Baidu has unveiled an automated research system that improves the software behind its own artificial intelligence models, describing it as a first practical step towards AI that helps to improve itself.

The system, called ERNIE Evolve, was introduced on Monday, 9 November 2026. Given a code repository, an objective and a set of evaluation criteria, it chooses research directions, edits the code, runs the experiments and keeps only the changes that survive an independent evaluator and a separate reviewer checking for goal drift and flaws in the tests.

Baidu presents the work as a move towards recursive self-improvement, in which AI systems take on the job of improving the software that builds and runs AI. The company says the system operates within human-defined repositories, objectives, evaluation rules, permissions and compute budgets, and must remain human-controlled and interruptible.

Public benchmark results

Baidu first tested ERNIE Evolve on three public benchmark suites: NanoChat AutoResearch, which measures model quality under a fixed training budget; NanoGPT Speedrun, which measures training efficiency; and SOL-ExecBench, which measures GPU kernel performance.

According to Baidu's reported results, the system beat the best public result on three of the four public metrics tested. On NanoChat AutoResearch it reached a validation score of 0.8763 bits per byte, against a previous public best of 0.8878, where a lower score is better. On NanoGPT Speedrun it cut the time to reach the target loss to 38.9 seconds from 39.9 seconds.

The results were not a clean sweep. On SOL-ExecBench, ERNIE Evolve scored a mean of 0.819 against the public best of 0.827, showing that the automated search did not lead on every task it attempted. Baidu has published the optimised benchmark code and measurement logs on GitHub, so developers can inspect the resulting code, although it has not released the ERNIE Evolve orchestrator itself, its prompts or its internal production environments.

Speed gains on live engineering work

Baidu then turned the system to live engineering work on ERNIE-5.5-Flash, a model still in development, running on the company's Kunlun P800 accelerator hardware. The company reports that four research rounds raised generation speed to 623.6 percent of the original baseline, equivalent to a 523.6 percent increase, while passing the stated correctness checks, with the model and deployment setup unchanged.

In a second task, deterministic inference for reinforcement-learning training, where each request must produce identical tokens and log probabilities regardless of the other requests in its batch, Baidu reports throughput rose 33.7 percent, time per output token fell 26 percent and time to first token fell 12 percent over about three days of work, with determinism preserved.

A third task, single-request decoding speed on graphics processors, improved by 14.9 percent over a service that engineers had already tuned by hand, with accuracy unchanged. The largest single improvement in the deterministic task replaced a sampling step that consumed 14 percent of GPU time per decoding step with a kernel that processes vocabulary chunks in parallel.

Claims remain company-reported

All of the figures are Baidu's own, measured by its own evaluator on a model that is not yet public, and they have not been independently reproduced. That matters: a system that grades its own homework needs outside verification before its numbers can be treated as settled fact.

Even so, the release is notable for what it makes public. Most talk of self-improving AI has been theoretical. Baidu has shown a working loop, published its benchmark code and logs, and been open about the task where it fell short. Its next step, the company says, is to apply the loop to training-system optimisation, then use model-capability measures to guide automated hypothesis generation, experiment design and validation.

Whether ERNIE Evolve keeps delivering gains as the research directions get harder will be the real test. For now, it is a concrete demonstration that AI research itself is becoming a job AI systems can help to do.