PERELMAN — AI4X-AC 2026 Poster: English Text Draft

Auto-prepared by Feanor, 2026-03-28 00:10 MSK
For use at AI for Accelerated Science (AI4X-AC 2026), Singapore, June 15–19
Submission #316 | Poster session

TITLE

PERELMAN: Automated Quantitative Meta-Analysis of Scientific Literature via Vision-Language Models

Daniil Merkulov, [co-authors TBD]
Skoltech / AI for Science (Sber)

ABSTRACT / TAGLINE (30 words max, for header)

PERELMAN extracts quantitative data from figures in hundreds of research papers automatically — no manual labeling, no domain expert instructions required.

SECTION 1: PROBLEM

Title: The Data Extraction Bottleneck in Scientific Literature

Scientific literature grows at >2 million papers per year. For materials science, drug discovery, and physics, critical experimental results are not in tables or abstracts — they live in figures. A typical meta-analysis of 500+ papers requires:

• Months of manual expert effort
• Reading, interpreting, and digitizing hundreds of plots
• Error-prone transcription of numerical values

Motivating example: A systematic review of NMC811 cathode materials required processing 548 papers with 950+ individual experiments — this is a months-long task for a PhD student.

Key insight: ~40–60% of quantitative results in experimental papers appear only in figures, not in text or tables. Existing tools (Elicit, SciSpace) perform Q&A over text — they miss figure-embedded data entirely.

SECTION 2: METHOD

Title: Pipeline Overview

PERELMAN is a multi-stage VLM pipeline for automated figure-based data extraction:

PDF Corpus (548 papers)
      ↓
 docling  (PDF → Markdown + extracted figures)
      ↓
 Mosaic Detector (VLM)
 → rows × cols, panel labels, target panel, rotation
      ↓
 Panel Splitter + Score Ranker
 → caption context (±1000 chars) for keyword matching
      ↓
 VLM Extractor (on target panel only)
 → voltage_range, C_discharge, C_charge, CE%
      ↓
 Text Fallback (LLM reads paper text if VLM fails)
      ↓
 Structured Database (JSON sidecar per paper)

Key design choices:

1. docling for PDF processing — preserves table structure and formula layouts better than PyMuPDF; outputs clean Markdown
2. Two-stage VLM: Mosaic Detector first identifies panel layout (rows × cols) and target panel letter (a/b/c/d); Extractor then receives only the cropped target panel → reduces noise, lowers cost
3. Context window (obvyzka): ±1,000 characters around each figure mention (caption + in-text references) is injected into the extraction prompt → improves domain-specific disambiguation
4. Zero-shot prompting: domain-specific (NMC811 synonyms, baseline-series disambiguation) but no few-shot examples required
5. Text fallback: when VLM returns no value, LLM reads the paper text → increases coverage

Cost: ~$0.01–0.05 per paper. Processing 548 papers ≈ $5–25 total — negligible vs. months of manual work.

SECTION 3: RESULTS

Title: Evaluation on 24-Paper Ground-Truth Set

We manually annotated 24 NMC811 papers (ground truth) and compared PERELMAN’s extracted values:

Accuracy progression (ablation):

Scale: 548 NMC811 papers processed → 950+ structured experimental records in the database.

SECTION 4: COMPETITIVE LANDSCAPE

Title: Where PERELMAN Fits

Key differentiator: PERELMAN extracts numbers from figures — the data source that text-based RAG systems cannot access.

                figure-based
                data extraction
                      ↑
                  PERELMAN

text-based ←──────────┼──────────── figure-based
(Elicit, SciSpace)    │
                      ↓
               literature writing
           (AutoResearchClaw, CLAW)

SECTION 5: APPLICATIONS

Title: Beyond Batteries — A General Framework

The PERELMAN architecture is domain-agnostic. To adapt to a new domain: write one extraction prompt describing the target graph types. Tested on NMC811 as proof-of-concept.

Other domains:

• Drug discovery: IC₅₀/EC₅₀ from dose-response curves; binding affinity from competition assays
• Physics: critical temperatures, band gaps, phase transition points from experimental spectra
• Quantum materials: magnetization curves, conductance vs. gate voltage
• General pattern: any field with characteristic plot types → one domain prompt → process entire literature corpus

SECTION 6: DEMO / CONCLUSION

Demo: NMC811 gigascheme — interactive visualization of 548-paper meta-analysis database. Search by synthesis method, voltage range, authors. Browse extracted figures with overlaid values.

Conclusions:
1. VLM pipelines can automate quantitative meta-analysis at scale with high accuracy (83% voltage, 74% capacity precision)
2. Two-stage mosaic detection is critical for multi-panel figures (~50% → 80% improvement)
3. Context window injection significantly improves extraction quality (~50% → 71% → 74%)
4. Cost is negligible ($5–25 for 548 papers) vs. manual effort (months)

Future work:
- Cross-domain generalization (extend beyond NMC811)
- Active learning loop: uncertain extractions → human-in-the-loop feedback
- Integration with synthesis parameter databases for ML-driven materials discovery

CONTACT / QR CODE

[QR code → GitHub repo or demo URL]

Daniil Merkulov | daniil.merkulov@skoltech.ru
Skoltech / AI for Science, Sber

POSTER PRODUCTION NOTES

Format: A0 portrait (841 × 1189 mm) or A1 — confirm with AI4X-AC organizers
Figures needed:
1. Pipeline diagram (adapt from README.md)
2. fig2_accuracy_bars.png — accuracy bar chart ✅ ready
3. fig1_scatter_capacity.png — GT vs extracted scatter ✅ ready
4. Competitive landscape table (above)
5. Capacity distribution histogram — needs 548-paper DB access (currently not local)

Early bird registration deadline: March 31, 2026 (S$1,100 academic)
Normal rate: S$1,200 (after March 31)

Draft prepared by Feanor worker 2026-03-28 00:10 MSK.
Based on: perelman_poster_outline.md, eval metrics, codebase analysis.