Multimodal AI models can answer complex questions about data in images and documents with startling accuracy, but a new research paper from Arizona State University reveals a critical flaw. The study introduces ViTaB-A, a benchmark designed to test if these models can cite the exact source of their answers within a table. The results show that even when an AI provides the correct answer, it frequently fails to identify the specific rows and columns used to generate it.
The Black Box Problem in Data Analysis
For AI to be useful in professional settings like finance, research, or compliance, getting the right answer is only half the battle. Users must be able to verify where the information came from, a process known as attribution. Without it, the AI is a 'black box' whose outputs cannot be trusted for high-stakes decisions, as reported in the academic paper titled "ViTaB-A: Evaluating Multimodal Large Language Models on Visual Table Attribution."
This lack of transparency creates a significant barrier to enterprise adoption. An analyst needs to know if an AI's conclusion is based on the correct data points or if it's a 'hallucination'—a confident but fabricated response. This research shows that an AI can be confidently correct in its answer but completely wrong about its justification.
Putting Models to the Test with ViTaB-A
The Visual Table Attribution Benchmark (ViTaB-A) was created to systematically measure this crucial capability. Researchers evaluated several leading multimodal Large Language Models (mLLMs) against a variety of structured data formats that professionals use every day.
Key formats tested in the benchmark include:
- Image-based tables: Screenshots of financial reports or spreadsheets.
- Markdown tables: Commonly found in technical documentation and code repositories.
- JSON data: A standard format for data interchange in web applications.
The benchmark tests models on their ability to not only answer a question based on the table but also to precisely cite the cell-level evidence supporting that answer. The findings indicate a widespread weakness in this fundamental skill across the models tested.
For more deep dives into AI research and what it means for the industry, subscribe to the AI Breaking Wire newsletter. We deliver essential analysis to over 50,000 AI professionals and developers weekly.
Why It Matters
The ViTaB-A benchmark is a vital reality check for the AI industry. As companies rush to deploy multimodal AI for data analytics, this research highlights that model explainability and trustworthiness cannot be an afterthought. Without reliable source attribution, AI-powered data analysis remains a high-risk tool unfit for mission-critical applications. This work provides a clear metric for developers to improve the transparency and accountability of future models, paving the way for AI that can not only answer questions but also show its work.