Vassilis Sioros at Sci-ImageMiner 2026 Tasks 2, 3, and 4

Abstract

In this work, we address two core challenges in the ALD-E domain: auxiliary evidence construction via summarization and data table extraction, and visual question answering (VQA). Our approach first generates auxiliary summaries and structured tables to support test-time VQA prompting, using Qwen3.5-9B vision-language models with parameter-efficient fine-tuning (LoRA). The model operates on cropped subfigures alongside sliding-window document context, including captions and neighboring text blocks. It then answers Yes/No, Paragraph, Factoid, and List questions under a shared training and inference setup. This staged design reveals three key observations: (1) summarization performs strongly across both standard and contextual evaluation metrics; (2) table extraction achieves competitive structural accuracy; and (3) overall VQA performance is moderate. Yes/No questions appear largely saturated, while Factoid and List questions remain the primary bottlenecks. Overall, our findings highlight the effectiveness of staged multimodal evidence integration and answer-type-aware optimization. Code is available at https://github.com/insane-group/staged-qwen3.5-scivqa.