Sep 2026· Open Conference Proceedings· 0 citations· 22 references
TL;DR
A cross-task context-injection pipeline that prepends the structured outputs of earlier tasks to downstream prompts at inference time, grounding summarizations and VQA answers in explicitly extracted values rather than visual estimation, without additional training is proposed.
Abstract
We describe our system for the ICDAR 2026 Sci-ImageMiner competition, covering figure panel classification (Task 1), data extraction (Task 2), summarization (Task 3), and visual question answering (Task 4) on scientific figures from atomic layerdeposition and etching literature. For Task 1, we progress from CNN and zero-shot VLM baselines through QLoRA fine-tuning of Qwen2.5-VL-7B-Instruct, finding that short disambiguation descriptions injected only for the most confusable class pairs improved fine-grained classification without degrading well-separated classes; a six model weighted ensemble achieved F1=0.8059 on the CodaBench evaluation phase. For Tasks 2–4, we fine-tune Qwen2.5-VL independently for each task and additionally propose a cross-task context-injection pipeline that prepends the structured outputs of earlier tasks to downstream prompts at inference time, grounding summarizations and VQA answers in explicitly extracted values rather than visual estimation, without additional training.
This paper presents a solution to the Sci-ImageMiner Competition: Task 1 (classification) at ICDAR 2026. The objective is to accurately classify individual scientific figure panels from Atomic Layer Deposition and Etching (ALD/E) literature into a domain-specific taxonomy. We formulate this challenge as a conditional t...
Soumyajyoti Mohanta, Shresth Kasyap, Sasmit Shashwat et al.· Open Conference Proceedings· 0 citations
This work presents Skill-Based Agentic DocMiner, a unified agentic workflow with three recurring components: multimodal context enrichment from captions and paper text, iterative skill refinement using development-set feedback, and collaborative verification by specialist reviewers.
Shu-Tao Gong, Jing Huang· Open Conference Proceedings· 0 citations
Scientific figures encode critical experimental findings that are often not fully captured in textual descriptions. Despite recent advances in MLLMs, their performance on domain-specific scientific figures remains limited due to insufficient grounding in relevant context. In this work, we focus on in Task 3 (summarizat...
Xiaole Lv, Wei Wu, Yang Tan et al.· Open Conference Proceedings· 0 citations
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...
Vassilis Sioros· Open Conference Proceedings· 0 citations
Results show that the two subtasks behave differently in practice: satirical headline detection is better handled by prompting a large instruction-following model, while LLM-generated joke detection benefits more from supervised learning over dataset-specific stylistic and length-related cues.