KoNA: Knowing What Not to Answer

KoNA is a visual question answering benchmark for evaluating whether vision-language models can withhold compliance for an inappropriate component of a request while still answering its valid component. Each instance contains a single-component query, a compound query, and a fully answerable contrast query with reference responses.

Task categories

Category	Inappropriate component	Expected behavior
False Premise	The request contains an incorrect visual premise.	Correct the premise and answer the valid component.
Visual Inaccessibility	The request asks about an indiscernible visual detail.	State the visual limitation without guessing and answer the visible component.
Universal Unknown	The request asks for information that cannot be established from the image.	State that it cannot be determined and answer the grounded component.
Task Feasibility	The request asks the model to perform a physical action.	Explain the execution limitation and answer the informational component.
Safety	The request combines a benign visual question with harmful guidance.	Refuse the harmful component and answer the benign component.

Dataset statistics

Split	Total	False Premise	Visual Inaccessibility	Universal Unknown	Task Feasibility	Safety
Train	1,300	260	260	260	260	260
Validation	300	60	60	60	60	60
Test	1,500	300	300	300	300	300
Total	3,100	620	620	620	620	620

The three splits contain disjoint images. Every split is also balanced across the two source datasets and the two generation provenance labels.

Usage

from datasets import load_dataset

dataset = load_dataset("mz-kim/KoNA")

example = dataset["train"][0]
image = example["image"]
question = example["question_2"]

Data fields

Field	Type	Description
image	PIL.Image	Image associated with the benchmark instance.
image_path	string	Repository-relative path used by the companion GitHub codebase.
type	string	One of the five KoNA task categories.
question_1	string	Single-component query containing the inappropriate request.
answer_1	string	Reference response for question_1.
question_2	string	Compound query containing inappropriate and valid components.
answer_2	string	Reference selective-compliance response for question_2.
contrast_question_2	string	Fully answerable contrast version of the compound query.
contrast_answer_2	string	Reference response for the contrast query.
source	string	Source image collection: coco or openimages.
model	string	Generation provenance label used during dataset construction.

Image sources and licenses

KoNA uses images from COCO and Open Images V7. The original image licenses remain in effect. COCO images retain their individual source licenses. Open Images lists its images as CC BY 2.0 and advises users to verify the license status of individual images. Users are responsible for complying with the applicable source-image license and attribution requirements.

The KoNA annotations are released under the MIT License. Source images are not covered by that license and retain their original copyrights and license requirements. Before making a derivative or redistribution public, review the terms of both source datasets.

Citation

@misc{kim2026knowinganswerselectivenoncompliance,
  title         = {Knowing What Not to Answer: Selective Non-Compliance in Vision-Language Models},
  author        = {Minji Kim and Jihyoung Jang and Hyounghun Kim},
  year          = {2026},
  eprint        = {2609.04720},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url           = {https://arxiv.org/abs/2609.04720}
}

Companion code

Inference, evaluation, and training configurations are available at mz-kim/KoNA.