Research
155d agoSynthetic Mixed Training Surpasses RAG
Researchers unveiled Synthetic Mixed Training, a new approach that blends synthetic question‑answer pairs with synthetic documents to surpass Retrieval‑Augmented Generation limits. By combining complementary signals, the method achieves log‑linear gains as data volume and generator strength grow, delivering a 2.6 % improvement on the QuaLITY long‑document benchmark.
The Signal
This breakthrough could reshape knowledge acquisition for language models worldwide.