Strategy/projects/files/perelman_ai4xac2026_poster_abstract.md
+

perelman_ai4xac2026_poster_abstract

PERELMAN — AI4X-AC 2026 Poster Abstract

Status: Draft (auto-generated by Feanor, 2026-03-29)
Submission #316 | AI4X-AC 2026, Singapore, June 15–19

Abstract (English, ~280 words)

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

Quantitative meta-analysis of scientific literature is a critical bottleneck in accelerated discovery: a researcher aiming to survey 500+ papers in a domain must manually extract numerical results from figures, tables, and captions — a process that takes months and is prone to human error. A key challenge is that 40–60% of reported experimental values appear only in graphs, not in text or tables, making text-based retrieval methods inadequate.

We present PERELMAN, an end-to-end pipeline for automated extraction of quantitative data from scientific PDF corpora. Given a collection of papers, PERELMAN: (1) converts PDFs to structured Markdown and images using docling; (2) detects multi-panel figures using a custom VLM Mosaic Detector that predicts grid layout and target panel labels; (3) extracts target numerical values from identified panels using a domain-specific VLM Extractor with context-aware prompting (±1000-character window around figure references); and (4) falls back to text extraction when figure-based extraction fails.

We evaluate PERELMAN on a corpus of 548 NMC811 lithium-ion battery papers (Savina & Abakumov meta-analysis), extracting first discharge capacity and voltage range. On a 24-paper ground-truth evaluation set, PERELMAN achieves MAPE = 4.6% for capacity extraction (n=20 matched pairs), 95% accuracy for voltage upper bound (±0.1 V, n=21), and 86% accuracy for voltage lower bound. Coverage (fraction of papers with successful extraction) is 83% (20/24). These results represent a 50%+ improvement over a baseline VLM without our mosaic detection and context-aware prompting modules.

PERELMAN is domain-agnostic at the architectural level: the extraction prompt is the only domain-specific component, enabling rapid adaptation to new fields (drug discovery, condensed matter physics, etc.). Our pipeline reduces quantitative literature curation from months to hours, enabling data-driven modeling at the scale of entire research domains.

Keywords

scientific literature mining, vision-language models, figure extraction, quantitative meta-analysis, materials science, NMC811, multi-panel detection, automated data curation

Author Contributions (draft, for Quantum-style submission if needed)

  • D. Merkulov: Pipeline architecture, VLM prompting, mosaic detector, evaluation
  • D. Sherki: Domain validation, NMC811 ground-truth annotation
  • [Add co-authors as appropriate]

Notes for Camera-Ready

  • Poster format: likely A0 Portrait (official instructions not yet published — request via info@ai4x.cc post-registration)
  • Camera-ready deadline: typically 3–4 weeks before conference (~May 20–25, 2026)
  • Figures available: /root/Strategy/projects/files/poster_figures/
  • fig1: capacity scatter (GT vs extracted)
  • fig2: error distribution histogram
  • fig3: accuracy bar chart by metric
  • fig4: pipeline flow diagram

Generated: Feanor worker, 2026-03-29 11:05 MSK

Choose icon