original labels are not supported by the image, it lists the model with the highest NOTA selection rate. HS is undefined for NOTA-only, which contains no non-NOTA questions.

Model	Original	Shuffle	Cautious	Noise NOTA sel.	Noise+G NOTA sel.	NOTA-only hit
Cambrian	55.56	56.56	44.79	82.32	97.86	60.67
Gemma-3	59.48	58.13	59.29	78.88	68.89	16.61
DeepSeek-VL2	61.97	36.02	35.32	71.04	5.69	15.41
VILA1.5	62.66	63.89	60.79	12.74	8.50	12.75
LLaVA-Next	65.29	65.00	63.66	67.42	92.45	13.51
LLaMA-3.2	65.31	65.26	65.45	19.00	18.68	8.90

Table 6: Model-level results across evaluation settings (percentages). The Original, Shuffle, and Cautious columns report overall accuracy; the two noise columns report the NOTA selection rate, for which higher is better; the last column reports the NOTA hit rate in the NOTA-only setting. The large-scale and proprietary models are evaluated in the Original setting (Table 4 ).

	Normal questions					Stress questions
Model	Overall	Non-NOTA	NOTA hit	False-NOTA	HS	NOTA hit
Human (mean of 5)	68.00	68.10	67.50	7.62	79.94	62.00
Gemini-3-Flash	72.80	76.67	52.50	3.33	74.58	–
Gemini-2.5-Pro	75.60	81.43	45.00	2.38	71.31	–
Qwen3-VL-235B	74.00	80.48	40.00	3.33	68.33	–
Claude-Haiku-5.5	70.80	74.76	50.00	6.67	71.67	–

Table 7: Human and model performance on the same questions (percentages). The normal columns are computed on the original HumbleBench questions answered by the annotators (50 per annotator, 8 of which have NOTA as the correct answer). The last column reports the NOTA hit rate on the stress questions; for models, it is computed from the NOTA-only predictions on the same questions, which are available for the open-source models. All values are averaged over the five annotators’ question sets.

Figure 4: Errors made by Qwen2.5-VL on HumbleBench. Correct options are marked in green with a check mark and the model’s predictions in red with a cross. Questions 1–4 show failures to select “None of the above”, where the model chooses an object, attribute, or relation that is not present in the image. Question 5 shows a counting error. Question 6, from the noise-image setting, shows the model answering from language priors when the image contains no visual information.

Benchmark	Venue	Size	Hallucination Types			Questions	Reject Option
			Object	Relation	Attribute
CHAIR	EMNLP’18	$\sim$ 5k	✓	✗	✗	yes/no	✗
POPE	EMNLP’23	$\sim$ 3k	✓	✗	✗	yes/no	✗
HallusionBench	CVPR’24	254	✓	✓	✓	yes/no	✗
Hallu-PI	MM’24	$\sim$ 8k	✓	✓	✓	yes/no	✗
Reefknot	ACL’25	$\sim$ 17k	✗	✓	✗	mixed	✗

Table 8: Comparison of HumbleBench with existing hallucination benchmarks. HumbleBench is large-scale and explicitly includes a reject option (“None of the above”) to evaluate false-option rejection in a forced-choice setting. mixed means a combination of yes/no, multiple-choice, and open-ended questions.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 5: Prompt used to generate the candidate questions with GPT-4-Turbo. It specifies the input information, the constraints on the questions and options, and the output format.

Figure 6: Interface of the filtering tool. Each image-question pair is shown with controls to keep, modify, or delete the candidate.

	Object	Attribute	Relation	Total
Candidates	12,216	14,507	15,120	41,843
Kept	41.47%	46.39%	43.04%	43.74%
Modified	17.67%	9.30%	6.75%	10.82%
Deleted	40.86%	44.31%	50.21%	45.44%
Final	7,224	8,079	7,528	22,831
Final (%)	59.14%	55.69%	49.79%	54.56%

Appendix

Table 9: Outcome of manual filtering by question type. Percentages are relative to the number of candidates of each type.

General-Purpose Models
Model	NOTA Selection	NOTA Hit	False-NOTA	Non-NOTA Acc.	$d^{\prime}$
Cambrian	33.98 $\rightarrow$ 56.58	65.32 $\rightarrow$ 82.57	28.64 $\rightarrow$ 52.17	53.90 $\rightarrow$ 38.34	0.96 $\rightarrow$ 0.88
Gemma-3	5.39 $\rightarrow$ 20.57	20.05 $\rightarrow$ 49.52	2.90 $\rightarrow$ 15.65	66.18 $\rightarrow$ 60.95	1.06 $\rightarrow$ 1.00
DeepSeek-VL2	2.81 $\rightarrow$ 12.22	11.68 $\rightarrow$ 20.08	1.30 $\rightarrow$ 10.88	70.53 $\rightarrow$ 37.92	1.04 $\rightarrow$ 0.39
VILA1.5	2.38 $\rightarrow$ 2.92	13.00 $\rightarrow$ 13.25	0.57 $\rightarrow$ 1.16	71.11 $\rightarrow$ 68.88	1.40 $\rightarrow$ 1.16
LLaVA-Next	3.55 $\rightarrow$ 10.25	15.44 $\rightarrow$ 32.00	1.52 $\rightarrow$ 6.55	73.78 $\rightarrow$ 69.04	1.15 $\rightarrow$ 1.04

Appendix

Table 1: Original

\rightarrow

Cautious prompt comparison of NOTA behavior for the 20 open-source models. All values except

d^{\prime}

are percentages.

d^{\prime}

is the signal-detection sensitivity of NOTA detection defined in Appendix 7.2 ; it does not depend on how often a model selects NOTA. The last row averages over the 20 models.

Explore similar work

May 28, 2026

ReactBench: A Cause-Driven Benchmark for Multimodal Hallucination via Systematic Evaluation

While multimodal large language models (MLLMs) have achieved rapid progress in vision-language understanding, they remain prone to multimodal hallucinations, producing responses that are inconsistent with the visual input. Existing benchmarks predominantly focus on detecting hallucination outcomes rather than evaluating the underlying causes of these failures. Moreover, many benchmarks rely on simplistic scenarios and limited evaluation formats that no longer challenge state-of-the-art models. To address these limitations, we introduce ReactBench, a cause-driven hallucination benchmark featuring multiple tasks and an exam-style evaluation format. By generating adversarial images and hallucination-inducing queries, ReactBench introduces four targeted tasks: Relational Erasure, Counterfactual Attribute, Alteration Tracing, and Dense Counting. These tasks systematically expose co-occurrence bias, language priors, cross-image comparative perception deficiencies, and fine-grained perceptual bottlenecks. Beyond standard accuracy-based evaluation, we leverage Chain-of-Thought reasoning to identify fine-grained sub-causes of hallucination within each task. Extensive evaluations reveal that current MLLMs remain notably vulnerable to cause-specific hallucination triggers, demonstrating the value of ReactBench as a systematic and interpretable testbed for diagnosing and improving multimodal model robustness. The project page is available at https://reactbench.github.io/.

Shizhe Zhou, Bohan Jia, Kai Wu +4

East China Normal University

VLM Robustness Hallucination Detection Hallucination in Language Models Multimodal Large Language Models Multimodal Hallucination Multimodal Model Evaluation

Ask PDF

Aug 4, 2026

KnowHal: A Knowledge-Driven Benchmark for Comprehensive Multimodal Hallucination Evaluation

Hallucination remains a critical challenge for developing trustworthy Multimodal Large Language Models (MLLMs). While existing benchmarks mainly focus on entity, attribute, and relation hallucinations, knowledge-related failures are often investigated separately, lacking a unified evaluation framework across different hallucination dimensions. To overcome this, we propose KnowHal, a benchmark that explicitly incorporates knowledge hallucination into multimodal hallucination evaluation spanning four dimensions: entity, attribute, relation, and knowledge. KnowHal constructs paired positive and negative questions over shared images and entities, enabling controlled comparisons among perceptual errors, knowledge-related errors, and false-premise acceptance. The benchmark contains 1,800 samples across 10 domains and 50 categories, constructed through a semi-automated pipeline combining LLM assistance, CLIP-based filtering, and human verification. We evaluate 14 representative MLLMs on KnowHal and conduct extensive analyses. Results show that the knowledge dimension consistently presents the greatest challenge for nearly all evaluated models, while most models exhibit substantial performance degradation on negative questions, revealing limited robustness to false premises. By unifying four hallucination dimensions with paired question design, KnowHal addresses an important gap in existing evaluation frameworks and enables a more comprehensive assessment of hallucinations in MLLMs.

Ruihan Li, Jiyang Tan, Kailin Jiang +5

1C-FAIR & School of software, Shandong University · University of Science and Technology of China · 3Jiangnan University

Multimodal Hallucination Multimodal Model Evaluation

Ask PDF

Mar 18, 2026

FINER: MLLMs Hallucinate under Fine-grained Negative Queries

Multimodal large language models (MLLMs) struggle with hallucinations, particularly with fine-grained queries, a challenge underrepresented by existing benchmarks that focus on coarse image-related questions. We introduce FIne-grained NEgative queRies (FINER), alongside two benchmarks: FINER-CompreCap and FINER-DOCCI. Using FINER, we analyze hallucinations across four settings: multi-object, multi-attribute, multi-relation, and ``what'' questions. Our benchmarks reveal that MLLMs hallucinate when fine-grained mismatches co-occur with genuinely present elements in the image. To address this, we propose FINER-Tuning, leveraging Direct Preference Optimization (DPO) on FINER-inspired data. Finetuning four frontier MLLMs with FINER-Tuning yields up to 24.2% gains (InternVL3.5-14B) on hallucinations from our benchmarks, while simultaneously improving performance on eight existing hallucination suites, and enhancing general multimodal capabilities across six benchmarks. Code, benchmark, and models are available at https://explainableml.github.io/finer-project/.

Rui Xiao, Sanghwan Kim, Yongqin Xian +2

Technical University of Munich · Munich Center for Machine Learning · Helmholtz Munich +2

VLM Hallucination VLM Hallucination Mitigation Multimodal Model Evaluation

Ask PDF