Source-linked AI summary
ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog
Lingao Xiao, Yalun Dai, Yangyu Huang, Qihao Zhao, Wenshan Wu, Hugo He, Ruishuo Chen, Jin Jiang, Qianli Ma, Jiahuan Zhang, Xin Zhang, Ying Xin, Yang Ou, Yan Xia, Scarlett Li, Longbo Huang, Zhipeng Zhang, Yang He, Yap Kim Hui, Yan Lu
TL;DR
Turning a paper into a poster, talk video, and blog remains labor-intensive, while practical workflows must preserve editability and connect the outputs. ResearchStudio-Reel addresses this with a native-editable, experience-level workspace whose posters achieve 3.56 vs. 3.03 aesthetics against authors’ posters and win overall quality on 74 and 95 of 100 papers under two judges.
Problem
Turning a paper into a coherent poster, talk video, and blog remains labor-intensive, while practical workflows need editable outputs connected in one navigable deliverable.
Method
ResearchStudio-Reel uses five composable skills to generate native-editable poster, video, and bilingual blog artifacts, then binds them through Paper2Reel.
Results
3.56 vs. 3.03 aesthetics: on 100 benchmark papers, the system beats authors’ posters and wins overall quality on 74 and 95 papers under two judges.
Takeaways & Limitations
The pipeline packages editable source artifacts with an aligned interactive viewer, supporting revision and reuse across poster, video, and blog formats.
Takeaways & Limitations
Evaluation is proxy-bound: the aesthetic and comprehension rubrics do not measure whether readers actually absorb the work, and the fill loop optimizes geometric density.
Abstract
from arXiv · showhide
Despite growing automation, turning a paper into a coherent poster, talk video, and blog piece often remains a labor-intensive last mile. Recent systems increasingly generate multiple dissemination formats, but a practical workflow must also keep the outputs editable in native tools and bound into one navigable deliverable for revision and reuse. We present ResearchStudio-Reel, a native-editable dissemination workspace that binds its three artifacts into one interactive deliverable at the experience level, implemented as five skills executable in Claude Code and Codex: one shared extractor, three editable artifact generators, and one interactive convergence layer. A shared asset bundle feeds a PowerPoint poster and video deck, plus a bilingual Word blog; rather than re-rendering the paper into a fourth format, Paper2Reel converges these already-produced artifacts at the experience level, binding poster regions, video segments, and blog passages into one interactive viewer. Artifact-specific release checks make this delivery contract testable, and Paper2Poster additionally uses a measured-fill loop. On the Paper2Poster benchmark, our Claude Code configuration achieves the best scores among automated systems on all three aesthetic sub-criteria and the best or tied-best scores on two of three information sub-criteria. Under two VLMjudges, it exceeds the authors' posters in average aesthetics (3.56 vs. 3.03) and wins on overall quality on 74 and 95 of the 100 papers under the two judges. The full pipeline additionally packages the native-editable source artifacts and their aligned viewer. Project is available at https://aka.ms/ResearchStudio
1 Introduction
ResearchStudio-Reel addresses the labor-intensive research dissemination layer with a native-editable workspace that generates and converges a poster, narrated video, and bilingual blog from one paper. Its shared extraction bundle, editable source artifacts, and alignment record support revision and cross-artifact navigation, while generated posters achieve strong benchmark and judge results.
- Motivation: The dissemination layer spans conference posters, talk videos, and audience-facing blog pieces, each requiring time-consuming figure selection, layout, narration, or adaptation.This work targets the last mile of research communication, which is structurally separate from the paper itself.
- System: ResearchStudio-Reel turns one paper PDF into a print-ready poster, narrated talk video, and bilingual blog through five composable skills.The skills include Paper2Assets, Paper2Poster, Paper2Video, Paper2Blog, and Paper2Reel, implemented in Claude Code and Codex.
- Convergence: Paper2Reel connects poster regions, video segments, and blog passages in one navigable HTML surface, converging already-produced artifacts at the experience level.A shared Paper2Assets pass provides stable section identifiers, figure handles, and claim anchors for mutual cross-referencing.
- Editable delivery: The delivery contract packages one shared Paper2Assets bundle with native-editable PowerPoint poster and video-deck sources and a Word bilingual blog.Native source files and the alignment record remain first-class deliverables so authors can revise artifacts without rerunning the entire generator.
- Evaluation: 3.56 vs. 3.03 aesthetics: generated posters exceed the authors’ posters by ∼17.5% under two VLM judges and win overall quality on 74 and 95 of 100 papers.On the Paper2Poster benchmark, the posters achieve the best automated-system scores on all three aesthetic sub-criteria and the best or tied-best scores on two of three information sub-criteria.
2 Skills
ResearchStudio-Reel implements five executable skills: one shared extractor, three native-editable artifact generators, and an interactive convergence layer. The skills enforce cross-artifact consistency and release quality through shared assets, measured poster filling, deterministic media and document checks, and an aligned viewer.
- Paper2Assets: Paper2Assets extracts text, captions, metadata, summaries, scripts, and cleaned figures once into a shared bundle that downstream skills consume without reopening the PDF.Stable figure naming keeps references consistent across the poster, video, and bilingual blog.
- Paper2Poster: Paper2Poster treats the poster as a fixed-size, readable, deliberately designed, editable page and uses a discrete measured-fill loop to size its sections.Each pass computes fullRatio from rendered content height and selects remediation from five categorical verdicts.
- Paper2Video: Paper2Video exports an editable, user-facing video.pptx and applies mandatory deterministic checks for audio, subtitles, highlights, timing, frames, and duration safety.The package gate verifies that audio, captions, slides, highlights, and timeline.json remain aligned for Paper2Reel.
- Paper2Blog: Paper2Blog produces two editable articles from one evidence base, keeping Chinese and English versions consistent in results, methods, figures, and links while adapting their editorial register.Both languages use the same selected figure set, reviewed for readability and caption compatibility after DOCX resizing.
- Paper2Reel: Paper2Reel reads completed artifacts and builds a self-contained viewer whose alignment record connects poster sections, slides, video times, captions, and blog blocks.Readers can navigate from poster regions to synchronized video and blog content, switch language, inspect slide context, and download native files.
3 Experiments
Experiments evaluate ResearchStudio-Reel’s poster quality on the 100-paper Paper2Poster benchmark and audit the full pipeline’s artifact capabilities and operational profile. The system achieves leading automated poster scores while producing editable, aligned dissemination artifacts, with quality gains tied to its runtime configuration and a visible coverage–legibility tradeoff.
- Experimental setup: The Paper2Poster evaluation compares single-shot vision-language models, agentic poster pipelines, ResearchStudio-Reel configurations, and author references under a shared scoring protocol.The benchmark contains 100 papers, uses two VLM judges, and applies six aesthetic/information criteria plus PaperQuiz reading comprehension.
- Poster quality: 3.56 Aesthetic Overall is best among systems and exceeds the authors’ 3.03; ResearchStudio-Reel leads automated systems on all aesthetic sub-criteria.It is best or tied-best on two of three information sub-criteria and wins overall on 74 papers under Claude and 95 under GPT.
- Analysis: PaperQuiz rankings nearly reverse aesthetic rankings, indicating a tension between dense content coverage for reading comprehension and selective, legible visual layouts.The analysis identifies P2P as leading both PaperQuiz splits while stronger reading-comprehension systems can score less well visually.
- Controlled comparisons: 0.64 aesthetic points separate single-shot Claude from the same model using composition and measured fill, which reaches 3.56 versus 2.92.Layout likewise rises from 2.97 to 3.99, showing that runtime configuration materially affects the aesthetic result.
- Capability coverage: The capability audit finds that ResearchStudio-Reel emits poster, video, and bilingual blog artifacts with editable source files, shared intermediate assets, and one navigable aligned viewer.The audit measures delivery-contract breadth rather than post-generation editing effort or the usability benefit of the aligned viewer.
- Operational profile: ∼89 minutes and ∼2.6M input / ∼276K output tokens are required for a full four-artifact bundle from one PDF in the instrumented runs.The shared Paper2Assets extraction is reused across downstream skills, while Paper2Poster and Paper2Video have the largest generator subtotals.
4 Applications
ResearchStudio-Reel targets three dissemination workflows: camera-ready authors, research organizations, and courses or reading groups. All use shared extraction, native-editable artifacts, and one Paper2Reel viewer, while differing in operators, artifact emphasis, and revision needs.
- Target contexts: Three target contexts are camera-ready authors, research or industrial-research organizations, and graduate courses, reading groups, or public-understanding outlets.These contexts span individual single-paper polishing, publication-intake workflows, and recurring paper-of-the-week briefing packs.
- Shared workflow: All use cases share Paper2Assets extraction, native-editable PowerPoint and Word artifacts, and the Paper2Reel viewer.Differences concern who runs the pipeline, which bundle artifacts they use, and what they revise afterward.
- Camera-ready authors: Accepted-paper authors can generate a print-ready poster, talk video, and public-facing blog within days through one Claude Code session per paper.One extraction pass populates Paper2Assets before the three generators emit editable artifacts.
- Research organizations: Research organizations can auto-generate draft dissemination artifacts for each accepted paper and hand them to editorial or communications teams for polishing.Editable PowerPoint and Word outputs support draft-and-revise handoff rather than black-box rendering.
- Courses and reading groups: Courses and reading groups can turn reading lists into briefing packs containing editable slides, talk videos, and Word blog pieces cross-linked in one viewer.Students can move between poster, video, and blog while reviewing, and instructors can revise editable artifacts.
5 Related Works
Prior work develops artifact-specific generators for posters, slides, videos, and long-form articles, alongside agentic execution mechanisms and multi-artifact presentation systems. ResearchStudio-Reel distinguishes itself through native-editable outputs, section-level alignment, and experience-level convergence in one interactive viewer.
- Scope of Related Work: Related work covers artifact-specific generation for posters, slides, videos, and long-form articles, as well as increasingly unified presentation suites.The paper organizes these works by output type before discussing agent and skill mechanisms and closely related multi-artifact systems.
- Poster Generation: Poster systems explore visual-in-the-loop pipelines, hierarchical content-layout reasoning, separated processing stages, specialized agents, benchmarks, and editable PPTX output.Examples include Paper2Poster, PosterForest, P2P, and PosterGen.
- Slides and Presentation Videos: Slide-generation systems progress from learned sentence selection and retrieval-based summarization to multimodal content-layout generation, edit-based workflows, and editable scientific PPTX decks.The cited systems include PPSGen, D2S, DOC2PPT, SlideSpawn, PPTAgent, and SlideGen.
- Long-Form Articles: Long-document work includes context repackaging, page-content alignment, discourse-graph summarization, reinforcement learning, and critique-augmented process supervision.These approaches are represented by HERA, PTSPI, GoSum, and LongDPO.
- Multi-Artifact Systems: PaperX and OmniPresent generate multiple dissemination artifacts, whereas ResearchStudio-Reel emphasizes native editability, section-level alignment, and experience-level convergence from one shared asset bundle.Its editable PowerPoint poster, video deck, and bilingual Word blog are bound into one interactive viewer mapping poster regions, video segments, and blog passages.
6 Future Work
Future work should move beyond proxy-based evaluation toward measuring whether readers actually absorb the work. In particular, the system should close its optimization loop on controlled human reading-and-recall signals rather than geometric density or competing aesthetic and comprehension proxies.
- Evaluation: The evaluation is proxy-bound: aesthetic scoring and PaperQuiz reward cleaner and denser posters, respectively, rather than directly measuring reader absorption.A denser poster can improve comprehension while a cleaner poster improves aesthetics.
- Evaluation: The measured-fill loop optimizes a geometric density target, not understanding.Its current objective therefore does not establish whether readers absorb the work.
- Future direction: The next step is to close the loop using a controlled human reading-and-recall signal instead of another proxy.This would directly target understanding rather than optimizing only presentation properties.
7 Conclusion
ResearchStudio-Reel presents native-editable research dissemination as a section-aligned workspace that generates a poster, video deck, and bilingual blog, then binds them through Paper2Reel. Its posters achieve the best scores among automated systems on all three reported benchmark criteria.
- Contribution: ResearchStudio-Reel reframes research dissemination as a native-editable, section-aligned workspace.The system is implemented as five composable skills.
- Contribution: One Paper2Assets bundle produces a native-editable PowerPoint poster, video deck, and bilingual Word blog.These are the three dissemination artifacts generated by the workspace.
- Contribution: Paper2Reel binds sections, slides, video times, captions, and blog passages across the three artifacts.The binding connects the artifacts into one section-aligned deliverable.
A Limitations
The system’s limitations include recurring end-to-end failure modes, untested transfer beyond ML, CV, and NLP, and evaluation coverage restricted mainly to poster capability and contract audits. Current audits do not measure human editing effort, post-edit fidelity, or section-level navigation.
- Recurring failure modes: Five failure modes recur in end-to-end runs, each paired with a one-sentence mitigation.One example is figure-cleanup residue, where a caption strip or body-text slice baked into a raster is re-cropped before downstream embedding.
- Domain scope: Transfer beyond ML, CV, and NLP venues is untested because poster conventions and evidence structures differ across domains.The underlying primitives are domain-agnostic, but each skill’s move catalogue may need extending for biomedicine, physics, or design-heavy fields.
- Evaluation coverage: Quantitative evaluation is reported only for the poster, using the Paper2Poster benchmark under two VLM judges.ResearchStudio-Reel has not been evaluated under the related Paper2Video protocol; video and blog are compared only on capability coverage.
- Evaluation coverage: Current audits document output contracts but do not measure human editing effort, fidelity after edits, or section-level navigation.The passages state that these omissions remain part of the evaluation coverage limitation.
B Ethics
Deployments should preserve provenance metadata because AI-generated dissemination artifacts can amplify scientific misunderstandings and may not appear AI-produced. Redistribution also requires checking text-to-speech service terms and the licenses governing paper figures and fetched logos.
- Provenance and disclosure: AI-generated posters, videos, and blogs can amplify correct understanding or confident misreadings of the underlying science.Narrated talk videos with burned-in subtitles are not obviously AI-produced at a glance.
- Provenance and disclosure: Deployments should preserve provenance metadata, including the skill and version, Paper2Assets bundle checksum, and artifact-specific rendering choices.The passage also identifies poster theme and header arrangement or TTS voice as relevant metadata.
- Licensing and redistribution: Redistribution should verify edge-tts service terms and confirm licenses for extracted paper figures and institution or venue logos.Figures follow the source paper’s license, while fetched logos follow their respective source licenses, including Wikimedia Commons and Wikidata where available.
C Reproducibility
ResearchStudio-Reel is released with pinned, installable dependencies and documented workflows, while its evaluation reuses the Paper2Poster benchmark protocol and supports configurable model endpoints. The reproducibility setup includes credential-free narration, host-provided language-model access, and standardized poster scoring across 100 papers.
- Code and dependencies: The project is open-source under the MIT license, with each skill’s SKILL.md defining its workflow and per-skill requirements.txt files pinning Python dependencies.The top-level install.sh symlinks skills into the host skills directory, while system dependencies include poppler-utils, libreoffice, ffmpeg, and headless Chromium.
- Models and credentials: Narration uses credential-free Edge TTS, while Claude Code and Codex obtain language-model access from their host runtime configurations.The authors’ runs routed credentials through a shared Copilot API proxy across several hosts; switching to first-party endpoints requires configuration changes rather than code changes.
- Benchmark protocol: 100 papers are evaluated using the Paper2Poster benchmark’s original six-criterion aesthetic and information rubric and PaperQuiz probe.Posters are downscaled to at most 2560 px on the long edge and scored by claude-opus-4.8 and gpt-5.5, with Table 1 reporting their mean.
D Per-judge Benchmark Scores
Table 5 provides an unaveraged, per-judge view of poster-quality evaluation, reporting six aesthetic and information sub-criteria plus two PaperQuiz comprehension splits for Claude and GPT judges. Aesthetic and information scores use a 1–5 scale, while Quiz reports answer accuracy as a percentage.
- Per-judge evaluation: Table 5 reports scores separately for Claude (claude-opus-4.8) and GPT (gpt-5.5), whose means form Table 1.The judges come from different model families, enabling inspection of either evaluator independently.
- Per-judge evaluation: Six aesthetic and information sub-criteria are reported for each judge, with overall columns averaging the three sub-criteria in each block.The abbreviated criteria include Element, Engagement, Layout, Low-level, Logic, Content, Overall, Detail, and Understanding.
- Per-judge evaluation: PaperQuiz comprehension is reported as answer accuracy (%) in two splits for each judge.Quiz is distinct from the aesthetic and information criteria, which are scored on a 1–5 scale.
E Poster Comparison Across Systems
Across systems rendered from the same source PDF, the single-shot LLM baselines recover comparable paper content but produce weaker poster layouts. Their main shortcomings are rigid grids, undifferentiated text, and figures placed without regard to whitespace.
- Content recovery: The single-shot LLM baselines recover the title, contributions, headline numbers, and sensible figures, keeping PaperQuiz accuracy close to ours.The passage attributes the performance gap primarily to layout rather than content recovery.
- Layout limitations: Their one-pass A0 posters use rigid grids with unequal columns, long undifferentiated text, and figures placed without regard to whitespace, making the sheet read flat.These layout choices are identified as the main source of the observed gap despite comparable information content.
- Comparison setup: Figure 11 compares our poster with three baseline systems and the authors’ poster, all rendered from the same benchmark-paper PDF.The comparison includes Paper2Poster Tool, PosterGen, portrait P2P, and single-shot Claude-4.8 Opus, GPT-5.5, and Gemini-3.1 Pro outputs.
F Poster Gallery Showcase · G Single-Shot LLM Baseline Prompt
The paper showcases 10 randomly sampled Paper2Poster outputs from a 100-paper benchmark and specifies a fixed single-shot HTML-poster prompt for comparing three LLM baselines. The prompt standardizes inputs, A0-landscape formatting, content fidelity, figure usage, and self-contained implementation constraints.
- F Poster Gallery Showcase: 10 posters are sampled uniformly at random from the 100-paper benchmark and displayed two per row.Each poster uses a distinct accent color, while column layout, visual style, and header are sampled per paper.
- G Single-Shot LLM Baseline Prompt: Three baselines—Claude-4.8 Opus, GPT-5.5, and Gemini-3.1 Pro—share one fixed prompt, isolating raw model capability from prompt engineering.Only the model identifier changes between runs.
- G Single-Shot LLM Baseline Prompt: Each baseline receives a structured paper summary, metadata, and available figure filenames with captions, then returns a single self-contained poster output.The summary is the Paper2Assets paper_spec.md with audio-narration lines removed.
- G Single-Shot LLM Baseline Prompt: The system prompt instructs the model to act as an expert academic conference poster designer and output only the complete A0-landscape poster HTML.The user prompt ends by requesting the complete poster HTML for the paper.
- G Single-Shot LLM Baseline Prompt: The poster must faithfully present Problem/Motivation, Method, Key Results with exact provided numbers, and a Takeaway without inventing numbers, results, or citations.This makes the structured summary authoritative for displayed content.
- G Single-Shot LLM Baseline Prompt: Figures must use exactly the provided filenames, appear near relevant sections with short captions, and never reference unavailable files.The per-paper prompt supplies title, authors, institutes, venue, structured summary, and available figures through placeholders.
- G Single-Shot LLM Baseline Prompt: Implementation constraints require inline CSS in one style block, no external stylesheets, web fonts, or JavaScript, and only system-safe fonts.The required fonts are Helvetica, Arial, Georgia, and Times.