本作品来自 MIT 许可的 PPT Master Examples。下方展示全部 18 页；点击页面可放大浏览。
01 / 18ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog02 / 18The last mile of research is a separate layer, prepared in the days after acceptance, and three gaps remain open03 / 18Prior work is organized by output type; contracts differ in editability, and VLM scores are mostly used for evaluation, not as release gates04 / 18The answer: a native-editable workspace of five skills — one shared extractor, three editable generators, one convergence layer05 / 18Paper2Assets reads the PDF once and emits one bundle every generator consumes verbatim; figure cleanup runs once06 / 18Paper2Poster composes on four orthogonal axes, fills the page with a measured loop, and exports native PowerPoint shapes07 / 18The fill loop is discrete: five verdict bands, one section edited per round, termination as a categorical fixed point08 / 18Paper2Video plans duration before rendering, anchors narration to visible objects, and gates an editable deck plus two MP4s09 / 18Paper2Blog writes the Chinese and English articles separately from one evidence map, then gates the Word files on layout10 / 18Paper2Reel is a presentation surface, not a fourth summary: one alignment sidecar makes four artifacts views of one section11 / 18Best automated system on all three aesthetic sub-criteria (3.56 vs the authors' 3.03) and on two of three information sub-criteria12 / 18The PaperQuiz ordering nearly reverses the aesthetic ordering: denser posters answer more; the fill loop packs to a target density13 / 18With the skill, prompt, and pipeline fixed, five configurations each produce a full page differing in figure choice, density, and accent14 / 18A full four-artifact bundle takes about 89 minutes of stage time and 2.6M input / 276K output tokens per paper15 / 18Three contexts span the dissemination spectrum on one pipeline: a camera-ready week, a lab's intake, a course's weekly packs16 / 18Five failure modes recur in end-to-end runs; the system is calibrated on ML / CV / NLP venues only; quantitative evaluation covers the poster17 / 18Supported: a native-editable, section-aligned workspace with testable gates. Open: reading-and-recall evaluation and bespoke diagrams18 / 18Sources
