Can AI truly understand and reason with numbers the way humans do? Large Language Models (LLMs) have shown impressive abilities in various tasks, but mathematical reasoning within documents containing tables has remained a challenge. Existing approaches often rely on massive models, raising cost and privacy concerns, especially for businesses dealing with sensitive data. A new research paper introduces MATATA, a clever way to empower smaller, more manageable LLMs to solve complex math problems in tables. Imagine an AI assistant that can analyze financial reports, extract key insights, and answer complex questions without needing access to the entire internet or a supercomputer. MATATA makes this possible through a tool-assisted approach, breaking down problems into smaller, manageable steps. Instead of relying on large, expensive models like GPT-4 to provide training data, MATATA uses a 'self-improvement' method. It starts with basic instructions and learns from its own attempts, progressively refining its abilities through a weak supervision method. This innovative approach also allows MATATA to leverage the collective knowledge from different datasets, leading to improved performance across diverse mathematical tasks. The results are compelling: MATATA achieves state-of-the-art performance on benchmarks like FinQA and TAT-QA, rivaling even GPT-4 powered systems on TabMWP while using significantly smaller, locally hosted models. This breakthrough paves the way for more accessible, efficient, and privacy-preserving AI solutions for businesses and individuals. The future of AI-powered math reasoning isn't about bigger models—it's about smarter strategies like MATATA.