Baidu Introduces ERNIE Evolve, an Automated AI Research System, Demonstrating Significant Performance Gains
Baidu has unveiled ERNIE Evolve, an automated research system designed to optimize AI training and inference code. This system operates through two nested loops: an outer loop that determines research directions and an inner loop that develops and tests improvements within those directions. ERNIE Evolve has been tested on three public benchmarks—NanoChat AutoResearch, NanoGPT Speedrun, and SOL-ExecBench—where it surpassed the best public results in several metrics. For instance, it achieved a validation BPB of 0.8763 on NanoChat AutoResearch and reached target loss in 38.9 seconds on NanoGPT Speedrun Track 1. The system also optimized inference for the in-development ERNIE-5.5-Flash, leading to substantial performance gains, including a 623.6% increase in generation speed on Kunlunxin P800 and a 33.7% rise in deterministic throughput. Baidu views ERNIE Evolve as a step towards recursive self-improvement in AI, where AI systems enhance the software that builds and serves them.
