Markdown-KV Format Boosts AI Accuracy 37% Over CSV in New LLM Performance Study
Summary
New research reveals that switching from CSV to Markdown-KV format dramatically improves AI language model accuracy by 37%, achieving 60.7% success rate compared to CSV's 44.3% when processing employee data, though the superior format requires nearly three times more computational tokens.
Key Points
- Researchers test 11 different data formats with GPT-4.1-nano using 1,000 employee records and 1,000 questions to determine which format LLMs understand best
- Markdown-KV format achieves highest accuracy at 60.7%, significantly outperforming CSV (44.3%) and JSONL (45.0%), though it requires 2.7 times more tokens than CSV
- Results show format choice significantly impacts LLM performance, with XML (56.0%) and INI (55.7%) also ranking high, suggesting simple transformations could improve AI system accuracy