Baidu Wenxin Discloses RSI Progress for the First Time: ERNIE Evolve Already Used in Latest Model R&D, Wenxin 5.5 Flash May Be Released Soon
- Baidu's ERNIE team has disclosed its RSI system ERNIE Evolve, which it says has been used in the development of the upcoming ERNIE 5.5 Flash model.
- ERNIE Evolve uses a two-layer loop structure: an outer loop proposes research directions and an inner loop iterates on code, evaluation, and knowledge accumulation.
- In a low-latency optimization on Kunlun Core P800, ERNIE Evolve ran for about 168 hours across four research directions, achieving a generation speed of 623.6% of the baseline.
- For deterministic inference throughput optimization, ERNIE Evolve explored three directions over three days, increasing throughput to 133.7% of baseline while maintaining determinism.
- Baidu says the same research loop will extend to training system optimization, model capability research, and evaluation system iteration, with safety assessments and human oversight.
Source details
- Latest report
- 9 Nov 2026 09:48 UTC
- First seen here
- 9 Nov 2026 09:48 UTC
- Outlets
- 2
