CounterCount: A Diagnostic Framework for Counting Bias in Vision Language Models

In three linesCounterCount is a diagnostic framework to evaluate counting bias in vision-language models. Tests show VLMs perform well on factual images but degrade significantly on counterfactual images where visual attributes contradict learned priors. An inference-time attention modulation strategy improves accuracy by up to 8% across multiple VLMs.

Summary generated by Claude — human-verified