This paper proposes a "Question Reconstruction before Answering" (QRA) prompting strategy for the ImageCLEF2025 multimodal reasoning task. The method first completes missing question stems using image information, then guides the language model through step-by-step reasoning and answering, thereby enhancing the model's comprehension and reasoning capabilities. On the EXAMS-V dataset, through our investigation of diferent prompts and their impact on accuracy, we found that the QRA prompting demonstrates strong cross-lingual adaptability compared to conventional Chain-of-Thought (CoT) prompting. Experimental results show that this method efectively improves visual question answering performance without requiring OCR or additional fine-tuning, ofering a new perspective for multimodal reasoning tasks.
Visual-Language Models (VLMs) have made significant progress in multimodal understanding tasks in
recent years. CLIP[
Although these methods have made progress in image-text alignment and language generation, their reasoning processes still perform limitedly under incomplete prompts. Our approach attempts to address this shortcoming by reconstructing question text combined with CoT (Chain-of-Thought) reasoning.
Chain-of-Thought (CoT) prompting significantly enhances the reasoning capabilities of large language
models in complex tasks by guiding the model to generate intermediate reasoning steps [
cnE iV o su red la
LLM
In image-only multimodal question answering tasks,such as the EXAMS-V dataset [
Inspired by this, we propose a "complete first, then reason" strategy, the Architecture shown in Figure 1. This strategy first uses image features to guide the language model to complete the missing question information, thereby reconstructing the complete question stem; subsequently, based on the reconstruction results, a Chain-of-Thought (CoT) reasoning mechanism is introduced to enhance the model’s cross-modal reasoning ability. This method not only enhances the model’s understanding of the task context but also efectively alleviates the semantic disconnection caused by modal diferences. Specifically, our method includes two key steps: 1) Question Background Information Prompt Embedding , 2) Question-Reasoning-Answer Prompting.
In practical multimodal question answering tasks, questions often involve specific languages and subject backgrounds, with language expressions that are highly specialized and context-dependent. Especially in scenarios containing only images, language models, lacking explicit context, are prone to misunderstandings of the question stem.
To address this issue, we introduce question background information embedding. Specifically, we extract the language category (e.g., English, French, etc.) and subject labels (e.g., physics, chemistry, etc.) of the question from the image’s metadata and use them as prior knowledge prompt words to guide the language model in context modeling. This approach efectively mitigates semantic ambiguity caused by language specificity, making the model more targeted and accurate when generating completion content.
Standard Prompting There is a question in the image, please provide the correct answer.
Upon rechecking, it appears there might be a typo in the problem or the options. However, based on the calculations, the closest match is: D(135±0.23)
Chain-of-Thought Prompting A conversation between User and Assistant...
Format: <think> reasoning </think> <answer> final answer</answer> <think> Step-by-Step Solution: The balance condition for the Wheatstone bridge is … <think/> <answer>A(60±0.15) </answer>
Question-Reasoning-Answer Prompting
A conversation between User and Assistant… The problem you should complete, the reasoning process, and the answer are enclosed within the following tags: <problem> problem </problem> <think> reasoning</think> <answer> answer here </answer> <problem> During an experiment with a metre bridge, the galvanometer shows a null point when the jockey is pressed at 40.0cm using a standard resistance of 90 … </problem> <think> This is a classic Wheatstone bridge setup using a metre bridge. The balance condition is given by: …… </think> <answer> C (60±0.25)</answer>
After completing the question stem, the model still requires strong reasoning capabilities to correctly perform the question-answering task. Traditional Chain-of-Thought (CoT) prompting, which guides language models to generate intermediate reasoning steps, has achieved significant success in textual reasoning tasks. However, directly applying the CoT mechanism to multimodal question-answering tasks involving only images can lead to information confusion or insuficient semantic alignment, resulting in the model’s inability to construct coherent and clear reasoning chains.
To address this, we propose a structured "Question Reconstruction before Answering" guided prompting strategy, aiming to explicitly separate the question comprehension process from the reasoning process to enhance the model’s ability to build reasoning chains. Specifically, we design a unified prompt template that introduces the <Question>...</Question> tag to guide the model in first understanding the question before engaging in step-by-step thinking and answering. We show the efects of three types of prompts in Figure 2.
To validate the efectiveness of our proposed QRA Prompting strategy, we participated in the ImageCLEF2025 multimodal reasoning task and submitted test results for both the Multilingual Track and the English Track. Table 1 lists our performance on the multilingual test set.
In the Multilingual Track, our method ranked 6th among all participating teams, achieving an accuracy of 0.5195. Compared to the oficial baseline method (accuracy of 0.2701), our approach improved performance by 24.9%, demonstrating the strong competitiveness of our method in practical tasks. This significant improvement indicates that our proposed structured strategy of "first completing the question, then reasoning" has clear advantages in alleviating inter-modal information misalignment and enhancing cross-modal understanding.
Notably, we achieved near-top-tier performance without relying on any additional OCR modules or ifne-tuning the model for multilingual tasks. This demonstrates that QRA Prompting possesses strong robustness and excellent transfer generalization capabilities, performing stably and reliably in complex real-world multimodal reasoning scenarios.
In the English Track, we also submitted model predictions based on QRA Prompting, achieving an accuracy of 0.5371 and ranking 6th,As shown in Table 2. Our method consistently delivered strong performance across both tasks, further validating its cross-language consistency.
To systematically evaluate the contributions of each key component in QRA Prompting, we conducted ablation experiments on the English validation set of the EXAMS-V dataset. The experiments used QwenVL-2.5-32B as the base model, employed a zero-shot setting, and compared against standard Prompting and Chain-of-Thought (CoT) Prompting methods. As shown in Table 3, the standard Prompting method achieved an accuracy of 0.458, demonstrating relatively weak performance. The CoT Prompting method, which guides the model through chain-of-thought reasoning, improved accuracy to 0.548. The QRA Prompting strategy further enhanced this performance, achieving an accuracy of 0.582, which represents a 12.4% improvement over the standard method and a 3.4% improvement over the CoT method.
These results indicate that QRA Prompting not only inherits the advantages of chain-of-thought reasoning from CoT but also efectively enhances the language model’s understanding of image semantics through explicit question stem completion, significantly boosting the model’s performance in complex reasoning tasks.
In this study, we addressed the Multimodal Reasoning task of the ImageCLEF2025 Multimodal Lab. By employing the QRA strategy, we enhance the inference accuracy of models in visual question answering tasks. Our approach involves constructing QRA prompt templates and integrating contextual information. These strategies efectively address two key challenges faced by traditional Chain-ofThought (CoT) in multimodal scenarios: they alleviate the "modality gap" problem caused by relying solely on visual features and enhance the ability to reconstruct missing question text from visual data.
Evaluation results demonstrate the feasibility and efectiveness of our method, achieving an accuracy of 0.5195 on the multilingual version of the EXAMS-V test set. These findings indicate that our approach provides a viable solution for visual question answering tasks that use only visual features, contributing to the field of multimodal reasoning.
This work is supported by the Research Projects of OrdinaryUniversities in Guangdong Province under Grant2023KTSCX133, the Guangdong Basic and Applied Basic Research Foundation under Grant 2022A1515140103.
During the preparation of this work, the author(s) used deepseek-v3 in order to: Grammar and spelling check. After using these tool(s)/service(s), the author(s) reviewed and edited the content as needed and take(s) full responsibility for the publication’s content.