How JD’s Vega v1 Model Dominated GLUE Benchmark, Surpassing Human Performance

In the recent GLUE benchmark for natural language processing, JD Explore Research Institute teamed up with the University of Sydney, Wuhan University and Beihang University, forming the “Dream Team” (d‑team). Their Vega v1 model achieved a top average score of 91.3, setting a new world record and surpassing competitors such as Microsoft, Facebook and Stanford.

Vega v1 ranked first in four of the nine GLUE sub‑tasks—SST‑2 sentiment classification, MRPC semantic similarity, QQP question‑pair equivalence, and WNLI coreference—and notably exceeded human performance on the SST‑2 and WNLI tasks.

The model follows a pre‑training‑fine‑tuning paradigm, leveraging efficient parallel training frameworks, billions of parameters, improved self‑supervised signals and multi‑granular sentence representations. Future upgrades will incorporate trustworthy AI features such as stability, interpretability, privacy protection and fairness.

JD Explore Research Institute has published over 520 papers at top AI conferences, won 33 international AI competition championships, holds more than 370 patents, and has built a next‑generation multimodal intelligent customer‑service platform covering the full service chain.

The Vega v1 achievement raises JD’s NLP capabilities, promising stronger support for smart city, supply‑chain, and intelligent retail applications, and accelerating the growth of the digital‑intelligence industry.