Benchmark RadarAI BENCHMARK PROFILE
VCIFBench
Evaluates complex instruction following for video understanding across content, format, style, and structure constraints with 306 test instructions.
- Released
- 2026-06-03
- Readiness
- Paper only
- Primary field
- General AI
Why it matters
Could address the need for video benchmarks testing explicit output constraints, but lacks public evidence of reuse path.
Motivation
Multimodal large language models have made rapid progress in video understanding, yet existing benchmarks largely rely on simple prompts and provide limited evidence about whether models can satisfy explicit output constraints.
Primary resources
Benchmark Radar records only publicly supported details and links back to primary sources for verification.