An arXiv paper introduces EvaOmni, a benchmark aimed at audio-visual reasoning. The abstract notes that omni large language models can naturally locate key audio and visual evidence and integrate information across modalities. EvaOmni appears designed to test that capability in a more rigorous way.
Existing audio-visual reasoning benchmarks, the paper argues, mostly rely on multiple-choice questions. The abstract begins to describe the limitations of that approach, though the provided excerpt cuts off mid-sentence. The new benchmark is described as "evidence-grounded," implying it expects models to point to specific audio and visual evidence when answering.
Because the source is a single abstract, details on task format, dataset size, and evaluation remain limited. The contribution is a benchmark that shifts focus from selecting an answer to grounding reasoning in evidence.