← News

Agents & MCPSources: arXiv · 19 Nov 2026

VideoACE benchmark tests how AI models use interaction to understand fast-action video

Researchers introduce VideoACE, a 1,000-question benchmark measuring whether video models can actively inspect footage, use tools, and revise judgments rather than answer from a single viewing.

In plain words

Most video AI tests show a model a clip once and score its answer. VideoACE instead checks whether a model can improve its understanding by choosing what to look at, adjusting detail, and reacting to feedback. The authors built 1,000 multiple-choice questions about fast, subtle actions for this purpose.

What was announced

The benchmark, described in an arXiv paper, targets high-frame-rate action videos, where the authors say rapid and subtle changes make targeted observation especially valuable. It contains 1,000 multiple-choice questions across six categories of fine-grained action understanding, with temporal evidence annotations and user hints screened to prevent answer leakage. The evaluation environment pairs video inspection tools with a user simulator, letting models adjust temporal coverage, sampling density, and spatial detail, then revisit evidence and revise judgments based on simulated user feedback.

The authors evaluated five models under three conditions: single-pass observation, autonomous tool use, and tool use combined with user feedback. They report distinct patterns in how the models gathered evidence and responded to chances for further interaction, and they position VideoACE as a testbed for agentic video understanding beyond fixed-observation performance. The paper excerpt does not name the models or give scores, so no specific rankings can be reported here.

What it does not show

  • The source is an abstract only; model names, scores, and detailed results are not available in the excerpt.
  • Results come from the benchmark's authors and have not been independently confirmed.

Sources

Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.

Coverage: arXiv: abstract.

Updates and corrections

No corrections so far.

First published 20 Nov 2026

New to the subject? AI from the Ground Up starts from the beginning.