Microsoft Unveils Two AI‑Powered Research Automation Papers

ResearchStudio‑Idea – Automating the First Mile of Innovation

LLMs can suggest research ideas, but current systems lack a structured bridge between retrieved evidence and executable research opportunities, leading to surface novelty that is often hollow. ResearchStudio‑Idea addresses this "missing middle" by converting acceptance/rejection outcomes of top‑tier conference papers into reusable "Ideation Patterns".

Data‑Driven Pattern Extraction

The team collected 1,947 papers from ICLR/ICML/NeurIPS (2021‑2025), labeling them as Oral (1,014), High‑Cited (260), or Reject (722). Each paper was annotated with 12 innovation fields, split into two stages: basic fields (innovation approach, key step, why non‑obvious, trigger condition, reviewer praise/concern, acceptance signal, contribution type) and domain‑agnostic rewrites to enable clustering by strategy rather than topic.

Using UMAP + HDBSCAN (OpenAI text‑embedding‑3‑large, silhouette 0.584), 31 fine‑grained sub‑clusters were aggregated into 15 Level‑1 innovation patterns such as "Audit and Pivot an Assumption" and "Substitute the Operator or Representation". Each pattern is an actionable card containing applicable contexts, success conditions (derived from Oral papers), failure modes (from Reject papers), reviewer expectations, and cognitive obstacles.

Key findings include a modal combination size of k=2: 59.2% of papers employ exactly two patterns, and the specific pattern combination better predicts acceptance than the number of patterns alone.

Pattern Adjacency and Domain Coverage

Cosine similarity of pattern centroids reveals strong adjacency (e.g., "Audit and Pivot" ↔ "Reframe as a Solvable Object" with 0.956) and isolated patterns (e.g., "Design a Confound‑Isolating Diagnostic" with 0.904). Domain analysis shows broad coverage, yet effectiveness varies by field; some patterns provide strong Oral signals across multiple domains.

Experimental Evaluation

On 100 held‑out ICLR 2026 Oral seeds, blind evaluation shows IdeaSpark occupies the "high quality + competitive novelty" region, while a GPT‑5.5 baseline falls into a "novel‑but‑empty" trap. IdeaSpark scores 3.87/4 in quality (88% win rate) versus 2.56 for Opus‑4.8 and 1.00 for GPT‑5.5. Novelty scores are 2.92 (Level 3) for IdeaSpark versus 3.73 (Level 4) for GPT‑5.5, indicating higher novelty but lower substance for the latter.

ResearchStudio‑Reel – Automating the Last Mile of Research Dissemination

Existing automation tools for research dissemination suffer from three structural gaps:

G1 Repeated extraction: Posters, videos, and blogs each re‑parse PDFs, breaking cross‑product factual consistency.

G2 One‑way rendering: Outputs are static PDFs/MP4s/Markdown, preventing direct editing in PowerPoint or Word.

G3 Soft quality gate: Reliance on VLM‑as‑judge scores (e.g., 7.8/10) can mask completely missing factual blocks.

ResearchStudio‑Reel answers these by constructing the final mile as a composable set of five Skills.

Five‑Skill Architecture

A single PDF input passes through a shared extraction layer (Paper2Assets) that produces a structured abstract (nine standard sections), cleaned figures, metadata, and narration scripts. Downstream Skills consume the same asset package, guaranteeing cross‑product consistency.

Paper2Poster – Filling the Page Exactly

The main engineering challenge is converging a fixed‑size poster to a "just‑full" state; a single edit can cause sections to overflow or leave empty space, leading to oscillation. Using Paper2Assets, the system selects primary method figures and secondary figures, arranges them along four axes (column layout, visual theme, header, QR block), then enters a Staged‑fill loop.

The Staged‑fill loop quantifies each section's fill rate into five discrete verdicts (EMPTY, SPARSE, FULL, SPILLAGE, OVERFLOW) rather than a continuous score. The target band is FULL (90–98%). Each iteration edits one section, adding or removing content based on the verdict, and rejects convergence on soft scores.

Render‑time expansion dynamically stretches sections from 90% to 98% visual fill, preventing endless oscillation. The process progresses from an uneven draft (red/orange‑marked sections) to a fully filled poster (green‑marked sections with correctly sized figures).

Editable PPT bridge: Instead of pixel‑reverse engineering, the system reads DOM geometry and styles, converting them to native PowerPoint objects (text boxes, images, MathJax → OMML, CSS colors → RGBA). Authors can edit text, swap images, adjust colors directly in PPT and re‑export without rerunning the loop.

Paper2Video – Auditable Media Packages

Instead of a single MP4, the Video Skill outputs an auditable media package: a PPT master deck, highlighted semantic anchors, Edge TTS audio, subtitles, and a timeline.json mapping section IDs to audio windows, subtitles, slide frames, and highlight geometry.

Duration pre‑planning: Estimate script length before TTS and rewrite semantics to match target duration.

Attention highlighting: Spotlight laser emphasizes discussed regions.

Dual MP4 variants: video_no_subtitles.mp4 for Reel interaction and video.mp4 with burned‑in subtitles.

Timeline sidecar: timeline.json links section IDs to media assets for direct navigation.

Paper2Blog – Bilingual Editable Blogs

The Blog Skill produces two Word .docx files (English research blog and Chinese public‑account article) rather than plain text. Using the shared evidence map, the system drafts multilingual content, then runs an edit‑QA loop for fact matching, figure fitting, and page‑flow checks.

Quality gates include factual consistency (numbers, claims, figure order aligned across languages) and layout awareness (detecting near‑blank pages, orphan tails, thumbnail‑sized images).

Paper2Reel – Unified Navigation Across Three Products

Reel is not a fourth product but the interactive convergence layer for the three outputs. The poster’s first screen highlights sections on hover; double‑click opens a synchronized modal with video + subtitles + slide thumbnail on the left and the corresponding blog paragraph on the right, plus language switching.

Alignment sidecar binds poster blocks, video segments, slide thumbnails, and blog passages to the same section ID, achieving section‑level convergence.

Experimental Results

On a benchmark of 100 papers, ResearchStudio‑Reel (Claude Code) outperforms all automatic systems and single‑shot LLMs across six aesthetic/information criteria, achieving an average aesthetic score of 3.52 vs. 2.94 for the human author (5‑point scale) and an overall win rate of 84–93%.

Reel uniquely produces three editable artifacts with unified navigation. In head‑to‑head comparisons, Paper2Video and Paper2Blog check all boxes for editability, subtitles, duration control, bilingual output, DOCX export, and layout checks.

Runtime cost: the full four‑product pipeline takes ~89 minutes, consumes ~2.6 M input tokens, and the three generators can run in parallel.

ResearchStudio-Reel: Automate the Last Mile of Research from Paper to Poster, Video, and Blog
https://arxiv.org/abs/2607.04438

ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes
https://arxiv.org/abs/2607.04439