```html

AI in Biomedical Research: A New Benchmark for Multimodal Reasoning

Biomedical research increasingly relies on complex datasets encompassing various modalities such as images, text, and numerical data. Effective analysis and interpretation of this data requires advanced multimodal reasoning capabilities. Artificial intelligence (AI), particularly multimodal large language models (MLLMs), offers the potential to accelerate scientific discovery by automating and supporting the analysis of large and complex datasets. A new benchmark called MicroVQA aims to evaluate and improve the capabilities of AI models in this domain.

MicroVQA: A Benchmark for Microscopy

MicroVQA is a Visual Question Answering (VQA) benchmark specifically designed for the challenges of microscopy-based research. It consists of 1,042 multiple-choice questions curated by biology experts, covering various microscopy modalities. The focus is on three core competencies essential for scientific workflows: expert understanding of images, hypothesis generation, and experiment proposal. MicroVQA ensures that the questions and answers reflect real-world scientific practice, thus enabling a practical evaluation of AI models.

Challenges in Creating Multiple-Choice Questions

The developers of MicroVQA encountered a particular challenge in creating the benchmark: standard methods for generating multiple-choice questions often led to linguistic shortcuts that allowed AI models to guess the correct answer through purely linguistic cues, without actually understanding the image. To address this issue, a two-stage pipeline was developed. In the first step, an optimized LLM prompt structures question-answer pairs into multiple-choice questions. In the second step, an agent-based "RefineBot" revises these questions to remove linguistic shortcuts and ensure that the correct answer can only be found through the combination of image understanding and logical reasoning.

Benchmarking Current AI Models

Initial tests with state-of-the-art MLLMs on MicroVQA show that the best performance is at 53%. Interestingly, models with smaller LLMs perform only slightly worse than the top models. This suggests that language-based reasoning is less challenging in this context than multimodal reasoning, which requires the integration of image and text information. Furthermore, it was found that training with scientific articles improves the performance of the models.

Error Analysis and Future Research

A detailed analysis of the AI models' answers, particularly the so-called "chain-of-thought" answers, reveals that perceptual errors occur most frequently, followed by knowledge errors and overgeneralization errors. These findings highlight the challenges in multimodal scientific reasoning and the importance of benchmarks like MicroVQA for the advancement of AI-driven biomedical research.

MicroVQA offers the research community a valuable tool to advance the development of AI models for biomedical image analysis. The results of the benchmark tests provide important insights into the strengths and weaknesses of current MLLMs and lay the foundation for future research to improve multimodal reasoning in scientific applications. Through close collaboration between AI experts and biologists, the potential of AI for biomedical research can be fully realized.

Bibliographie: Burgess, J., et al. "MicroVQA: A Multimodal Reasoning Benchmark for Microscopy-Based Scientific Research." arXiv preprint arXiv:2503.13399 (2025). jmhb0.github.io/microvqa. huggingface.co/datasets/jmhb/microvqa. huggingface.co/papers. cvpr.thecvf.com/Conferences/2025/AcceptedPapers. arxiv.org/abs/2502.16033. openreview.net/forum?id=eRleg6vy0Y&referrer=%5Bthe%20profile%20of%20Serena%20Yeung-Levy%5D(%2Fprofile%3Fid%3D~Serena_Yeung-Levy1). arxiv.org/abs/2403.00231. aclanthology.org/2024.findings-emnlp.904.pdf. openaccess.thecvf.com/content/CVPR2024/papers/Yue_MMMU_A_Massive_Multi-discipline_Multimodal_Understanding_and_Reasoning_Benchmark_for_CVPR_2024_paper.pdf. huggingface.co/papers/2410.10783. ```