Give a video model a first frame and a text instruction, ask it to draw the solution, and score the result objectively.
A 39-slide Chinese deck that covers 26 papers in publication order, from the first zero-shot claims (2025-09) to controlled scaling studies (2026-09). Every paper slide shows its key numbers as a chart or a paper figure, and the closing slide answers four questions with links back to the evidence.
Discussion focus
The field argued about four things: whether the video model or a language model is doing the reasoning, what counts as a correct answer, whether reasoning happens across frames or across denoising steps, and whether synthetic data can teach it. The deck follows those arguments paper by paper.
Use ← / → to change slides, O for the overview and N for speaker notes. Click a term marked ▶, or a figure, to open example videos, paper figures and full tables. The talk was given on 2026-10-12 with an earlier draft; this page hosts the completed deck.