CounterCount: Diagnosing Knowledge Bias in How Vision-Language Models Count Attributes

Name: CounterCount: Diagnosing Knowledge Bias in How Vision-Language Models Count Attributes
Uploaded: 2026-10-22T12:26:50.902418+00:00
Duration: 2366 s
Description: CounterCount diagnoses failures in Vision-Language Models (VLMs) when they count objects based on visual evidence that contradicts their internal knowledge of what an object should look like. The framework uses edited images and attention modulation to expose these prior-driven counting errors, showing that models often ignore crucial visual details needed for accurate counting.

CounterCount diagnoses failures in Vision-Language Models (VLMs) when they count objects based on visual evidence that contradicts their internal knowledge of what an object should look like. The framework uses edited images and attention modulation to expose these prior-driven counting errors, showing that models often ignore crucial visual details needed for accurate counting.

Open the full episode