Research
156d agoNew Synthetic Training Beats Retrieval‑Augmented Models
Researchers unveil a synthetic mixed training method that outperforms retrieval‑augmented generation, a benchmark technique. By combining synthetic question‑answer pairs with synthetic documents, the approach yields log‑linear gains as data volume grows.
The Signal
On the global QuaLITY benchmark, the model surpasses RAG by 2.6 % relative, underscoring the power of scalable synthetic data for large‑scale language understanding.