Preserving Acoustic Cues for Video Reasoning: An Efficient Uniqueness-Driven Token Compression Framework
Flash-VAReason is a visual-audio guided video reasoning framework that preserves informative acoustic cues while compressing audio tokens efficiently for MLLM-based reasoning.
- Project page: https://thuhcsi.github.io/interspeech2026-Flash-VAReason/
- Demo video asset path:
assets/demo_haiwei_v5.mp4 - Demo poster path:
assets/demo-poster.jpg
- Upload demo video to
assets/demo_haiwei_v5.mp4 - Upload paper PDF to
papers/flash-vareason.pdf - Add code and reproduction instructions