Скачать или смотреть VDR-Bench: New Benchmark for Multimodal LLMs

In this AI Research Roundup episode, Alex discusses the paper: 'Vision-DeepResearch Benchmark: Rethinking Visual and Textual Search for Multimodal Large Language Models' This research introduces VDR-Bench, a new evaluation framework designed to test how Multimodal Large Language Models handle complex visual and textual search. The authors identify that existing benchmarks often allow models to bypass visual verification through textual shortcuts or language priors. To address this, VDR-Bench provides 2,000 multi-hop instances across ten diverse domains to simulate real-world, iterative fact-finding. The dataset uses a rigorous curation pipeline including manual cropping and knowledge-graph-based expansion to ensure genuine visual reasoning. This benchmark sets a new standard for assessing the deep research capabilities of modern multimodal systems. Paper URL: https://arxiv.org/abs/2602.02185 #AI #MachineLearning #DeepLearning #MultimodalLLM #VDRBench #ComputerVision #VisualSearch