perelman_poster_talking_points
PERELMAN — Poster Talking Points (AI4X-AC 2026, Singapore)
Составлено Феанором 2026-04-10. Конференция 15–19 июня 2026, Submission #316.
Источник: poster_outline.md + poster_abstract.md + gigascheme pipeline
ОДНА ФРАЗА (elevator pitch, 10 сек)
“We built a pipeline that reads 548 battery papers and automatically extracts quantitative data from figures — achieving 4.6% error with no manual annotation.”
5 ГЛАВНЫХ ТЕЗИСОВ
T1. Проблема — боль аудитории (открыть разговор)
Тезис: “40–60% of experimental values in materials science papers appear only in graphs — not in text. Text-based AI tools miss them entirely.”
- Конкретный пример: обзор NMC811 (Savina & Abakumov) — 548 статей, 950+ экспериментов. Вручную: месяцы работы аспиранта.
- Открытый вопрос к собеседнику: “How long does your team spend extracting data from plots?”
T2. Наш подход — 2 ключевых элемента
Тезис: “Two novel components make PERELMAN work: (1) a Mosaic Detector for multi-panel figures, and (2) context-aware prompting using ±1000-char windows around figure references.”
- Mosaic Detector: VLM предсказывает rows×cols + target panel letter → точная нарезка панели
- Context window: caption + все упоминания фигуры → LLM «знает» что ищет
- Без этих двух: ~50% accuracy. С ними: 86–95%
T3. Результаты — числа для плаката
Тезис: “On a 24-paper ground-truth set: voltage accuracy 95% (max) / 86% (min), capacity MAPE 4.6%, coverage 83%.”
- Если спросят “это хорошо?”: MAPE 4.6% для автоматической экстракции из графиков = отличный результат (human annotation error ~2–5%)
- 50%+ improvement over baseline (without mosaic + context)
T4. Чем отличаемся от аналогов
Тезис: “We’re not competing with Elicit or SciSpace — they do text Q&A. We do figure-based numerical extraction, which is a different problem.”
- Vs Elicit/SciSpace: они работают с текстом/абстрактами → не видят данные в графиках
- Vs AutoResearchClaw: они пишут статьи, мы строим базы данных
- Уникальная ниша: figure → structured number, at corpus scale
T5. Применения (для AI4X-AC аудитории)
Тезис: “The architecture is domain-agnostic: swap out the extraction prompt and PERELMAN works for drug IC50 curves, band gaps in condensed matter, phase diagrams.”
- Materials (это мы): capacity, voltage → fit regression “synthesis → performance”
- Drug discovery: IC50/EC50 из pharmacology graphs
- Physics: critical temperatures, band gaps из experimental curves
- “One domain-specific prompt = entire corpus processed in hours, not months”
ОЖИДАЕМЫЕ ВОПРОСЫ И ОТВЕТЫ
Q: What is MAPE and why 4.6%?
A: “MAPE = Mean Absolute Percentage Error. 4.6% means on average our extracted capacity is within 4.6% of the manually annotated value. For comparison, inter-annotator agreement in manual annotation is typically 2–5%. So we’re within human-level precision on most papers.”
Q: What is your model / which VLM?
A: “We use off-the-shelf VLMs via OpenRouter API — currently GPT-4o-mini class models for the mosaic detector and extractor. The key contribution is the prompting architecture and pipeline, not a custom model. Cost: ~$0.01–0.05 per paper, so 548 papers ≈ $5–25 total.”
Q: Why better than DFT / how does this relate to simulation?
A: “We don’t replace DFT — we complement it. DFT predicts properties from first principles. PERELMAN mines experimental results from literature at scale, giving you ground-truth benchmarks to validate DFT predictions, or input features for ML potentials trained on experimental data.”
Q: How does it generalize to other domains?
A: “The architecture is domain-agnostic. The only domain-specific component is the extraction prompt — which describes what types of graphs to look for and what values to extract. Writing a new prompt takes hours, not months. We demonstrated on NMC811 as proof-of-concept.”
Q: How do you handle figures with multiple panels (a, b, c, d)?
A: “That’s exactly what the Mosaic Detector solves. It takes the figure image, predicts the grid layout (e.g., 2×2), and identifies which panel contains the target data (e.g., ‘panel b’). Then we crop just that panel before extraction. This alone improved accuracy from ~50% to ~80%.”
Q: What about figures that are tables or text-heavy?
A: “We have a text fallback: if VLM extraction from the figure fails or returns low-confidence values, the pipeline falls back to reading the surrounding text. This recovers some of the 17% non-extracted papers (coverage gap from 83% to 100%).”
Q: OpenRouter cost? Is this scalable?
A: “~$0.01–0.05 per paper. 548 papers ≈ $5–25 total. Negligible vs. months of PhD student time. And costs are dropping — the same quality is getting 10× cheaper each year.”
Q: Can we try it / is it open source?
A: “Not yet public, but we’re planning to open-source after the paper. The NMC811 database is available — happy to share contacts.”
DEMO TALKING POINTS (если есть ноутбук/QR)
- Показать: gigascheme visual (548 papers, capacity distribution)
- Показать: example figure before/after (raw plot → extracted number)
- Показать: scatter GT vs extracted (fig1_capacity_scatter.png)
КОРОТКИЕ ВЕРСИИ ПО ВРЕМЕНИ
30 сек (в лифте):
“We automate quantitative meta-analysis of scientific papers. The key insight: most data is in figures, not text. We built a VLM pipeline that reads figures directly — 4.6% error on 548 battery papers, no manual annotation needed.”
2 мин (у постера):
T1 (проблема) → T2 (наш подход) → T3 (результаты) → вопрос “what field are you in?” → T5 (применения в их области)
5 мин (развёрнуто):
T1 → T2 → T3 → T4 (vs конкуренты) → T5 → предложить посмотреть demo
Подготовлено Феанором 2026-04-10 на основе poster_outline.md + poster_abstract.md