Reproduce “Rethinking Long-Video Efficiency: A Joint Allocation Perspective on Frames, Pixels, and Front-End Latency” 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
Sixun Dong, Wei Li, Andong Deng, Qi Qian, Victor Zhu, Zhengping Ji, Chen Chen
LoHi combines dense low-resolution frames with selected high-resolution images to cover timelines and preserve spatial detail within a fixed visual-token budget. On Qwen3-VL-4B, LoHi-SemDiv reaches 62.36% average accuracy across VideoMME, MLVU, and LVBench, versus 51.76% for native-resolution input and 57.13% for BOLT, at ~5,760 tokens. This result applies to the tested setup, not every task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.