Reproduce “R$^2$-Omni: Reducing Structural and Precision Redundancy in Omni-LLMs for Long-Form Generation” 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
Bolong Feng, Pengfei Zhou, Zhenglin Wan, Wei Wang, Wangbo Zhao
R²-Omni compresses long audio-video contexts by consolidating structurally redundant multimodal tokens before prefill, then assigning cached values different precisions according to estimated attention-readout error and semantic signals. Across three Omni-LLMs and three long-form generation benchmarks, it reports average 5.0× KV-cache compression while retaining 97.98% of baseline generation quality; these are aggregate results, not guarantees for each metric or model.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.