Berkeley AI Team Unveils T³ Corpus, Boosting RAG Performance by 43.9%

The Berkeley AI team has introduced the T³ corpus, a new dataset designed to enhance retrieval-augmented generation (RAG) in reasoning tasks. Built on chain-of-thought principles, the T³ corpus has demonstrated a significant impact on performance, achieving a 43.9% improvement in the AIME 2025-2026 mathematics competition. This development underscores the importance of corpus quality in optimizing RAG methods for complex reasoning tasks.