WorldSense is a real-world omni-modal benchmark for evaluating multimodal LLMs on audio-visual-text video understanding. The benchmark contains synchronized audio, visual, and text inputs and is designed to require synergistic use of audio and video signals. WorldSense comprises a diverse collection of 1,662 audio-visual synchronized videos organized into 8 primary domains and 67 fine-grained subcategories, with 3,172 multiple-choice QA pairs spanning 26 distinct task types. Annotations were produced and quality-checked by expert annotators. The benchmark targets grounded reasoning and comprehensive evaluation of models that must integrate vision, audio, and textual cues.
No results indexed yet — be the first to submit a score.
Submit a checkpoint and a reproduction script. We will run it, publish the score, and — if it takes the top — annotate the step on the progress chart with your name.