This article was developed as part of the ImageCLEF 2025 competition. We adapted the BLIP-Base image-captioning model for the Multimodal Reasoning task, integrating the SmolLM-360M model for question answering and training on the MBZUAI EXAMS-V dataset (16 724 training and 4 208 validation examples). We then conducted a prompt-ablation study using three diferent templates to evaluate their impact on answer-key accuracy, measured by case-insensitive substring matching against the correct option within the provided set of three to five answers. Finally, we analyzed the distributions of generated caption lengths.
Modern vision–language systems combine powerful image encoders with autoregressive text decoders
to perform tasks such as image captioning, visual question answering, and multimodal reasoning. Early
works such as CLIP demonstrated that contrastive pretraining on large-scale image–text pairs yields
embeddings that transfer well to downstream classification and retrieval tasks. Building on this, BLIP
introduced a dual objective of contrastive alignment and generative captioning, producing models such
as Salesforce/blip-image-captioning-base and -large that achieve state-of-the-art results
on COCO and other benchmarks [
At the same time, recent advances in compact causal language models (with less than 500 M
parameters) show that mid-scale Transformers can deliver strong generative performance under tight compute
budgets. SmolLM-360M [
The ImageCLEF 2025 Multimodal Reasoning task challenges systems to select the correct answer
from 3–5 provided options, given an image of a science exam question, covering topics from chemistry
to physics, across multiple languages [
We use the MBZUAI EXAMS-V dataset [
To generate image captions, we use the following encoder-decoder models: • BLIP-Base(Salesforce/blip-image-captioning-base) • BLIP-Large (Salesforce/blip-image-captioning-large,CLIP-ViT-L/14 backbone) These encoders extract visual features from the exam images and decode them into textual descriptions. The captions are later used as inputs for question-answering via a language model.
We assess how prompt phrasing afects caption content and downstream Accuracy. Each image is paired with one of three prompt templates: • None: The image is passed without additional text. • "A picture of": encourages concise object-focused captions.
• "Describe what you see:": encourages detailed, descriptive captions.
3.4. Model
To perform reasoning over our generated image captions, we employ a lightweight yet powerful
language model:
• SmalLM 360M - a compact transformer-based language model optimized for low-resource
inference. With only 360 million parameters and eficient deployment on hardware with as little
as 12 GB of GPU memory, it enables practical experimentation without sacrificing performance.
Despite its small size, SmolLM-360M is currently the best-performing model in its category
(sub-500M parameters). According to the Hugging Face benchmark [
Answer: This zero-shot approach allows us to simulate realistic, low-resource deployment conditions while assessing how well the model can reason over image-derived text alone.
Answer-key accuracy is the percentage of validation samples whose generated caption contains the correct option letter (A–E) as a standalone token. Formally:
accuracy = 1 ∑︁ ⊮[token() ∈ tokens()] × 100%.
=1 Here = 4208 for the validation split. We also report the distribution of caption token lengths.
Our system was oficially evaluated as part of the ImageCLEF 2025 Multimodal Reasoning task [
Unlike many participating systems that focused on individual languages, we ran our model on the entire multilingual test set, which includes science exam questions in multiple languages such as English, Bulgarian, Arabic, and others. This multilingual setup allowed us to evaluate the generalization capabilities of our lightweight models across a diverse range of inputs. Importantly, our oficial submission used the bare image input without any additional prompt text for captioning—i.e., we did not prepend instructions like “A photo of” or “Describe what you see:”. This minimal setup demonstrates the capability of our pipeline to extract useful semantic information from images alone.
And here are the results that we achieved:
• English: placed 11th with 25.20 % accuracy • Bulgarian: placed 6th with 23.50 % accuracy • Multilingual: placed 10th with 21.88 % accuracy
In other words, our strongest relative showing was in the Bulgarian track (rank 6), even though the absolute highest accuracy was in the English track.
• Caption Conciseness Correlates with Accuracy Across both BLIP-Base and BLIP-Large, the shortest outputs consistently yield the best match against the answer key. For BLIP-Base, omitting any leading prompt (“None”) produces the briefest captions (11.4 tokens) and delivers the highest accuracy (22.0%). Likewise, for BLIP-Large, the “Describe what you see:” template—despite being wordier than no prompt—actually results in the most concise captions (12.7 tokens) of the three setups and achieves the top performance (22.0%). • Prompt Wording Matters—But Only Modestly Swapping among “A photo of,” “Describe what you see:,” or no explicit prefix shifts accuracy by at most 1.6 points. In contrast, average caption lengths vary by as much as 3 tokens. This gap suggests that while prompt phrasing reliably inflates or trims verbosity, it only marginally influences the model’s ability to generate an answer–key match. In other words, template choice can nudge performance but is not the dominant factor.
We present a prompt-ablation study for BLIP-Base on ImageCLEF 2025 [
During the preparation of this work, we used OpenAI GPT-4o to assist with grammar and spelling improvements. All suggestions were reviewed and edited by the authors, who take full responsibility for the final content of the publication.