Summary of "ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog"

Summary (Overview)

• ResearchStudio-Reel is a five-skill composition that turns a single research paper PDF into three editor-ready artifacts: a print-ready conference poster, a synchronized narrated talk video, and a bilingual (Chinese + English) blog post.
• All artifacts are editable in their native tools (PowerPoint for poster and video deck, Word for blog), not one-way renders. A shared Paper2Assets extraction layer ensures cross-artifact factual consistency.
• Paper2Reel binds the three artifacts into a self-contained interactive HTML viewer: section-level clicks jump the video, slides, captions, and blog to matching content.
• On the 100-paper Paper2Poster benchmark, posters generated by ResearchStudio-Reel outperform prior automated systems and single-shot frontier LLMs on every aesthetic and information sub-criterion, surpassing the authors' own posters on aesthetics (3.52 vs. 2.94 on a 5-point scale) and winning overall on 84–93% of papers.
• The pipeline is the only system to ship all three editable dissemination artifacts from a single run, replacing soft VLM-preference quality gates with hard categorical fill verdicts.

Introduction and Theoretical Foundation

Background and Motivation

The "last mile" of research dissemination—converting an accepted paper into a poster, a talk video, and a blog post—is still done manually by authors in the narrow window after acceptance. This is time-consuming, error-prone, and duplicated across artifacts. The same three artifacts are needed by industrial labs for public communication, graduate courses for weekly briefing packs, and multilingual organizations for cross-region outreach.

Three Recurring Gaps in Prior Work

The authors identify three structural gaps in existing automated systems:

1. (G1) Isolated extraction: Each artifact is solved by a separate monolithic system that re-extracts figures, captions, and metadata from scratch, leading to cross-artifact factual inconsistency (e.g., figure numbers on the poster not matching those referenced in the blog).
2. (G2) One-way renders: Prior systems ship as PDF poster, MP4 video, or HTML/markdown blog that cannot be reopened and tweaked in PowerPoint or Word. Any post-hoc correction means re-running the generator.
3. (G3) Soft quality gates: Quality is judged by continuous VLM-as-judge or aesthetic scores, so a layout scoring 7.8/10 is accepted even when load-bearing sections read as empty. The agent-versus-deterministic-primitive boundary is drawn inconsistently.

Theoretical Approach: Composition of Skills

The paper treats the last mile as an architectural problem. The solution is a composition of skills—thin agent-readable contracts that share one upstream extractor and wrap deterministic primitives in a measured-fill loop with hard pass/fail render gates. This responds to G1–G3 respectively:

• Paper2Assets extracts the paper once into a shared bundle (G1).
• Native-tool authoring ensures every artifact is editable in its source tool (G2).
• Measured-fill loops with categorical quality verdicts replace soft scores (G3).

The authors note three convergences making this tractable: (i) mature agent-driven skill runtimes (Claude Code, Codex), (ii) mature deterministic primitives (headless Chromium for HTML→PDF, LibreOffice+ffmpeg for slides→video, python-docx for .docx), and (iii) Edge TTS closing the gap on narration quality.

Methodology

The system is composed of five skills organized into three layers:

1. Shared Extractor: Paper2Assets

Reads a paper PDF once and produces a single bundle containing:

• Full body text with page breaks
• Detected figure captions and per-figure manifest
• Cleaned figure images (deterministic cleanup + visual-AI verifier)
• Paper metadata (title, authors, institutes, venue, paper/code links)
• Structured nine-section summary (Problem, Motivation, Contribution, Method, Dataset/Benchmark, Key Result, Ablation, Headline Numbers, Takeaway)
• Institution logos (from Wikimedia Commons) and venue logos (from Wikidata)
• QR codes for paper/code links
• Narration clip list in reading order

Key property: Single-owner extraction — downstream skills read the bundle and never re-open the PDF. A stable figure-naming invariant keeps artifacts factually consistent without cross-skill coordination. The pipeline is debuggable: a single correction propagates to all artifacts.

2. Three Generators

Paper2Poster (Poster Generator)

Turns the shared asset bundle into a print-ready conference poster in four coupled formats: self-contained HTML, print-resolution PDF, thumbnail PNG, and editable PowerPoint (.pptx).

Design requirements addressed:

• (A1) Composition without template explosion: Poster assembled at build time from four orthogonal axes (column layout, visual style, title-band header, Scan-to-Read block). Only layout fixed by hard rule; others sampled reproducibly.
• (A2) Fill loop that converges: A staged-fill loop measures each section's fullRatio = h_content / h_card and quantizes it into five verdicts (EMPTY, SPARSE, FULL, SPILLAGE, OVERFLOW) with corresponding remediation moves. Loop terminates only when every section reads FULL (90–98%) and no figure is below 70% of its card on one axis. Three mechanisms damp oscillation: move sized by pixel delta, refusal to re-apply overshot moves, circuit breaker.
• (A3) Page too large to re-read: The refinement loop never pulls the whole poster into context; only off-target sections' verbatim source is edited.
• (A4) Figures that fill their cards: Figures sized through their height cap, with hard floor enforced.
• (A5) Faithful editable export: The PowerPoint bridge walks the live DOM node by node, reads geometry and appearance, and converts to native PowerPoint shapes (text frames, bullets, pictures, SVG, OMML equations, rounded rectangles with gradients). Equations become editable native OMML rather than flat images.

Quality gates: Mandatory deliverables gate refuses release until:

• Every section in FULL band (90–98%)
• No figure below 70% of its card on both axes
• Optional reading-comprehension gate (RRP) verifies a held-out reader can still answer questions from the poster
• PDF matches the intended fixed-size canvas

Paper2Video (Video Generator)

Turns the shared bundle into a synchronized video package with:

• Editable PowerPoint deck (video.pptx)
• Two MP4 variants: with and without burned-in subtitles
• Narration-aligned visual highlights (spotlight/laser cues)
• Timeline sidecar for downstream navigation

Design requirements addressed:

• (B1) Duration planning before rendering: Narration shaped before TTS, then checked after render. Large errors trigger rewriting, not post-hoc cutting.
• (B2) Attention guidance: Visual-cue resolver combines script, word timings, slide geometry, and authored anchors to produce timed highlights.
• (B3) Caption delivery: Two MP4 variants: one with burned-in subtitles, one without (for interactive reel to avoid double subtitles).
• (B4) Addressable video: timeline.json maps each paper section to audio window, subtitle cues, slide frame, and visual cues. Paper2Reel reads this directly.
• (B5) Deterministic media checks: Package gate checks existence, playability, audio stream, duration tolerance, non-empty subtitle sidecar, highlight coverage, blank frames, text overflow, etc.

Paper2Blog (Blog Generator)

Produces two Word .docx files from the shared bundle:

• Chinese article (WeChat public-account register)
• English article (research-blog register)

Design requirements addressed:

• (C1) Single evidence base: One evidence map records hook, problem, method components, claims, results, limitations, links, figure roles. Both language drafts read from this map.
• (C2) Register control: Separate language-specific editorial style guides control tone at generation time.
• (C3) Shared figure set: 3–7 evidence visuals selected and placed near the paragraph that explains them.
• (C4) Editable DOCX: Fixed filenames, embedded media, stable fonts, editor-friendly formatting.
• (C5) Layout as rendered artifact: Gate inspects DOCX for underfilled images, near-blank pages, orphan tails, and large bottom whitespace (using rendered page images in strict mode).

3. Convergent Layer: Paper2Reel

Binds the three artifacts into a self-contained interactive HTML viewer. Poster-first design: first screen is the generated poster. Double-clicking a section opens a modal with video (left) and blog content (right) with draggable splitter. Slide thumbnails below video allow seeking. Language switch toggles between English and Chinese blog blocks.

Content alignment: An alignment sidecar maps each canonical section ID to poster block, slide targets, video start/end times, subtitle tracks, slide thumbnails, and blog blocks. The viewer reads this sidecar instead of guessing.

Quality gates: Both static file checks and browser-based validation (checks poster-first loading, hover, modal opening, split-pane layout, subtitle toggling, slide seeking, video seeking, downloads, and blog rendering using headless browser with range-capable local server).

Empirical Validation / Results

Benchmark and Setup

• Dataset: 100 papers from the Paper2Poster benchmark.
• Comparison families:
  • Single-shot frontier LLMs: Claude-4.8 Opus, GPT-5.5, Gemini-3.1 Pro (prompted once for full poster)
  • Prior poster pipeline systems: Paper2Poster tool, P2P, PosterGen (under their released configs, reproduced with Claude Code)
  • ResearchStudio-Reel in two settings: Claude Code (claude-opus-4.8) and Codex (gpt-5.5)
  • Author ground-truth posters
• Scoring: Two held-out VLM judges (claude-opus-4.8 and gpt-5.5) on a six-criterion aesthetic/information rubric (1–5 scale). PaperQuiz reading-comprehension probe (answer accuracy by exact match).

Key Quantitative Results (Table 1)

• ResearchStudio-Reel (Claude Code) wins every aesthetic and information sub-criterion among automated systems, and surpasses authors' own posters on aesthetics (3.52 vs. 2.94) and information (3.86 vs. 3.63).
• On a per-paper basis, wins overall score on 84–93% of papers and loses outright on at most four.
• Ablation results:
  • The skill machinery adds 0.76 on aesthetics and 1.14 on Layout compared to single-shot Claude-4.8 Opus.
  • Swapping harness and model (Claude Code vs. Codex) retains most quality (3.52 vs. 3.36 aesthetic), confirming the skill architecture is the primary driver.

Analysis of PaperQuiz Trade-off

PaperQuiz (reading comprehension) rewards text density, while aesthetics reward selective, legible layout. P2P leads on Quiz Detail (75.40%) because its full-height portrait canvas packs nearly verbatim prose, but has low aesthetic scores. Author posters are the opposite: clean but discard detail. ResearchStudio-Reel sits deliberately between poles—its fill loop packs each column to target density (90–98%) rather than exhaustion.

Capability Coverage (Tables 2 and 3)

Table 2 (Video) and Table 3 (Blog) compare against representative product categories and research systems. ResearchStudio-Reel is the only system that:

• Produces editable PPTX deck with narration
• Delivers timed highlight cues with duration control
• Provides bilingual (CN/EN) delivery with embedded figures and layout-aware DOCX repair

No prior system ships all three editable dissemination artifacts from a single run.

Operational Cost (Table 4)

• The full four-artifact bundle takes ~89 minutes and ~2.6M input / ~276K output tokens per paper.
• Most traffic is cached-context re-reads (billed at ~10% of fresh input).
• Poster and video are heaviest skills; blog is driven by bilingual QA gate.
• Paper2Reel binding is comparatively cheap.
• Three generators can run in parallel, reducing wall-clock time significantly.
• When LaTeX source is available, Paper2Assets runs substantially faster (bypasses PDF figure extraction).

Theoretical and Practical Implications

Theoretical Implications

• Last mile as architectural problem: The paper argues that dissemination automation is best built as a composition of skills with shared extraction, not as monolithic per-artifact systems. This is a principled architectural contribution.
• Categorical quality gates: Replacing continuous soft scores with hard pass/fail gates on geometric density (fill ratio) provides deterministic convergence and diagnosability, avoiding the plateau problem of learned preference scores.
• Measured-fill loop: The staged-fill loop with quantized verdicts and bounded step size offers a reproducible, engineerable alternative to black-box agent optimization.

Practical Implications

• For authors: Collapses the three-day manual last mile into a ~1.5-hour automated session. Artifacts are editable, so authors can polish rather than regenerate.
• For labs/orgs: Can be wired into publication intake pipelines. Editable outputs make handoff to non-author editors practical. Bilingual blog covers cross-region outreach.
• For education: Enables paper-of-the-week briefing packs with lecture slides, talk video, and blog post from a reading list.
• Open-source availability: Project at https://aka.ms/ResearchStudio.

Limitations Acknowledged

1. Evaluation proxy-bound: The aesthetic rubric and PaperQuiz pull in opposite directions. Neither proxy measures actual human understanding. The fill loop optimizes geometric density, not comprehension.
2. Generative gap: Pipeline reuses only existing paper figures; cannot draw bespoke method/overview diagrams that human designers create. Closing this gap requires faithful figure synthesis, which reintroduces hallucination risk.

Conclusion

ResearchStudio-Reel demonstrates that the last mile of research dissemination is best built as a small composition of skills: thin agent-readable contracts that share one upstream extractor and gate their output on categorical fill verdicts rather than on soft preference scores. The instantiation produces a print-ready poster, a synchronized talk-video package, and a bilingual Word blog from one paper PDF inside a single Claude Code or Codex session. Quantitative results show the poster quality leads every prior automated system and single-shot frontier LLM while surpassing authors' own posters on aesthetics. The pattern is extensible to any dissemination target with a deterministic render and a categorical fill verdict. Future work should focus on closing the evaluation loop on human reading-and-recall signals and on faithful figure synthesis to match the generative capabilities of human designers.