Benchmarking VLM Scene Understanding: How Well Do Vision-Language Models Interpret Video Narratives?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

This study benchmarks three vision-language models Gemini 1.5 Pro, LLaMA 3.1 70B (via Groq), and LLaVA-NeXT-Video 7B , on structured video scene annotation across six dimensions: subject identification, action description, emotional tone, spatial relationships, scene transitions, and lighting/atmosphere. Results show proprietary models lead on temporal tasks, but all models struggle with emotional tone and spatial reasoning.