Physix FrontierPhysix Frontier · News Briefing Card (QbitAI · Nov 9, 2026)
Baidu ERNIE unveils ERNIE Evolve, already used in ERNIE 5.5 Flash R&D
KEY FACTS
- Baidu ERNIE publicly unveiled its automated research system ERNIE Evolve on November 9.
- The system can autonomously propose research directions, run experiments, and verify results.
- ERNIE Evolve achieved public state-of-the-art on three metrics across two AutoResearch tasks.
- The system has been used for inference optimization of ERNIE 5.5 Flash and has completed production-grade validation.
- Baidu open-sourced the ERNIE 5.5 Flash Base foundation model on November 5.
KEY DATA
increased to 623.6% of baselineKunlunxin P800 generation speed in low-latency scenarios
increased to 133.7% of baselineDeterministic inference throughput
increased to 114.9% of baselineInference performance with extreme decoding speed optimization
approximately 168 hoursERNIE Evolve low-latency optimization time
PHYSIX OBSERVATION
RSI has moved from a paper concept into the production environment. Baidu has landed the closed loop of AI autonomously modifying code first on inference optimization, and the gains are quantifiable. For the industry, this says more about how R&D itself is changing than any model benchmark score does; but automated changes still need evaluation and human oversight as a backstop, and the safety boundary remains the key variable before scaling.
Source: QbitAI report