Coling-UniA at SciVQA 2025: Few-Shot Example Retrieval and Confidence-Informed Ensembling for Multimodal Large Language Models

TL;DR

This paper presents a system for scientific visual question answering that combines multimodal large language models with few-shot example retrieval and confidence-based answer selection, achieving third place in the SciVQA 2025 challenge.

Contribution

It introduces a novel ensemble approach with adaptive model and few-shot setting selection based on question type, and confidence-informed answer aggregation.

Findings

Achieved third place with an average F1 score of 85.12.

Demonstrated effectiveness of confidence-based answer selection.

Showcased the utility of adaptive model selection for different question types.

Abstract

This paper describes our system for the SciVQA 2025 Shared Task on Scientific Visual Question Answering. Our system employs an ensemble of two Multimodal Large Language Models and various few-shot example retrieval strategies. The model and few-shot setting are selected based on the figure and question type. We also select answers based on the models' confidence levels. On the blind test data, our system ranks third out of seven with an average F1 score of 85.12 across ROUGE-1, ROUGE-L, and BERTS. Our code is publicly available.