Polish Medical VQA Reveals Vision-Language Models Underutilize Visuals

Jakub Pokrywka, {\L}ukasz Grzybowski, Antoni Lasik, Marek Kubis, Jeremi Ignacy Kaczmarek, Wojciech Kusa· August 14, 2026 View original

Key takeaways

  • New Polish medical VQA benchmark highlights challenges for current vision-language models.
  • Models often underutilize visual evidence, relying more on text and answer choices.
  • Robust visual grounding is crucial for reliable AI in specialized medical domains.
  • Performance gaps exist between AI and human experts in complex medical VQA tasks.

Who benefits

HealthcareMedical ImagingAI DevelopmentEdTech

Summary

A new Polish medical visual question answering (VQA) benchmark, derived from physician and dentist certification exams, shows current vision-language models struggle with visual evidence. Models perform better on text-only questions and often derive answers from choices alone, indicating a lack of robust visual grounding.

Researchers have introduced a new benchmark for medical visual question answering (VQA) specifically tailored for the Polish language. This dataset uses questions from Polish Board Certification Examinations for medical professionals, covering a wide range of specialties and visual content. The goal is to assess how well vision-language models can interpret medical images in conjunction with text. Initial evaluations of various models, including Polish-specific, general-purpose, and commercial options, reveal significant challenges. The top-performing model achieved 79% accuracy, but most models fell short of human performance, especially on image-dominant questions. A key finding is that models tend to rely more heavily on the question text than on the visual information, often achieving above-chance accuracy even without the image, suggesting a weakness in visual grounding.

Why it matters

Professionals developing AI for healthcare need to understand current model limitations, especially in specialized domains like medical VQA, to build more reliable and visually grounded diagnostic or assistive tools.

How to implement this in your domain

  1. 1Prioritize visual grounding: Develop and train models with explicit mechanisms to ensure visual information is deeply integrated and critical for decision-making, not just supplementary.
  2. 2Augment training data: Create diverse datasets that emphasize the importance of visual cues, potentially by generating adversarial examples where text alone is misleading.
  3. 3Implement explainability tools: Use techniques to visualize which parts of an image and text a model is attending to, helping diagnose visual underutilization.
  4. 4Benchmark with domain-specific data: Regularly test models against specialized benchmarks like this Polish medical VQA to ensure performance translates to real-world clinical scenarios.

Original post by Jakub Pokrywka, {\L}ukasz Grzybowski, Antoni Lasik, Marek Kubis, Jeremi Ignacy Kaczmarek, Wojciech Kusa

"arXiv:2608.12928v1 Announce Type: new Abstract: We introduce a Polish-language medical visual question answering (VQA) benchmark, built from Polish Board Certification Examination questions for licensed physicians and dentists pursuing specialist certification. The benchmark comp…"

View on X

Originally posted by Jakub Pokrywka, {\L}ukasz Grzybowski, Antoni Lasik, Marek Kubis, Jeremi Ignacy Kaczmarek, Wojciech Kusa on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI Engineering & DevToolsAI ResearchAI Investing

FlowLOB Generates Realistic, Controllable Limit Order Books Efficiently

This paper introduces FlowLOB, a conditional flow-matching generator for Limit Order Book (LOB) trajectories that offers realistic market dynamics, efficient sampling, and controllable scenario generation, outperforming existing agent-based and deep generative simulators. FlowLOB achieves high fidelity with significantly fewer computational steps than diffusion models and transfers effectively to unseen instruments.

Zhuohan Wang, Andreea Bacalum, Ollie Olby, Carmine Ventre, Namid StillmanAug 14, 2026
AI Engineering & DevToolsAI Research

Auditing Reveals Bias in Neural Combinatorial Optimization Benchmarks

This paper audits test-time budget allocation in Neural Combinatorial Optimization (NCO) solvers, revealing that reported gains from non-uniform sampling often stem from "sampling luck" rather than true allocation benefits on in-distribution data. It proposes a correction procedure and demonstrates real gains under distribution shift, emphasizing the need for rigorous evaluation.

Jinhyung BaeAug 14, 2026
AI Engineering & DevToolsAI Research

Diffusion Models Solve Mixed-Integer Optimization Problems Faster

This paper introduces Constrained Graph Diffusion (CGD), a novel learning-based approach that uses a graph-based generative diffusion model to approximately solve mixed-integer optimization problems (MIPs). CGD integrates a training-free feasibility projection operator into the diffusion process, significantly improving solution quality and feasibility while achieving substantial speedups over traditional numerical solvers.

Vincenzo Di Vito, Mehdi Taghizadeh, Deepjyoti Deka, Kaarthik Sundar, Ferdinando FiorettoAug 14, 2026