We interact with tables daily, from checking sports scores to analyzing financial reports. Tables are everywhere, compactly presenting information in a way that's easy for *us* to understand. But what about AI? Can large language models (LLMs) truly grasp the information locked within those rows and columns? New research suggests they might struggle more than we think. A groundbreaking study introduces “TabIS,” a benchmark designed to rigorously test how well LLMs can extract and reason with tabular data. Unlike previous tests relying on potentially unreliable text comparisons, TabIS uses a simple multiple-choice format. This approach presents the LLM with a table and a question, asking it to select the correct answer from two options. The results are surprising. While industry giants like GPT-4 perform reasonably well, most LLMs, including many open-source models, struggle. They often fail to understand the relationships between different cells, particularly in complex or hierarchical tables. Imagine an AI trying to summarize a financial report but misinterpreting key figures due to its limited table comprehension – the consequences could be significant. Even more concerning, the study reveals that many LLMs are easily fooled by the presence of additional, irrelevant tables, a common scenario in real-world document retrieval. They get distracted by similar-looking data, demonstrating a lack of true table understanding. One key problem area is table structure understanding (TSU). While humans can easily pinpoint information based on row and column positions, LLMs struggle to interpret this spatial arrangement. This weakness makes it difficult for them to extract precise information from tables using only positional clues. The researchers also explored ways to improve LLM performance on table reasoning tasks. They found that fine-tuning a model on a large set of table-related questions can significantly boost its abilities. However, even the fine-tuned models couldn't reach the level of GPT-4, highlighting the ongoing challenges in teaching AI to truly understand tabular data. This research underscores the importance of developing more robust methods for evaluating and improving how LLMs process tabular information. As AI becomes increasingly integrated into our lives, ensuring it can reliably interpret structured data is crucial. The next generation of table-savvy AI could revolutionize how we interact with information, automating data analysis and making complex datasets accessible to everyone. But first, we need to help them crack the code of the table.