analiza aiqbench
VCIFBench addresses a critical gap in video understanding evaluation by shifting focus from simple prompts to complex instruction following with explicit output constraints. The benchmark reveals that jointly satisfying multiple constraints remains a significant challenge for current multimodal models, though data-driven DPO training offers a viable path to improvement. For practitioners, this new resource provides a rigorous framework to assess and enhance model reliability in real-world applications requiring strict adherence to content, format, and structural requirements.
VCIFBench: Evaluating Complex Instruction Following for Video Understanding
arXiv:2606.04588v1 Announce Type: new Abstract: 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. We introduce VCIFBench, a benchmark for evaluating complex instruction following in video understanding. VCIFBench constructs constraint-rich instructions from both benchmark-adapted and directly video-grounded prompts, covering content, format, style, and structure requirements, and evaluates model outputs with a hybrid verification pipeline. The benchmark contains 306 satisfiable test instructions, a 540-pair DPO preference dataset, and a 30-item conflict diagnostic subset. Experiments on 10 MLLMs show that joint constraint satisfaction remains challenging. We further show that DPO training on VCIFBench data can improve instruction-following performance.
Abstract: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. We introduce VCIFBench, a benchmark for evaluating complex instruction following in video understanding. VCIFBench constructs constraint-rich instructions from both benchmark-adapted and directly video-grounded prompts, covering content, format, style, and structure requirements, and evaluates model outputs with a hybrid verification pipeline. The benchmark contains 306 satisfiable test instructions, a 540-pair DPO preference dataset, and a 30-item conflict diagnostic subset. Experiments on 10 MLLMs show that joint constraint satisfaction remains challenging. We further show that DPO training on VCIFBench data can improve instruction-following performance.
Computation and Language (cs.CL)
arXiv:2606.04588 [cs.CL]
(or arXiv:2606.04588v1 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2606.04588 arXiv-issued DOI via DataCite (pending registration)
Text integral sub licența deschisă a publicației · 132 cuvinte
Subiecte conexe
VCIFBench: Evaluating Complex Instruction Following for Video Understanding
arXiv · 25 aug. 2026
NVIDIA Nemotron 3 Nano Omni Powers Multimodal Agent Reasoning in a Single Efficient Open Model
NVIDIA · 28 apr. 2026
Links #5: 2026/07
LessWrong · 31 iul. 2026
NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
NVIDIA · 15 oct. 2026