Reproduce “DASH: Dynamic Audio-Driven Semantic Chunking for Efficient Omnimodal Token Compression” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Bingzhou Li, Tao Huang
DASH compresses audio-video inputs before they reach a language model, using changes between neighboring audio embeddings to create variable-length segments and guide video timing, then ranking audio tokens by boundary, uniqueness, and attention signals. On WorldSense with Qwen2.5-Omni-7B, it reports 44.9% average accuracy at 25% target retention, versus OmniZip’s 44.7% at 35%; full-token accuracy is 46.8%.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.