← News

ResearchSources: arXiv · 18 Nov 2026

FrameRouter routes frame budgets per question for long-video QA

The FrameRouter authors propose giving each video question a different frame budget while keeping the average equal to a uniform baseline, reporting gains on Video-MME-v2.

In plain words

Video question-answering systems usually get the same number of video frames to look at for every question. The FrameRouter paper argues some questions need more visual evidence than others, so it sends each question to a frame budget suited to it while keeping the overall average unchanged. The method needs no extra training and works with existing frame-sampling policies.

What was announced

Researchers writing on arXiv analyzed Video-MME-v2, a benchmark for question answering over long videos, and found a mismatch in how it evaluates systems. Benchmarks typically assign a uniform frame budget to every question, but the authors report that different question subsets perform best under different frozen frame-budget policies, and that simply raising the frame count does not reliably improve all metrics. Their response is fixed-mean frame-budget routing: an inference setting where each question may receive a different number of frames, as long as the average budget matches a uniform reference.

The FrameRouter framework itself is described as training-free and operating at test time. It combines three parts: an evidence-demand router that decides how much visual evidence a question needs, a budget-constrained allocation rule, and a frozen bank of frame-sampling policies. Because the underlying samplers stay unchanged, the authors position FrameRouter as orthogonal to query-aware frame selection, meaning it could layer on top of such methods rather than replace them. Across Video-MME-v2 and several other long-video understanding benchmarks, they report that the system improves long-video question answering by spending the same average visual budget more selectively across questions. Code has been released on GitHub.

What it does not show

  • The source is the paper's abstract only; no full-method details, benchmark scores, or third-party evaluation are available to verify the reported improvements.

Sources

Source-support review completed 19 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 19 Nov 2026

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