Baidu ERNIE Evolve automates AI research and inference optimization

What is changing?

Baidu introduced ERNIE Evolve, a system that automates parts of AI research and inference engineering. Under explicitly defined objectives and constraints, it selects research directions, modifies code and runs experiments in isolated environments. Candidate results go through a separate evaluation process, and only changes that demonstrate improvement are retained.

In production tests reported by the company for ERNIE-5.5-Flash, deterministic throughput increased by 33.7%, time per output token fell by 26%, and time to first token fell by 12%. Baidu also reported a 14.9% speedup for single-request decoding. These are provider measurements rather than independent verification and reflect specific hardware and configurations.

ERNIE Evolve does not deploy unrestricted changes directly to production. Baidu says people define the objectives, the process can be monitored and interrupted, and results intended for production receive additional validation. The approach may shorten experimentation cycles in model development and inference optimization, but gains should be measured again across different infrastructure.