π²: AI Data Pipeline Boosts LLM Long-Context Reasoning

Researchers developed π², a structured data curation pipeline that improves large language models' ability to reason over long contexts by generating high-quality QA pairs from Wikipedia tables and verified reasoning traces. Fine-tuned models showed consistent gains of +2.7% to +4.3% across benchmarks, with potential for self-distillation. The open-source approach demonstrates how structured reasoning data can enhance AI reasoning capabilities.