Baidu unveils ERNIE Evolve automated research system
Baidu's ERNIE Evolve runs AI R&D experiments on its own and has already sped up inference for the in-development ERNIE-5.5-Flash, an early step toward what Baidu calls recursive self-improvement.
Baidu has described ERNIE Evolve, an automated research system. It takes a code repository, an optimization objective and evaluation criteria, then edits code and runs experiments in isolated environments with no human in the loop. The company calls it a "first step" toward recursive self-improvement, meaning AI systems that improve the models they run on.
The system uses two nested loops. In the outer loop, a Director proposes research directions. It draws on an Experience Bank of past candidates and results and on a Knowledge Base of papers, open-source code and domain knowledge. Each direction gets its own research lane, and several lanes run in parallel. Inside a lane, a Selector picks a candidate code version and a hypothesis to test, and an Executor writes the change.
Two checks decide whether a candidate is kept. The Evaluator is a fixed program that sits outside the repository and that no agent can see. Gains count only if they exceed a noise threshold. A Reviewer then reads the code changes. It discards candidates that exploit loopholes in the evaluation and blocks candidates that wander off the assigned direction from being chosen as the best version.
Baidu tested the system on three public tasks: NanoChat AutoResearch, NanoGPT Speedrun and SOL-ExecBench. These cover training algorithms, training speed and GPU kernel optimization. Baidu says ERNIE Evolve beat the baselines on all three. On NanoChat AutoResearch and NanoGPT Speedrun, it also says the system beat the best publicly verified results available at the time. The optimized results are open-sourced on GitHub.
Baidu also used the system to optimize inference for ERNIE-5.5-Flash, which is still in development. It says the changes passed production-grade validation and brought "substantial performance gains." The source material does not include specific benchmark figures or the size of those gains.