THE ESSENTIALS
  • The system separates research-direction selection, code changes and evaluation, according to Baidu.
  • Reported benchmark and inference improvements are company results, rather than independently replicated performance guarantees.

A research loop beyond code generation

Baidu described ERNIE Evolve in a November 9 research post as a system that chooses research directions, changes code, runs experiments and retains useful findings. The company says it has applied the system to inference optimization for ERNIE-5.5-Flash, which remains in development.

An outer loop selects directions from accumulated experiments and reference material. Parallel research lines then refine candidate implementations. Baidu says a fixed evaluator outside the editable repository checks performance and correctness, while a separate reviewer checks whether changes exploit evaluation loopholes.

Company-reported results and limits

Baidu reports results on NanoChat AutoResearch, NanoGPT Speedrun and SOL-ExecBench, alongside internal inference experiments. In one deterministic-inference workload, it reports a 33.7% throughput improvement after approximately three days of search, with determinism preserved. That result concerns a specified engineering setting, not a general improvement across models or hardware.

The post says expansion should remain controllable and pausable, within defined goals, permissions and resource boundaries, with human oversight. Its reported gains have not been independently replicated for this brief; they do not establish unrestricted autonomous research or universal production readiness.

Sources & context

Go to the original material. Company claims remain attributed to their sources.

01
ERNIE Evolve:让自动研究走进真实 AI 研发 · 文心大模型 ↗Retrieved and checked by the China AI Wire agent; evidence ledger retained internally.
Last updated November 15, 2026.Spotted an issue? Let us know ↗