PERELMAN — AI4X-AC 2026 Speaker Notes (Poster Session)

Conference: AI4X-AC 2026, Singapore, June 15–19
Poster: Submission #316
Pitch duration: 2–3 minutes
Prepared: Feanor worker, 2026-04-10

Opening Hook (15 sec)

“How long did your lab’s last meta-analysis take?”

Pause. Let them answer or react.

“Ours covered 548 papers. Manually, that’s 3–6 months of a PhD student’s time — just extracting numbers from graphs. We cut it to a few hours. That’s PERELMAN.”

The Problem (30 sec)

Point to the problem panel on the poster.

“Scientific literature is growing exponentially — over 2 million papers per year. The problem isn’t finding papers. It’s that 40–60% of the key quantitative results only exist inside figures — not in text, not in tables. Standard RAG systems, LLM Q&A, Elicit — they all miss this data completely.

For our domain — NMC811 lithium-ion batteries — a single meta-analysis by Savina and Abakumov covered 548 papers and 950+ experiments. Each datapoint was extracted by hand.”

The Method (45 sec)

Point to the pipeline diagram on the poster.

“PERELMAN is a four-step pipeline:

First — we convert PDFs to structured Markdown and images using docling, which preserves figure references and captions.

Second — our Mosaic Detector, a custom VLM prompt, identifies multi-panel figures. It predicts the grid layout — say, 2×2 — and tells us which panel contains the target data.

Third — the VLM Extractor reads the target panel. The key innovation here is context-aware prompting: we feed the model ±1000 characters of text surrounding every figure mention — the caption plus inline references. This is what improved our accuracy from ~50% to 80%+.

Finally — if the figure extraction fails, we fall back to text extraction.”

The Results (30 sec)

Point to the accuracy figures on the poster.

“On our 24-paper ground-truth evaluation set:

— Capacity MAPE: 4.6% — that’s excellent precision for automated extraction.
— Voltage accuracy: 95% for the upper bound, 86% for the lower bound.
— Coverage: 83% of papers successfully processed.

Compare to baseline VLM without our mosaic detection and context prompting — we see a 50%+ improvement.”

Unique Value and Scope (20 sec)

“The architecture is domain-agnostic. The only domain-specific component is the extraction prompt — a description of what you’re looking for and what your graphs look like. For a new domain, that’s an hour of work to adapt.

This pipeline scales to any field where quantitative data lives in figures: drug IC50 values, critical temperatures in condensed matter, binding affinities in protein design.”

Close (15 sec)

“We’ve turned a 3-month manual process into a few hours of automated extraction. The database from 548 papers is now available for downstream modeling — regression, correlation analysis, materials design.

I’m happy to walk through any part of the pipeline in detail, or discuss how this might apply to your domain.”

Anticipated Questions — Quick Answers

“How does it generalize to other domains?”

“The pipeline is domain-agnostic. You write one extraction prompt describing your graph types — that’s the only domain-specific piece. We’ve focused on NMC811 as proof-of-concept; next step is generalization experiments.”

“Why not just use GPT-4o on full PDF pages?”

“Two reasons. First, cost — we extract one targeted panel rather than processing full pages. Second, precision — targeted panel + domain-specific prompt significantly outperforms generic page description. And mosaic detection is a separate non-trivial problem that generic models don’t handle.”

“What about tables? Do you extract those too?”

“Docling already handles tables excellently — that’s actually the easy part. The hard part, and our contribution, is figure-based extraction. About 40–60% of values in this domain appear only in graphs.”

“How much does it cost to run?”

“Roughly $0.01–0.05 per paper via OpenRouter VLM APIs. For 548 papers, total cost was under $25. Negligible compared to months of manual annotation.”

“Is the code available?”

“Yes — github.com/art-int-4-science/perelman. Happy to discuss collaboration for domain adaptation.”

Poster Layout Reference (quick orientation for visitors)

Top-left: Problem (exponential literature growth, data locked in figures)
Top-center: Method (pipeline diagram)
Top-right: Results (accuracy plots)
Bottom-left: Competitive landscape table
Bottom-right: Applications (batteries, drugs, physics)
Bottom-full: Demo — NMC811 gigascheme visualization

Speaker notes by Feanor worker, 2026-04-10. Based on poster outline, abstract, and evaluation metrics (MAPE=4.6%, voltage 95%/86%, coverage 83%).