Baidu introduced ERNIE Evolve to automate recurring research work, from choosing a direction and changing code to testing results. The system separates decision making, implementation, evaluation and review, retaining failed attempts as information for later choices. Baidu identifies work on its still developing ERNIE 5.5 Flash among the applications.
Baidu, the Chinese technology company, introduced ERNIE Evolve, an automated research system that experiments with changes to artificial intelligence code. Its workflow combines choosing a research direction, implementing changes and evaluating the results. Baidu identifies work on ERNIE 5.5 Flash, a model still under development, among its applications.[1], [2]
At the upper level, a Director chooses the research direction and working version using experience and knowledge from earlier attempts. A Selector then chooses a hypothesis and ancestor version. An Executor implements the change in an isolated environment. An Evaluator checks correctness and performance, while a Reviewer examines the code before candidates are selected.[1]
Baidu says the evaluator stays fixed, with some checks hidden from the component proposing changes. Multiple research lanes explore experiments, and failed attempts inform later choices. The company also distinguishes hardware starting points: gains on its Kunlunxin P800 chip partly involved enabling existing optimized paths, while the graphics-processor environment had already received more tuning. Its acceleration results are developer measurements tied to those setups.[1]
