VLM4D: Spatiotemporal Reasoning Benchmark with 1,000 Video Samples

VLM4D is a benchmark of approximately 1,000 real-world and synthetic videos designed to evaluate spatiotemporal reasoning in Vision Language Models. Developed by Shijie Zhou and researchers at UCLA in 2025, the dataset provides curated video-text pairs to test model awareness of motion and time.

Use Cases

• Evaluating temporal sequencing in VLMs using video-text-to-text tasks
• Testing spatial awareness by querying object trajectories across video frames
• Benchmarking model reasoning on synthetic vs. real-world motion patterns

Strengths

• First benchmark specifically for VLM spatiotemporal reasoning
• Includes ~1,000 curated video samples
• MIT licensed for open research

Limitations

• Small sample size of ~1,000 records
• Potential domain gap between synthetic and real-world video logic

Provenance

Source

UCLA (Shijie Zhou et al.), Arxiv 2508.02095

Collection Method

Curated from real-world sources and synthetic generation

Freshness

Last updated February 2026.

Geography

United States