← News

ResearchSources: arXiv · 19 Nov 2026

Quicksviewer compresses video early with annealed Gumbel-Softmax cubing

The paper proposes a video LMM that groups frames into variable-size cubes and resamples them, reporting large compression and faster inference.

In plain words

Video language models usually process every frame equally, which wastes computation on repetitive footage. Quicksviewer instead splits a video into unevenly sized segments based on how much frames change, then resamples each segment. The authors say this cuts computation sharply while keeping understanding accuracy.

What was announced

The arXiv abstract describes a new perceiving paradigm for Large Multimodal Models. Instead of perceiving frames uniformly, the model partitions a nonuniform, redundant video into varying cubes using Gumbel Softmax, then applies unified resampling within each cube. A Gumbel-noise annealing strategy is proposed to make end-to-end training with this discrete cubing network stable, and compression happens online according to the momentum of frame differences.

The authors report an overall 45x compression rate and say the front-loaded compression in encoding reduces latency in the expensive prefill and decoding stages, which they argue suits mainstream EPD disaggregated inference deployment. Trained from a language backbone in three progressive stages with videos averaging 420 seconds at 1 fps, the model used only 0.8M video samples and, they claim, outperforms contemporary LMMs with the same backbone and comparable parameter counts. In an A/B pretraining comparison with identical backbone and data, cubing yielded lower training loss and 2.1 points higher benchmark accuracy than fixed partitioning. A project page links the code.

What it does not show

  • Evidence is limited to the paper's abstract, so the reported compression, accuracy and latency figures are author claims not independently verified.
  • The abstract does not specify which benchmarks, model sizes or baselines the comparison covers, nor the release status of the code.

For security and IT teams

Status
A code link is provided at quicksviewer.github.io, but the abstract does not state a deployment-ready release.

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.