Baidu introduces ERNIE Evolve, a system that runs its own experiments to improve SI
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
SI is beginning not only to perform individual research tasks but also to run a cycle of selecting hypotheses, experimenting and validating improvements for subsequent models.
ERNIE Evolve selects research directions, modifies code and runs experiments in an isolated environment. An evaluator external to the agents checks correctness and performance gains, while a separate reviewer rejects solutions that exceed the specified boundaries or exploit weaknesses in the evaluation. On NanoGPT Speedrun, the system reduced the time needed to reach the target quality to 38.9 seconds, compared with the best public result of 39.9 seconds; on NanoChat AutoResearch it achieved 0.8763 BPB versus 0.8878. In GPU kernel optimization, its score of 0.819 was close to the SOL-ExecBench leader's 0.827. While optimizing the still-in-development ERNIE-5.5-Flash, the system increased deterministic inference throughput on GPUs by 33.7% in about three days. On Kunlunxin P800, generation speed rose to 623.6% of the baseline; a significant part of the gain came from enabling optimized execution paths that were already available. All figures are reported by Baidu.
Sources 2