Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos
AV-Flamingo is an open audio-visual LLM built for reasoning across long, complex videos, not just short clips.
The paper introduces a ~7M-instance training set for captions and question-answering over real-world videos. It uses a three-stage curriculum moving from short-range perception to long-horizon multi-event reasoning. Its chain-of-thought method grounds intermediate steps to timestamps, aiming to make temporal reasoning more aligned and interpretable. The authors report gains over similarly sized open models across more than 15 benchmarks, with competitive results against larger open-weight and closed systems. HF Daily Papers' note
The paper introduces a ~7M-instance training set for captions and question-answering over real-world videos. It uses a three-stage curriculum moving from short-range perception to long-horizon multi-event reasoning. Its chain-of-thought method grounds intermediate steps to timestamps, aiming to make temporal reasoning more aligned and interpretable. The authors report gains over similarly sized open models across more than 15 benchmarks, with competitive results against larger open-weight and closed systems. HF Daily Papers' note
score 7