Research notes, benchmark teardowns, and interactive visualizations
How a vision-language model learns to picture a scene before it reasons: 3DThinker imagines a static room from a few views as latent tokens aligned with VGGT's 3D features; 4DThinker imagines how objects move through a video, with no depth, camera or overlay input at inference.
A second multimodal LLM scores every query–candidate pair: the scores drop false negatives, pick diverse hard negatives and replace InfoNCE's one-hot target with soft labels.
A brief introduction to the Partial FC algorithm — how partial class center sampling makes ultra-large-scale parallel classification feasible.