Baidu introduced ERNIE Evolve, an automated research system intended to help improve AI software through repeated experiments. The company describes it as a system that can choose what to investigate, modify code, run tests and use the results to guide later experiments. Its stated goal is to automate parts of AI research and development, including optimisation work that normally requires engineers to profile systems, propose changes and compare results. The announcement is significant because it moves AI agents closer to the development process of the models and infrastructure that power them.
The system separates research planning from code-level experiments
ERNIE Evolve is organised around two loops. An outer loop selects research directions and assigns work to parallel lanes, while an inner loop makes changes and evaluates candidate implementations. The system uses an experience bank to retain previous candidates and results, and a knowledge base containing papers, code and other domain material. Baidu says experiments run in isolated environments and that independent evaluators determine which changes should be kept. This separation is important because an automated system should not be allowed to approve its own changes solely on the basis of a metric it can manipulate.
Baidu says it tested the system on public research tasks covering training algorithms, training speed and GPU kernel optimisation. The company also applied it to inference optimisation for its in-development ERNIE-5.5-Flash model. On a GPU workload that had already been tuned by engineers, Baidu reports a 33.7 per cent throughput improvement after roughly three days of work, while preserving deterministic execution requirements. It also reports a further 14.9 per cent improvement in a decoding-speed experiment. Those figures are tied to the specific workloads and baselines described in the technical post, not a general claim that the system makes every model 33.7 per cent faster.
Baidu's reported speedups depend on workload and starting conditions
The Kunlunxin results need separate interpretation. Baidu has described a larger speedup on its own accelerator platform, but the starting configuration and amount of prior tuning affect the comparison. A system can produce a large percentage gain when it activates an unused feature or removes an obvious bottleneck, while producing smaller improvements on a highly optimised baseline. Baidu’s detailed write-up describes the changes it retained and the bottlenecks it tested, which is more useful than a single headline multiplier. Independent reproduction would help establish how much of the improvement comes from the agent’s research process and how much from the particular starting conditions.
Automating research introduces questions about reproducibility and control. The system must preserve source changes, record experiment settings, distinguish real improvements from measurement noise and reject changes that break correctness. In model optimisation, a faster result is not useful if it changes output quality, violates deterministic requirements or introduces a security issue. Baidu says its evaluator checks candidate changes independently and that the system operates under defined objectives and constraints. The next step will be to see whether external researchers can reproduce the results, whether the system can improve training and inference and how much human supervision remains necessary for high-impact changes.
A research agent that edits the software used by AI models also needs safeguards against optimising the wrong objective. Throughput can improve while latency, memory use or output quality gets worse, so evaluation must include constraints beyond a single speed measurement. Baidu says ERNIE Evolve uses independent evaluators and defined objectives, but teams adopting similar systems should preserve baselines, run repeated measurements and retain rejected experiments. The process should also be reproducible by human engineers. The broader direction is important: automated systems may increasingly contribute to chip design, model serving and training infrastructure. Their value will depend on whether they produce verifiable improvements rather than simply generating more code changes.