When Vision AI Hallucinates Details

A photo shows a dog on a sofa. You ask an AI assistant: Is there also a cat sitting next to it? The assistant confidently says yes, even though there is no cat at all. In many everyday situations, that may be annoying. In areas like healthcare, robotics, or safety systems, it can become a serious problem.

When Specific Questions Cause Trouble

Today’s AI systems can describe images impressively well. But when questions become highly specific, they often claim to see details that do not exist. A wrong color, a missing object, or an incorrect position can already lead to a confident but false answer.

In their paper “FINER: MLLMs Hallucinate under Fine-grained Negative Queries,” selected for an oral presentation at CVPR 2026, MCML Junior Members Rui Xiao and Sanghwan Kim, together with MCML PI Zeynep Akata and collaborators, investigate why multimodal AI systems make auch mistakes, and how they can be trained to become more reliable.

Why Existing Tests Fall Short

Many benchmarks ask only simple questions such as “Is there a bicycle in the image?” But real users are usually more specific:

• Is the red bicycle next to the tree?
• Is the cat wearing a blue collar?
• What object is beside the chair with wooden legs?

To answer correctly, a model must understand objects, attributes, and relationships at the same time.

A New Benchmark for Fine-Grained Errors

The team introduced FINER (FIne-grained NEgative queRies), a new benchmark designed to test whether image AI can recognize small but important mistakes.

Instead of changing one obvious detail, FINER inserts subtle mismatches into otherwise plausible descriptions. These include:

• replacing one object with a similar but absent one
• changing a color or posture
• altering a spatial relation such as “behind” vs. “next to”
• asking “what” questions based on false assumptions

The researchers also developed FINER-Tuning, a training method that helps models reject wrong claims while preserving correct answers.

What the Experiments Showed

The study found that performance drops as questions become more detailed. Strong models that do well on simple tasks often fail when several details must be checked together.

After FINER-Tuning:

• leading models improved substantially on the new benchmark
• gains transferred to other hallucination tests
• general visual skills remained stable or improved
• models became better at saying “no” when the claim was false

©Xiao et al.

Figure 1: Qualiative examples of FINER-CompreCap MCQs for each category together with MLLM answers.

Why This Matters Beyond Benchmarks

As AI systems are used more widely, trust becomes essential. A helpful assistant should not only answer quickly, it should also recognize when a question contains a false assumption.

This work shows that better evaluation and better training can make multimodal AI more dependable in real-world settings, where small details often make the biggest difference.

Further Reading & Reference

If you would like to learn more about reducing hallucinations in multimodal AI systems, check out the full paper and the project page. The work will be presented as an oral presentation at CVPR 2026 (IEEE/CVF Conference on Computer Vision and Pattern Recognition), one of the leading conferences in computer vision and AI research.

R. Xiao • S. Kim • Y. Xian • Z. Akata • S. Alaniz
FINER: MLLMs Hallucinate under Fine-grained Negative Queries.
CVPR 2026 - IEEE/CVF Conference on Computer Vision and Pattern Recognition. Denver, CO, USA, Jun 03-07, 2026. Oral Presentation. To be published. Preprint available. arXiv GitHub