Baidu's ernie evolve system achieves new benchmarks in automated AI research

Baidu has unveiled ERNIE Evolve, a novel automated research system designed to enhance both the training and inference code of artificial intelligence models. This system operates by accepting a code repository, a specific objective, and evaluation criteria, subsequently selecting research paths, modifying code, and executing experiments. Only modifications that successfully pass a hidden evaluator and a separate reviewer are retained. ERNIE Evolve functions through two nested loops: an outer loop determines the area of investigation, while the inner loop develops and tests improvements within that direction. The system runs multiple research lanes concurrently, maintaining separate version trees for each, and logs failed experiments to prevent redundant testing. Baidu tested the system against NanoChat AutoResearch, NanoGPT Speedrun, and SOL-ExecBench. Furthermore, the system was applied to optimize inference for ERNIE-5.5-Flash, which is currently under development. Baidu presented this technology as a practical step toward recursive self-improvement (RSI), where AI systems iteratively improve the software that creates and runs them. The initial results were compared against older records, such as the ~79.7 s mark for Hyra and Recursive Speedrun times.