Long-term, open-source efforts that keep growing.
A system-aware, complete-cost evaluation harness. It compares long-video pipelines under fixed backbones, matched evidence budgets, and one shared cost ledger spanning video-side and model-side work. Every method plugs into one stage of a shared pipeline, so a new idea can be compared with every baseline from its first run.
Interactive pages built around individual papers. Move the sliders, run the small experiments, and see the idea before reading the paper.
Dense low-resolution video plus a few high-resolution images. Half the decoding, lower latency and higher accuracy for long-video VLMs.
Two models can share an accuracy and make opposite mistakes. FlipAct splits when-to-act accuracy into a sensitivity and a criterion, and moves the criterion alone.
Selects the vision tokens worth keeping by maximizing multimodal coverage, for faster VLM inference at the same accuracy.
Aligns time series with visual and textual views of the same data to improve short-term and long-term forecasting.