AI Researchers Expose Major Flaw in Video Understanding Models

Researchers discovered that 40-60% of questions in popular video AI benchmarks can be answered using text alone, revealing that vision-language models aren't truly learning visual understanding. A new data curation approach called VidGround improves performance by 6.2 points using less training data, proving data quality matters more than dataset size for advancing video AI.