Measuring Epistemic Humility in Multimodal Large Language Models

The gist The authors present HumbleBench, a new hallucination benchmark designed to evaluate false-option rejection in multimodal large language models (MLLMs) under a forced-choice multiple-choice setting with a “None of the above” option, filling a key gap in current evaluation suites by assessing an abstention-oriented behavior relevant to trustworthy multimodal reasoning.

HumbleBench Design and Construction

HumbleBench provides a testbed for evaluating false-option rejection and abstention-oriented behavior in multimodal AI systems, consisting of 22,831 multiple-choice questions, each with five choices, one of which is set to “None of the above” (Page 2). The benchmark is constructed from a panoptic scene graph dataset [6], which provides fine-grained annotations for objects and relations; candidate attribute cues are generated separately and verified during manual filtering (Page 3). The construction pipeline involves several steps, starting with using the Panoptic Scene Graph (PSG) dataset as the data source because of its high-quality annotations (Page 2). Information extraction uses InstructBLIP to obtain candidate attribute cues by cropping objects and passing them to InstructBLIP with the prompt “This is a ⟨object⟩, describe it briefly please” (Page 3). Next, GPT-4-Turbo is prompted to generate multiple-choice questions along with distractor options using this information as context (Page 3). This automated process yields an initial dataset of approximately 42,000 questions before rigorous manual filtering (Page 3). The manual filtering process involves annotators checking five criteria: whether the question is clear and answerable from the image alone, whether the referred object, relation, or attribute is visually identifiable, whether there is exactly one correct answer among the five options, whether the “None of the above” option is correct when none of other options are visually supported, and whether distractors are plausible but clearly incorrect (Page 4). This process resulted in a final set of 22,831 questions (Page 4).

Evaluation Metrics and Behavioral Focus

The benchmark tests three types of hallucinations: object, relation, and attribute (Page 2). Unlike traditional benchmarks that focus on recognition, HumbleBench requires models not only to recognize correct visual information but also to identify when none of the provided answers are correct (Page 2). The evaluation metrics include Overall accuracy over all samples, Non-NOTA accuracy restricted to samples where yi!= ni, NOTA hit rate P(ˆyi = ni yi = ni), False-NOTA rate P(ˆyi = ni yi!= ni), NOTA selection rate P(ˆyi = ni) over all samples, and a Balanced accuracy for NOTA detection: 1/2 [NOTA hit + (1 - False-NOTA)] (Page 7). The authors interpret HumbleBench as a benchmark of false-option rejection and abstention-oriented visual reasoning rather than a complete measure of epistemic humility (Page 2).

Experimental Setup and Model Evaluation

Experiments were conducted on a variety of state-of-the-art MLLMs, including general-purpose models, specialized reasoning models, and the proprietary Gemini-2.5-Pro model (Page 3). The evaluation settings include the Original setting (standard HumbleBench), Shuffled NOTA randomly permutes the five options, Cautious prompt prepends an instruction encouraging the model to choose NOTA when visual evidence is insufficient, Noise image replaces each input image with Gaussian noise while keeping the original question and label, Noise + grounding combines the noise image with an instruction to rely strictly on visual evidence, and NOTA-only removes the original correct non-NOTA option (Page 3). The results show that accuracy alone can hide severe failures in false-option rejection (Page 3).

Key Empirical Findings

The results demonstrate that accuracy alone cannot distinguish between correct recognition and conservative NOTA behavior (Page 6). For instance, Qwen2.5-VL obtains strong non-NOTA accuracy (78.98%) but only a 32.33% NOTA hit rate, whereas Cambrian has much lower overall accuracy (55.56%) but a much higher NOTA hit rate (65.32%) (Page 6). The noise-image experiments expose a different failure mode where models revert to parametric knowledge when visual evidence is destroyed, with the average overall accuracy dropping sharply to 27.58% with noise images and further to 22.53% with the forced-grounding prompt (Page 6). Furthermore, the NOTA-only column in Table 6 isolates the ability to select NOTA when all nonNOTA answers are removed, showing that even strong models often over-commit to a concrete answer instead of selecting the explicit reject option (Page 9).

Limitations and Implications

The study concludes that recognition accuracy alone is insufficient for assessing hallucination robustness and highlights the importance of evaluating false-option rejection as a specific behavioral facet related to epistemic humility (Page 10). The findings suggest that option-position robustness varies substantially across models, while the cautious prompt experiment shows a trade-off between higher humility and lower recognition accuracy (Page 9). Moreover, the noise-image experiments expose a serious issue with visual grounding as many models with reasonable performance on HumbleBench fail sharply when visual evidence is destroyed (Page 10). The authors emphasize that model size cannot be isolated from other factors in these comparisons, since the evaluated models differ in architecture, training data, visual encoders, and post-training strategy (Page 8).

Data Availability

All data supporting the findings of this study are publicly available. HumbleBench, including code and data, is available at maifoundations/HumbleBench (Page 12). The authors provide basic cost indicators to make the construction and evaluation workload easier to reproduce (Page 13).

--- Page 1 ---

Measuring Epistemic Humility in Multimodal Large Language Models

Bingkui Tong, Jiaer Xia, Sifeng Shang, Kaiyang Zhou. Mohamed bin Zayed University of Artificial Intelligence, United Arab Emirates. Hong Kong Baptist University, Hong Kong. kyzhou@hkbu.edu.hk; bingkui.tong@mbzuai.ac.ae; xiajiaer@life.hkbu.edu.hk; cssfshang@comp.hkbu.edu

--- Page 2 ---

Question: How many snowboarders are performing tricks? A, B, C, D, E

Question: What is parked on the side of the road? A, B, C, D, E

Object Hallucination Question: What is the primary material of the airport floor visible? A-E

Attribute Hallucination Question: What color is the horse that a police officer is riding? A-E

Relation Hallucination Question: Where is the cow standing that is not in the water? A-E

--- Page 3 ---

Fig. 1: Examples from HumbleBench where correct options are marked green. This benchmark is derived from a panoptic scene graph dataset with rich object and relation annotations, together with candidate attribute cues that are manually verified. Different from existing hallucination benchmarks, HumbleBench has a “None of the above” option in each question to test whether models can identify when no provided answer is valid in a forced-choice multiple-choice setting. recall, these setups overlook an equally important ability for trustworthy AI: recognizing when none of the presented answers are correct and refusing to commit to a false choice [3]. The ability to abstain from giving incorrect answers aligns with the concept of epistemic humility[4, 5], which reflects awareness of one’s own knowledge limitations. However, in this work, we use the notion of epistemic humility in an operational and task-specific sense. HumbleBench does not aim to measure calibrated uncertainty, confidence calibration, or principled refusal in their full generality. Instead, it evaluates one directly observable behavioral facet related to epistemic humility: whether a model can reject all provided options when none is supported by the image in a forced-choice multiple-choice setting with a “None of the above” option. We therefore interpret HumbleBench as a benchmark of false-option rejection and abstention-oriented visual reasoning, rather than a complete measure of epistemic humility.

--- Page 4 ---

Table 1 contains more detailed statistics, such as question length and image size. Fig. 2 shows the distributions over answer choices and different hallucination types. Overall, the non-NOTA answer options A–D are relatively balanced. This ensures diversity in correct answers while preventing models from overfitting to one non-NOTA choice (Page 4). The three question types are fairly balanced, with a slight dominance in attribute-related questions (Page 4).

--- Page 5 ---

…

--- Page 6 ---

Table 3: Evaluation metrics used in the experiments. For sample i, yi is the ground-truth label, yˆi is the predicted label, and ni is the option label whose text corresponds to “None of the above”. Metric Definition Overall accuracy Accuracy over all samples. Non-NOTA accuracy Accuracy restricted to samples where yi!= ni. NOTA hit rate P(ˆyi = ni yi = ni), i.e., recall on true NOTA samples. False-NOTA rate P(ˆyi = ni yi!= ni), i.e., the rate of incorrectly abstaining on non-NOTA samples. NOTA selection rate P(ˆyi = ni) over all samples. Humility score Balanced accuracy for NOTA detection: 1/2 [NOTA hit + (1 - False-NOTA)] (Page 6).

--- Page 7 ---

Table 4: Performance of open-source and proprietary models on the original HumbleBench setting. All values are percentages. HS denotes the humility score in Table 3.

1. A new evaluation framework that explicitly measures false-option rejection in MLLMs by incorporating a None of the above option into multiple-choice questions, thus assessing an abstention-oriented behavior. This allows researchers to evaluate whether models can reject all provided options when none is supported by the image in a forced-choice multiple-choice setting with a “None of the above” option, moving beyond simple recognition accuracy.

2. A rigorous data construction pipeline leveraging Panoptic Scene Graphs (PSG) for high-quality annotations and using GPT-4-Turbo to generate questions, followed by a rigorous manual filtering process to ensure validity, resulting in 22,831 high-quality questions per benchmark. This pipeline ensures that models are tested on verifiable ground truth, minimizing dataset noise.

3. A comprehensive set of stress tests for MLLMs including Shuffled NOTA randomly permutes the five options and Noise image replaces each input image with Gaussian noise, to probe how models rely on visual evidence versus linguistic priors. This reveals critical failure modes, such as Qwen2.5-VL's failure in Question 6 where it reverts to the parametric knowledge of its language backbone.

4. A performance assessment that distinguishes between recognition accuracy and true behavioral robustness by reporting metrics like False-NOTA rate and a balanced NOTA-detection humility score. This moves evaluation beyond traditional accuracy, as the paper shows that accuracy alone can hide severe failures in false-option rejection.

5. Model selection strategies informed by the findings that reasoning orientation does not guarantee better performance, such as noting that R1-Onevision, which is fine-tuned from Qwen2.5-VL with the explicit goal of enhancing reasoning, performs significantly worse than its base model on HumbleBench. This guides researchers away from assuming reasoning oriented models exhibit stronger robustness against hallucinations compared to generalpurpose models.