Research
156d agoSynthetic Mixed Training Surpasses RAG
Researchers introduced Synthetic Mixed Training, a method that fuses synthetic question‑answer pairs with synthetic documents to exceed Retrieval‑Augmented Generation (RAG). Scaling data volume and generator quality yields log‑linear gains, outpacing RAG by 2.6 % on the global QuaLITY benchmark for long‑document comprehension.
The Signal
This shows that orchestrated synthetic data can unlock higher knowledge acquisition in language models worldwide.