Investigating Anisotropy in Visual Grounding under Controlled Counterfactual Perturbations

Visual grounding models fail in realistic scenarios: When image captions don't match the actual objects present (a common real-world situation), models often produce plausible-looking but incorrect results—for example, identifying the right type of object while ignoring important descriptive details that have changed.

Current benchmarks don't test for these failures: Existing visual grounding benchmarks assume the described object is always in the image, meaning models are rarely evaluated on mismatched captions, leaving a significant gap in how we assess their reliability.

The problem isn't simply about how embeddings are structured: The researchers tested whether "embedding anisotropy" (a geometric property of how AI models organize information) caused these failures by testing two different model architectures. They found no meaningful correlation between this property and approximation errors, suggesting the root cause is more complex than previously thought.

Explainability and trustworthiness are at stake: These approximation behaviors raise concerns about whether we can trust or understand why visual grounding models make their decisions, which is critical for deploying them in real-world applications where accuracy matters.

Deeper investigation of model geometry is needed: Rather than focusing on embedding anisotropy alone, researchers need to examine finer-grained geometric properties of how models organize information to truly improve robustness and faithfulness.