Video large multimodal models have improved at identifying visual evidence that spans multiple frames and at conducting open-web retrieval, according to the abstract of a new arXiv paper. These capabilities are reflected in recent video deep research (VDR) benchmarks, which evaluate how well models combine video understanding with web search.
The paper introduces GroundedVDR, a benchmark that appears to focus on grounding the dependencies between video evidence and web sources. The abstract notes that existing VDR benchmarks primarily—though the sentence is cut off—suggesting that current evaluation methods miss something important. GroundedVDR is positioned as a way to fill that gap.
Because this is a single source, the article can only report what the abstract states. The full construction of the benchmark, its evaluation methodology, and any experimental results are not available from the abstract alone.