Vision language models, the AI systems that can analyze images and answer questions about them, are failing at a surprisingly fundamental task: actually looking at what's in the image instead of relying on what they expect to see. A new research framework called SABRE has exposed this weakness across six leading models, with accuracy scores ranging from just 17.8% to 31.3% when visual evidence conflicts with learned assumptions.

The problem sounds simple but reveals something troubling about how these models work. When you show a vision language model (VLM) an image of a dog made of wood, or a cat in a kitchen where cats don't normally appear, the model often defaults to what it learned during training rather than what the image actually shows. This tendency to rely on "world priors," or learned expectations about how objects and scenes normally look, undermines the core purpose of visual AI: to see and understand what's actually there.

Why Can't AI Models Just Look at What's in Front of Them?

The challenge isn't that these models are broken; it's that building reliable tests to catch these failures has been nearly impossible. Traditional benchmarks take months or years to create because researchers must manually collect images, annotate them, and verify that the test actually measures what it claims to measure. By the time a benchmark is ready, the models have often already improved past it, making it useless for identifying remaining weaknesses.

SABRE solves this by automating much of the benchmark construction process. A researcher provides a simple task description in plain language, along with rules about what should be tested. The system then generates images, creates questions and answers, filters out easy cases, and has humans verify the results. This pipeline can be refreshed and adapted as models improve, rather than becoming obsolete.

What Exactly Are These Models Getting Wrong?

The researchers tested SABRE by creating a benchmark called SABRE-Prior, which specifically targets whether models follow visual evidence when it contradicts their expectations. The benchmark includes 600 images and 1,000 questions organized into four categories:

• Context: Unexpected objects placed in familiar scenes, like a bicycle in a bedroom, to test whether models notice what doesn't belong.
• Texture: Counterfactual materials, such as a wooden dog or a glass car, to see if models can recognize objects even when they're made of unusual materials.
• Attribute: Noncanonical component counts, like a three-legged chair or a car with five wheels, to test whether models accurately count and describe what they see.
• Language Elicitation: Questions worded in ways that suggest an answer not actually present in the image, to see if models are influenced by language rather than visual evidence.

Across all six models tested, the macro-average accuracy was just 22.6%, with the worst performers scoring below 18%. Even the best-performing model struggled significantly with these deliberately challenging but valid visual tasks.

How to Evaluate Vision Language Models More Effectively

The SABRE framework offers a reusable approach for continuously testing and improving VLMs as they evolve. Rather than treating benchmarks as fixed endpoints, the system treats them as living tools that can be updated and expanded. The process includes several key steps:

• Automated Image Generation and Editing: The system creates images that match specific task requirements, then applies targeted edits to ensure the visual evidence aligns with the test design.
• Model-in-the-Loop Filtering: A filtering VLM automatically removes questions that are too easy, keeping only those that expose potential failures in the models being tested.
• Human Verification and Repair: Every retained test case undergoes human review to confirm it's valid, with reviewers able to fix localized image problems or refine questions without regenerating entire images.

The researchers also tested SABRE on two other types of reasoning tasks: counting and spatial reasoning. These pilot tests showed that the framework can adapt to different types of visual challenges without requiring a complete redesign for each new task.

What makes SABRE particularly valuable is that it addresses a growing problem in AI development. As fixed benchmarks become saturated and models perform well on them, researchers struggle to identify what weaknesses remain. A benchmark that takes a year to build is nearly useless if the models have already moved past it by the time it's published. SABRE's automated pipeline makes it possible to keep pace with model development and catch failures that might otherwise go unnoticed.

The low accuracy scores on SABRE-Prior suggest that even the most advanced vision language models available today have significant room for improvement in their core function: actually grounding their answers in the visual evidence provided, rather than in learned assumptions about how the world typically looks. As these models become more integrated into real-world applications, from medical imaging to autonomous vehicles, the ability to reliably test whether they see what's actually there becomes increasingly critical.