ERNIE Evolve

ERNIE Evolve is Baidu ERNIE's automated research system, first described on 9 November 2026 — a concrete step toward putting recursive self-improvement (RSI) into real R&D. Given a clear goal and constraints, it proposes research directions, edits code, runs experiments and stores what it learned; only improvements that pass independent evaluation and review are kept.

It runs two nested loops. The outer loop uses past results and a knowledge base to pick new directions and run several research lanes in parallel; the inner loop keeps editing, evaluating and writing up experience inside one direction. Roles are split across agents: a Director chooses the next direction, a Selector picks the starting code and hypothesis, an Executor edits code in an isolated environment, an Evaluator scores the run with a fixed procedure the agents cannot see, and a Reviewer checks for goal drift or eval gaming.

Highlights:

• Introduced on 2026-11-09 on the ERNIE site and by Baidu's official account @Baidu_Inc;
• Reported publicly verifiable state-of-the-art numbers on NanoChat AutoResearch and NanoGPT Speedrun, with the optimization artifacts open-sourced;
• Already used on upcoming ERNIE 5.5 Flash inference work, including a Kunlunxin P800 low-latency path (~168 hours, generation speed to 623.6% of baseline), deterministic-inference throughput (~3 days, 133.7% of baseline) and a decode-speed pass (114.9% of baseline at matching accuracy);
• No public API or self-serve product yet — it is aimed at research and infra teams working on training, inference and kernel optimization;
• Write-up: https://ernie.baidu.com/blog/ernie-evolve-toward-automated-research-for-real-world-ai-rd .

Why it matters

Letting a model take part in its own R&D is something many labs tried in 2026. ERNIE Evolve is interesting less as "another coding agent" and more because Baidu says it already sits in the ERNIE team's daily engineering path: changes must pass the same load tests and correctness checks as a human engineer's patch before they reach production. Inference infrastructure is the first landing zone they name; the same loop is meant to extend to training systems, capability research and evals, always inside stated goals, permissions and resource limits, interruptible, and under ongoing safety review and human oversight.

Limits

There is no consumer access, pricing or signup. The public-benchmark leads are Baidu's own reports; the mean score on SOL-ExecBench was still slightly below the then-best public result. Whether automated research stays honest depends on how hard the eval is to game and whether human review keeps up.

1. Read the official note: https://ernie.baidu.com/blog/ernie-evolve-toward-automated-research-for-real-world-ai-rd (ERNIE blog, 2026-11-09);
2. Treat it as a research loop: give it a repo, a goal and an eval, let it edit code in isolation, and let an independent Evaluator/Reviewer decide what stays;
3. Follow the landings they already cite — ERNIE 5.5 Flash inference work, plus open-sourced results on NanoChat AutoResearch and NanoGPT Speedrun;
4. Do not look for a chat API or a consumer product page; none has been announced.