Benchgen

WorldSense — Results

RankModelScore
1lfm3-omni-2b47.1
W

WorldSense

1 phaseActive

A 1,662-video, 3,172-question multiple-choice omnimodal (visual + audio + text) benchmark spanning 8 primary domains and 67 fine-grained subcategories across 26 distinct tasks, requiring tight audio-video coupling.

Overview

WorldSense

Category Metric Saturation Created

Paper

Quick answer: WorldSense ("Evaluating Real-world Omnimodal Understanding for Multimodal LLMs") is a benchmark introduced by Jack Hong, Shilin Yan, Jiayin Cai, Xiaolong Jiang, Yao Hu, and Weidi Xie (submitted February 2025, revised through v3 in March 2026) with 1,662 audio-visual synchronised videos across 8 primary domains and 67 fine-grained subcategories, and 3,172 multiple-choice QA pairs spanning 26 distinct tasks.

What WorldSense Measures

Most omnimodal benchmarks allow a model to answer from either the visual or audio stream alone. WorldSense is designed around tight coupling of audio and video — its evaluation tasks require models to effectively use the synergistic perception of both modalities together, rather than relying on a single channel. All 3,172 QA pairs were manually labeled by 80 expert annotators across multiple rounds of correction to ensure annotation quality, and the benchmark's 1,662 videos were systematically categorized into 8 primary domains and 67 fine-grained subcategories to cover a broad range of real-world scenarios.

Public Benchmark Scores

ModelScoreSourceDate
LFM3-Omni-2B47.10%Liquid AI model card2026-10-27
  • LFM3-Omni-2B — currently the only Benchgen-tracked model with a public WorldSense score.
  • Daily-Omni — a related audio-visual temporal-alignment benchmark also scored by LFM3-Omni-2B.

Sources

  • Hong, Jack, Yan, Shilin, Cai, Jiayin, Jiang, Xiaolong, Hu, Yao, Xie, Weidi. "WorldSense: Evaluating Real-world Omnimodal Understanding for Multimodal LLMs." arXiv:2502.04326, submitted 2025-02-06, revised 2026-03-01 (v3).