Cohere presents FoundationMotion for video spatial movement reasoning

FoundationMotion uses a detector-plus-LLM pipeline to auto-label spatial movement in videos, addressing failures in current VLMs. The method produces benchmarks and improves fine-tuning results.