HumbleBench: Testing Epistemic Humility in MLLMs

• HumbleBench is a benchmark for multimodal LLMs that measures epistemic humility by testing both visual grounding and the ability to reject unsupported answers.
• It constructs high-quality, multiple-choice questions from the PSG dataset using GPT-4-Turbo and rigorous manual filtering to ensure clarity and correctness.
• The evaluation protocol focuses on both selection and rejection accuracy, with stress tests highlighting models’ risks of overconfident hallucinations in safety-critical settings.

Searching arXiv for the specified paper and closely related benchmark context. HumbleBench is a hallucination benchmark for multimodal LLMs (MLLMs) that evaluates epistemic humility: the ability to ground answers in the visual input and to decline to choose any provided answer when none is supported by the image. It was introduced in “Measuring Epistemic Humility in Multimodal LLMs” (Tong et al., 11 Sep 2025) to address a limitation of existing benchmarks, which primarily test recognition accuracy by asking models to select a correct answer among distractors, but do not test whether models can correctly reject all options. In this formulation, HumbleBench is designed to measure not only recognition but also correct abstention, especially in settings where overconfident hallucinations can produce unsafe errors.

1. Conceptual basis: epistemic humility in MLLMs

The benchmark is motivated by the claim that hallucinations in MLLMs generate content inconsistent with the input image and therefore pose risks in real-world applications, including misinformation in visual question answering and unsafe errors in decision-making. The central concern is that standard evaluation protocols incentivize guessing over honesty: if a benchmark always contains one correct option among plausible distractors, a model is never required to demonstrate that it can refuse unsupported answers.

Within this framework, epistemic humility is defined formally as an MLLM’s capacity to do two things: faithfully ground answers in the visual input, and decline to choose any of the given answers when none is supported by the image. This contrasts with pure recognition accuracy, because humility emphasizes correct abstention rather than only successful selection (Tong et al., 11 Sep 2025).

A plausible implication is that HumbleBench reframes hallucination assessment from a purely discriminative task into an evaluation of reliability under uncertainty. In that sense, it operationalizes a specific form of trustworthy behavior: the model must not merely identify what is present, but also recognize when the answer set is globally invalid.

2. Dataset construction and annotation pipeline

HumbleBench is built in three stages: data sourcing from the Panoptic Scene Graph (PSG) dataset, information extraction, and automatic question generation followed by manual filtering. PSG provides pixel-level panoptic segmentation and scene graph annotations for ∼40 K images. Each image has 11.04 objects on average, 5.65 relation triples, and multiple captions. From this source, 4 500 images are randomly sampled as the seed set (Tong et al., 11 Sep 2025).

For information extraction, objects and relations are read directly from PSG’s scene graph labels. Attributes require a separate procedure because PSG does not annotate colors or shapes. Each object is cropped by its mask and fed to InstructBLIP (Vicuna-7B) with the prompt, “This is a <object>, describe it briefly please.” The returned text snippets are then used as attribute labels.

Question generation is performed with GPT-4-Turbo, specifically gpt-4-turbo-2024-02-15-preview. The prompt engineering goals are explicit: each question must target exactly one hallucination type—object, relation, or attribute—must contain four plausible but incorrect distractors plus one “None of the above” choice, and must be answerable only with the image. For each image, GPT-4-Turbo is asked to produce 10 multiple-choice questions in strict JSON format, yielding ∼42 000 raw questions.

Manual filtering and quality control are then applied through a custom PyQt5 GUI that displays the image, question, five options, and buttons {Keep, Modify, Delete}. Annotators use three criteria: clarity and grammaticality, requiring well-formed and unambiguous questions; ground-truth checking, requiring that exactly one of A–D match PSG; and plausibility, requiring distractors that are plausible yet false.

The filtering outcomes are summarized below.

Outcome	Count	Share
Kept	18 304	43.7 %
Modified (answer relabeled)	4 527	10.8 %
Deleted	19 012	45.4 %

The final benchmark contains 22 831 high-quality questions. This suggests that the benchmark’s construction is not a purely synthetic pipeline; the manual intervention is substantial and functions as a quality-control stage rather than a minor post-processing step.

3. Hallucination taxonomy and question design

Each HumbleBench question probes exactly one of three hallucination categories. The first is object hallucination, in which the model claims that an object exists when it does not, or misses a real object. The second is relation hallucination, in which the model describes an incorrect spatial or semantic relation between two real objects. The third is attribute hallucination, in which the model assigns a wrong property such as color, shape, or material to an object (Tong et al., 11 Sep 2025).

The benchmark’s item format is multiple-choice with five options. A typical item includes the image, a question, four ordinary answer options labeled A–D, and an option E corresponding to “None of the above.” In some cases, one of A–D is correct; in other cases, none of A–D matches the image, and “None of the above” is the correct answer. The benchmark therefore makes false-option rejection an explicit part of the task.

The examples given in the source material illustrate the intended error modes. For object hallucination, a question such as “How many TVs are in the image?” may have the correct answer “0,” while a model may hallucinate the presence of a television. For relation hallucination, a question may ask whether a dog is sitting under a table. For attribute hallucination, a question such as “What color is the car?” requires selection of the visually supported color rather than a plausible but unsupported alternative.

A plausible implication is that the benchmark distinguishes between two failure modes that are often conflated: failure to recognize the correct visual fact, and failure to reject an answer set that contains no valid fact. HumbleBench is explicitly organized to make that distinction observable.

4. Evaluation protocol and metrics

The paper focuses on overall accuracy, but the source material specifies a broader evaluation framework for humility. Selection Accuracy is defined as

$$\mathrm{Acc} = \frac{1}{N} \sum_{i=1}^{N} \mathbf{1}(\hat{y}_i = y_i).$$

Rejection Accuracy is defined on the subset of questions whose true answer is “None of the above.” Let $M$ denote the number of such questions. Then

$$\mathrm{RejAcc} = \frac{1}{M} \sum_{j=1}^{M} \mathbf{1}(\hat{y}_j = \text{“None of the above”}).$$

The source also defines Precision, Recall, and F1 for the rejection decision by treating “None of the above” as the positive class:

$$P = \frac{TP}{TP + FP}, \qquad R = \frac{TP}{TP + FN}, \qquad F1 = \frac{2 \cdot P \cdot R}{P + R}.$$

Calibration is addressed with Expected Calibration Error (ECE). Predictions are partitioned into $K$ confidence bins $B_k$, with $\mathrm{conf}(B_k)$ the average confidence and $\mathrm{acc}(B_k)$ the accuracy in bin $k$. Then

$$\mathrm{ECE} = \sum_{k=1}^{K} \frac{|B_k|}{N} \cdot |\mathrm{acc}(B_k) - \mathrm{conf}(B_k)|.$$

These definitions indicate that HumbleBench can be used not only for categorical accuracy measurement but also for abstention-sensitive and calibration-sensitive evaluation (Tong et al., 11 Sep 2025). This suggests that the benchmark is intended to support analyses of both discriminative performance and decision reliability.

5. Experimental results and stress tests

The experimental evaluation covers two model groups: 12 general-purpose MLLMs and 7 specialized reasoning MLLMs. The general-purpose group includes Qwen2.5-VL-7B, LLaVA-NEXT-7B, Molmo-D-7B, DeepSeek-VL2-3B, InternVL3-8B, LLaMA3.2-11B, Phi-4-5B, Gemma-3-4B, Cambrian-8B, Pixtral-12B, Idefics3-8B, and VILA1.5-8B. The specialized reasoning group includes Ovis2-8B, Mulberry-7B, R1-Onevision-7B, Visionary-R1-4B, LLaVA-CoT-11B, R1-VL-7B, and GLM-4.1V-Thinking-9B.

Among the reported entries, Qwen2.5-VL (7B) attains 67.77 on object questions, 70.43 on relation questions, 77.81 on attribute questions, and 72.20 overall. InternVL3 (8B) attains 65.85, 68.13, 76.00, and 70.19. Pixtral (12B) attains 60.23, 64.73, 74.12, and 66.63. Among reasoning models, GLM-4.1V-Thinking records 69.30 on object questions, 71.25 on relation questions, 79.24 on attribute questions, and 73.46 overall, identified as the best reasoning result in the table. R1-VL (7B) records 63.59, 67.96, 74.03, and 68.73 (Tong et al., 11 Sep 2025).

Two stress tests are central to the benchmark’s analysis. In HumbleBench-E, all original correct options are removed, so “None of the above” is the only valid answer. Most models fall below the random baseline of 20%; GLM-4.1V scores 0.06%, while Cambrian is reported as an exception at 60.68%. In HumbleBench-GN, every image is replaced with Gaussian noise, again making “None of the above” the only correct answer. Under this condition, Qwen2.5-VL achieves 90.53%, R1-Onevision reaches 83.38%, and some models collapse under 30%.

These results separate ordinary benchmark competence from false-option rejection behavior. A model may perform competitively on standard multiple-choice recognition while failing severely when the task requires explicit abstention.

6. Empirical patterns, recommendations, and significance

Several aggregate findings are emphasized. Top general-purpose models can surpass 70% overall accuracy, but they typically guess incorrectly when forced to reject all distractors. Even leading reasoning models such as GLM-4.1V-Thinking fail near-catastrophically on “None of the above” questions. Across hallucination types, attribute questions are the easiest, with average ∼75% accuracy; object questions are the hardest, often involving counting and existence errors; and relation questions lie in between at ∼65%.

The reported interpretation is that high recognition accuracy alone is insufficient for safety-critical use. Overconfident hallucinations under noise or false distractors indicate a lack of true visual grounding, and without built-in humility, MLLMs risk propagating misinformation in domains where errors have serious consequences (Tong et al., 11 Sep 2025).

The benchmark also comes with practical recommendations. When using HumbleBench, the source recommends reporting both selection accuracy and rejection accuracy; calibrating model confidences, for example via temperature scaling, before measuring ECE; and incorporating a “check for None” decision rule in deployment, so that if no option exceeds a confidence threshold, the model abstains. Proposed extensions include moving beyond multiple-choice into open-ended and mixed-format questions with forced abstention, generating adversarial distractors via gradient-guided or human-in-the-loop methods, introducing domain-specific variants such as medical imaging and satellite imagery, and researching architectural or training modifications that jointly optimize accuracy and calibrated abstention, including auxiliary “I’m not sure” tokens in decoding.

Taken together, these elements position HumbleBench as a benchmark centered on explicit false-option rejection rather than recognition alone. Its “None of the above” design, its final scale of 22 831 questions, and its use of stress tests suggest a benchmark intended to measure whether MLLMs can admit uncertainty instead of hallucinating.