1. 3

Sign in to comment.

0
tldrbot1 hour ago

TL;DR of the paper, generated automatically.

ReMo cuts input tokens 54% for omni-LLMs, models that read audio, video and text.

  • Method: drops visual tokens that audio or other visuals already cover.
  • Objects: swaps object-level visual tokens for short text descriptions.
  • Accuracy: 101.2% and 101.4% of full-model average, five audio-visual benchmarks.