Team TeleOCR-VL’s Solution for Sci-ImageMiner 2026: Competition on Scientific Chart Understanding at ICDAR 2026

Abstract

This report presents Team TeleOCR-VL’s system for Tasks 2 and 3 of the ICDAR 2026 Sci-ImageMiner Competition. Sci-ImageMiner focuses on scientific chart understanding in real ALD/ALE research papers, requiring systems to extract structured data from complex figures and generate summaries consistent with scientific context. To address the challenges of complex layouts, dense numerical information, strong structural dependencies, and domain specific visual patterns, we build a structure-aware multimodal OCR system based on TeleOCR-VL. The system combines large-scale synthetic data augmentation, structure-content decoupled training, a vision-plus-context modeling strategy, and multi-model voting. Experimental results show that our system achieves aweighted score of 41.81 in Task 2 Data TableExtraction, ranking first, and a weighted score of 0.54 in Task 3 Chart Summary Generation, ranking third. Theseresults indicate that a specialized OCR-VL system designed for scientific chart understanding is more stable than general-purpose vision-language models in table structure recovery, numerical extraction, and chart summary generation.